Communications in Computer and Information Science
162
Abdulkadir Özcan Jan Zizka Dhinaharan Nagamalai (Eds.)
Recent Trends in Wireless and Mobile Networks Third International Conferences WiMo 2011 and CoNeCo 2011 Ankara, Turkey, June 26-28, 2011 Proceedings
13
Volume Editors Abdulkadir Özcan Girne American University Girne, TRNC, Turkey E-mail:
[email protected] Jan Zizka Mendel University Brno, Czech Republic E-mail:
[email protected] Dhinaharan Nagamalai Wireilla Net Solutions PTY Ltd Melbourne, VIC, Australia E-mail:
[email protected] ISSN 1865-0929 e-ISSN 1865-0937 ISBN 978-3-642-21936-8 e-ISBN 978-3-642-21937-5 DOI 10.1007/978-3-642-21937-5 Springer Heidelberg Dordrecht London New York Library of Congress Control Number: 2011929887 CR Subject Classification (1998): C.2, H.3.4-5, G.2.2
© Springer-Verlag Berlin Heidelberg 2011 This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India Printed on acid-free paper Springer is part of Springer Science+Business Media (www.springer.com)
Preface
The Third International Conference on Wireless, Mobile Networks (WiMo 2011) and the Third International Conference on Computer Networks and Communications (CoNeCo - 2011) were held in Ankara, Turkey, during June 26-28, 2011. They attracted many local and international delegates, presenting a balanced mixture of intellects from all over the world. The goal of this conference series is to bring together researchers and practitioners from academia and industry to focus on understanding wireless, mobile networks and communications and to establish new collaborations in these areas. Authors are invited to contribute to the conference by submitting articles that illustrate research results, projects, survey work and industrial experiences describing significant advances in all areas of wireless, mobile networks and communications. There were 202 submissions to the conference and the Program Committee selected 40 papers for publication. All the submissions underwent a strenuous peerreview process which comprised expert reviewers. These reviewers were selected from a talented pool of Technical Committee members and external reviewers on the basis of their expertise. The papers were then reviewed based on their contributions, technical content, originality and clarity. The entire process, which includes the submission, review and acceptance processes, was done electronically. All these efforts undertaken by the Organizing and Technical Committees led to an exciting, rich and a high-quality technical conference program, which featured high-impact presentations for all attendees to enjoy, appreciate and expand their expertise in the latest developments in wireless, mobile networks and communications research. The book is organized as a collection of papers from the Third International Conference on Wireless and Mobile Networks (WiMo-2011), the Third International Conference on Computer Networks and Communications (CoNeCo 2011), the Third International Workshop on Grid Computing (GridCoM - 2011) and the Second International Workshop on Communications Security and Information Assurance (CSIA- 2011). Finally, we would like to thank the General Chairs, local organizing team and Program Committee members and reviewers for arranging and organizing this conference. ¨ Abdulkadir Ozcan Jan Zizka Dhinaharan Nagamalai
Organization
The Third International Conference on Wireless, Mobile Networks (WiMo 2011)
General Chairs Sevki Erdogan Michael R. Peterson Natarajan Meghanathan
University of Hawaii, USA University of Hawaii, USA Jackson State University, USA
General Co-chairs Raja Kumar, M. Abdulkadhir Ozcan
Universiti Sains Malaysia, Malaysia The American University, North Cyprus, Turkey
Steering Committee Selma Boumerdassi Chih-Lin Hu Dhinaharan Nagamalai Krzysztof Walkowiak Atilla Elci Aysegul Alaybeyoglu Muhammed Salamah Jan Zizka
CNAM/Cedric, France National Central University, Taiwan Wireilla Net Solutions PTY LTD, Australia Wroclaw University of Technology, Poland Eastern Mediterranean University (TRNC), North Cyprus Celal Bayar Universitesi, Turkey Eastern Mediterranean University, KKTC, Turkey Mendel University, Brno, Czech Republic
Program Committee Members Kayhan Erciyes TurkeyDerya Birant Hesham El Zouka
Juha-Matti Vanhatupa
Izmir University, Turkey Dokuz Eylul University, Turkey Arab Academy for Science and Technology and Maritime Transport(AAST), Egypt Tampere University of Technoloy, Finland
VIII
Organization
Strassner John Charles Sahin Albayrak Jeong-Hyun Park Vishal Sharma H.V. Ramakrishnan Yeong Deok Kim Andy Seddon Balasubramanian Karuppiah Bong-Han, Kim Cho Han Jin David W. Deeds Girija Chetty Henrique Joao Lopes Domingos Jacques Demerjian Jose Enrique Armendariz-Inigo Krzysztof Walkowiak Marco Roccetti Michal Wozniak Phan Cong Vinh Yannick Le Moullec John Karamitsos Khoa N. Le Lu Yan Nidaa Abdual Muhsin Abbas Kamalrulnizam Abu Bakar Doina Bein M. Rajarajan Mohammad Momani Mohamed Hassan Salman Abdul Moiz Lakshmi Rajamani Amr Youssef Wichian Sittiprapaporn
Pohang University of Science and Technology, South Korea Technische Universit¨ at Berlin, Germany Electronics Telecommunication Research Institute, South Korea Metanoia Inc, USA Dr. MGR University, India Woosong University, South Korea Asia Pacific Institute of Information Technology, Malaysia Dr. MGR University, India Chongju University, South Korea Far East University, South Korea Shingu College, South Korea University of Canberra, Australia University of Lisbon, Portugal CS, Homeland Security, France Universidad Publica de Navarra, Spain Wroclaw University of Technology, Poland Universty of Bologna, Italy Wroclaw University of Technology, Poland London South Bank University, UK Aalborg University, Denmark University of the Aegean, Samos, Greece Griffith School of Engineering, Gold Coast Campus, Australia University of Hertfordshire, UK University of Babylon, Iraq Universiti Teknologi Malaysia, Malaysia The Pennsylvania State University, USA City University, UK University of Technology Sydney, Australia American University of Sharjah, UAE Centre for Development of Advanced Computing, India Osmania University, India Concordia University, Canada Mahasarakham University, Thailand
Organization
IX
The Third International Conference on Computer Networks and Communications (CoNeCo - 2011)
General Chairs Jan Zizka Khoa N. Le Selma Boumerdassi
SoNet/DI, FBE, Mendel University in Brno, Czech Republic University of Western Sydney, Australia Conservatoire National des Arts et Metiers (CNAM), France
Steering Committee Natarajan Meghanathan Jacques Demerjian Nabendu Chaki Henrique Jo˜ao Lopes Domingos
Jackson State University, USA CS (Communication & Systems), France University of Calcutta, India University of Lisbon, Portugal
Program Committee Members Abdul Kadir Ozcan Adalet N. Abiyev Andy Seddon Balasubramanian Karuppiah Bong-Han, Kim Bulent Bilgehan Cho Han Jin Dhinaharan Nagamalai Farhat Anwar Girija Chetty Henrique Jo˜ao Lopes Domingos Hoang, Huu Hanh Hwangjun Song Jacques Demerjian Jae Kwang Lee Jan Zizka, SoNet/DI, FBE Jos´e Enrique Armend´ariz-Inigo Jungwook Song Krzysztof Walkowiak Marco Roccetti Michal Wozniak Murugan D. N. Krishnan
The American University, Cyprus The American University, Cyprus Asia Pacific Institute of Information Technology, Malaysia Dr. MGR University, India Chongju University, South Korea The American University, Cyprus Far East University, South Korea Wireilla Net Solutions PTY LTD, Australia International Islamic University, Malaysia University of Canberra, Australia University of Lisbon, Portugal Hue University, Vietnam Pohang University of Science and Technology, South Korea CS, Homeland Security, France Hannam University, South Korea Mendel University in Brno, Czech Republic Universidad P´ ublica de Navarra, Spain Konkuk University, South Korea Wroclaw University of Technology, Poland Universty of Bologna, Italy Wroclaw University of Technology, Poland Manonmaniam Sundaranar University, India Manonmaniam Sundaranar University, India
X
Organization
Natarajan Meghanathan Paul D. Manuel Phan Cong Vinh Ponpit Wongthongtham Prabu Dorairaj Rajeswari Balasubramaniam Rakhesh Singh Kshetrimayum Ramayah Thurasamy Sarmistha Neogy SunYoung Han Thandeeswaran R. Yannick Le Moullec Yeong Deok Kim Boo-Hyung Lee Cynthia Dhinakaran Ho Dac Tu John Karamitsos Johnson Kuruvila Doina Bein Dario Kresic Salah M. Saleh Al-Majeed Virgil Dobrota Ahmed M. Khedr Polgar Zsolt Alfred Cristina Serban Genge Bela Haller Piroska Alejandro Garces Razvan Deaconescu
Jackson State University, USA Kuwait University, Kuwait London South Bank University, UK Curtin University of Technology, Australia NetApp Inc., India SriVenkateshwara Engineering College, India Indian Institute of Technology, Guwahati, India Universiti Sains Malaysia, Malaysia Jadavpur University, India Konkuk University, South Korea VIT University, India Aalborg University, Denmark Woosong University, South Korea KongJu National University, South Korea Hannam University, South Korea Waseda University, Japan University of the Aegean, Samos, Greece Dalhousie University, Halifax, Canada The Pennsylvania State University, USA University of Zagreb, Croatia University of Essex, UK Technical University of Cluj-Napoca, Romania Sharjah University, UAE Technical University of Cluj Napoca, Romania Ovidius University of Constantza, Romania Joint Research Centre, European Commission, Italy Petru Maior University, Tirgu Mures, Romania Jaume I University, Spain University Politehnica of Bucharest, Romania
Organized By
ACADEMY & INDUSTRY RESEARCH COLLABORATION CENTER (AIRCC)
Table of Contents
The Third International Conference on Wireless and Mobile Networks (WiMo-2011) A Survey of Reinforcement Learning Based Routing Protocols for Mobile Ad-Hoc Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Saloua Chettibi and Salim Chikhi
1
Detection of Denial of Service Attack Due to Selfish Node in MANET by Mobile Agent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Debdutta Barman Roy and Rituparna Chaki
14
A Novel Power-Balancing Routing Scheme for WSN . . . . . . . . . . . . . . . . . . Ayan Kumar Das and Rituparna Chaki
24
SRCHS – A Stable Reliable Cluster Head Selection Protocol . . . . . . . . . . . Ditipriya Sinha and Rituparna Chaki
35
Handover Latency Reduction Using Integrated Solution Scheme for Proxy Mobile IPv6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Md. Mahedi Hassan and Kuan Hoong Poo Modeling and Simulation Analysis of QPSK System with Channel Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . T.P. Surekha, T. Ananthapadmanabha, C. Puttamadappa, and A.P. Suma
45
57
Combating Sybil Attacks in Vehicular Ad Hoc Networks . . . . . . . . . . . . . . Khaled Mohamed Rabieh and Marianne Amir Azer
65
TCSAP: A New Secure and Robust Modified MANETconf Protocol . . . . Abdelhafid Abdelmalek, Zohra Slimane, Mohamed Feham, and Abdelmalik Taleb-Ahmed
73
Highly Resilient Communication Using Affine Planes for Key Predistribution and Reed Muller Codes for Connectivity in Wireless Sensor Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Samiran Bag, Amrita Saha, and Pinaki Sarkar A Cyclic-Translation-Based Grid-Quadtree Index for Continuous Range Queries over Moving Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hao Chen, Guangcun Luo, Aiguo Chen, Ke Qin, and Caihui Qu Two-Stage Clustering with k -Means Algorithm . . . . . . . . . . . . . . . . . . . . . . Raied Salman, Vojislav Kecman, Qi Li, Robert Strack, and Erick Test
83
95 110
XII
Table of Contents
An Energy and Delay-Aware Routing Protocol for Mobile Ad-Hoc Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jihen Drira Rekik, Le¨ıla Baccouche, and Henda Ben Ghezala Energy-Aware Transmission Scheme for Wireless Sensor Networks . . . . . . Abdullahi Ibrahim Abdu and Muhammed Salamah PRWSN: A Hybrid Routing Algorithm with Special Parameters in Wireless Sensor Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Arash Ghorbannia Delavar, Javad Artin, and Mohammad Mahdi Tajari
123 135
145
Cone Tessellation Model for Three-Dimensional Networks . . . . . . . . . . . . . G¨ ozde Sarı¸sın and Muhammed Salamah
159
Post Disaster Management Using Delay Tolerant Network . . . . . . . . . . . . . Sujoy Saha, Sushovan, Anirudh Sheldekar, Rijo Joseph C., Amartya Mukherjee, and Subrata Nandi
170
The Performance Comparison between Hybrid and Conventional Beamforming Receivers in a Multipath Channel . . . . . . . . . . . . . . . . . . . . . . Rim Haddad and Ridha Bouallegue A Qualitative Survey on Multicast Routing in Delay Tolerant Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sushovan Patra, Sujoy Saha, Vijay Shah, Satadal Sengupta, Konsam Gojendra Singh, and Subrata Nandi
185
197
The Third International Conference on Computer Networks and Communications (CoNeCo-2011) Integrating RFID Technology in Internet Applications . . . . . . . . . . . . . . . . Simon Fong
207
BPSO Algorithms for Knapsack Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . Amira Gherboudj and Salim Chikhi
217
Systematic Selection of CRC Generator Polynomials to Detect Double Bit Errors in Ethernet Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Behrouz Zolfaghari, Hamed Sheidaeian, and Saadat Pour Mozafari
228
Security Analysis of Ultra-lightweight Protocol for Low-Cost RFID Tags: SSL-MAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mehrdad Kianersi, Mahmoud Gardeshi, and Hamed Yousefi
236
File Distribution Algorithm from Multiple Road Side Units in Vehicular Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Saleh Yousefi, Amin Amini Maleki, and Reza Hashemi
246
Table of Contents
XIII
SMART-IP: A Multi-Agent System for Network Analysis and IP Addressing Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Samia Boucherkha and Mohamed Nadjib Djeghri
256
Adaptive QoS Resource Management by Using Hierarchical Distributed Classification for Future Generation Networks . . . . . . . . . . . . . . . . . . . . . . . Simon Fong
266
0.18um CMOS Technology in Implementation of S Box and a Modified S Box . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . K. Rahimunnisa, K. Rajeshkumar, and S. Sureshkumar
279
A Survey of Middleware for Internet of Things . . . . . . . . . . . . . . . . . . . . . . . Soma Bandyopadhyay, Munmun Sengupta, Souvik Maiti, and Subhajit Dutta
288
New Framework for Dynamic Policy Management in Grid Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tariq Alwada’n, Helge Janicke, Omer Aldabbas, and Hamza Aldabbas
297
Zone Based Seamless Vertical Handoff Technique between WiFi and WiMax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Abhijit Sarma and Sukumar Nandi
305
The Third International Workshop on Grid Computing (GridCom-2011) A Cluster-Based Dynamic Load Balancing Protocol for Grids . . . . . . . . . . ¨ Re¸sat Umit Payli, Kayhan Erciyes, and Orhan Dagdeviren
315
A P2P Based Scheduler for Home Grids . . . . . . . . . . . . . . . . . . . . . . . . . . . . Erick Lopes da Silva and Peter Linington
325
A Multi-Agent System-Based Resource Advertisement Model for Grid Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Muntasir Al-Asfoor, Maria Fasli, and Salah Al-Majeed
337
Grid-Enabled Framework for Large-Scale Analysis of Gene-Gene Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Moez Ben Haj Hmida and Yahya Slimani
348
A Context-Based Cyber Foraging Approach to Mobile Computing Empowerment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Somayeh Kafaie, Omid Kashefi, and Mohsen Sharifi
358
Model Driven Prototyping with Modelibra . . . . . . . . . . . . . . . . . . . . . . . . . . Dzenan Ridjanovic
368
XIV
Table of Contents
The Second International Workshop on Communications Security and Information Assurance (CSIA 2011) An Implementation of Axml(T ) : An Answer Set Programming Based Formal Language of Authorisation for XML Documents . . . . . . . . . . . . . . . Sean Policarpio and Yun Bai
378
On Cloud Computing Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yun Bai and Sean Policarpio
388
PAPR Reduction in OFDM by Using Modernize SLM Technique . . . . . . . Ashutosh K. Dubey, Yogeshver Khandagre, Ganesh Raj Kushwaha, Khushboo Hemnani, Ruby Tiwari, and Nishant Shrivastava
397
Application of Integrated Decision Support Model in Tendering . . . . . . . . Fadhilah Ahmad and M. Yazid M. Saman
406
Query and Update Support for Indexed and Compressed XML (QUICX) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Radha Senthilkumar and A. Kannan
414
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
429
A Survey of Reinforcement Learning Based Routing Protocols for Mobile Ad-Hoc Networks Saloua Chettibi and Salim Chikhi SCAL Group, MISC Laboratory, Mentouri University, Constantine, Algeria {sa.chettibi,slchikhi}@yahoo.com
Abstract. Designing mobility and power aware routing protocols have made the main focus of the early contributions to the field of Mobile Ad-hoc NETworks (MANETs). However, almost all conventional routing protocols for MANETs suffer from their lack of adaptivity leading to their performance degradation under varying network conditions. In fact, this is due to both simplistic conception hypotheses they made about the network and to the use of some prefixed parameters in protocols implementations. Currently, artificial intelligence methods like Reinforcement Learning (RL) are widely used to design adaptive routing strategies for MANETs. In this paper, we present a comprehensive survey of RL-based routing protocols for MANETs. Besides, we propose some future research directions in this area. Keywords: Mobile Ad-hoc Networks, Routing, Reinforcement learning.
1 Introduction A MANET is a transient network dynamically formed by a collection of arbitrarily located wireless and mobile nodes communicating without any pre-established network infrastructure. Since no base station is available in a MANET, then all nodes must cooperate to ensure routing service. Hence, each node runs as a router by forwarding its neighbors’ traffic. In reality, this ad-hoc functioning mode is advantageous because it ensures a rapid, an easy and economic network deployment. Furthermore, it offers fault tolerance property since no central node is designated. Particularly, MANETs are suitable to guarantee communication when the deployment of a fixed infrastructure is impossible or does not justify its cost, or simply when conventional fixed infrastructures are destroyed. Salient characteristics of MANETs can be summarized as follows: multi-hop communications, very dynamic topology, limited link capacity and quality, limited energy resources. Obviously, these characteristics make conventional routing protocols for wired networks which are based on hypothesis of fixed topology and predictable communication medium and energy-unconstrained hosts, inappropriate to MANETs. To deal with all the aforementioned challenging characteristics, many routing protocols have been proposed in the literature for MANETs. Currently, researchers focus is on the design of adaptive routing protocols that are built on the A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 1–13, 2011. © Springer-Verlag Berlin Heidelberg 2011
2
S. Chettibi and S. Chikhi
top of one or many artificial intelligence techniques. In this survey, we deal particularly with RL-based routing protocols where routing choices and/or routing parameters adjusting are formulated as decisions making problems. To the best of our knowledge, this paper is the first one in the literature dedicated to survey RL-based routing protocols for MANETs. The remainder of this paper is organized as follows: section 2 introduces the general formulation of the RL problem. Then, some basic RL algorithms are outlined. In section 3, we motivate the modelization of adaptive network routing problem in MANETs as a RL problem. Section 4 describes most major contributions to the field of RL-based routing for MANETs. In the light of this description, we state our conclusions and we draw some future research directions in section 5.
2 Reinforcement Learning The reinforcement learning [1] is a sub-area of machine learning concerned with learning from interaction by trials and errors how to behave in order to achieve a goal. Important notions in RL problem formulation as a Markov Decision Process and its resolution can be summarized as follows [1]: Markov property. An environment satisfies the Markov property if the state signal compactly summarizes the past without degrading the ability to predict the future. If the Markov property holds, then the RL environment is called a Markov Decision Process (MDP). Markov Decision Process (MDP). Formally, a finite MDP is a tuple <S,A,T,R> where is a finite set of environment sates, is a set of actions available at the agent, T:S×A→Π(S) is the state transition function giving for each state and action a probability distribution over states, R: S×A×S→ is the reinforcement function that indicates the real-value obtained when transiting from a state to another taking a particular action. Return. The return Rt is function of future rewards that the agent seeks to maximize. It can be defined as a simple finite sum of rewards when the agent-task breaks to finite-episodes. Instead, for continuing tasks, Rt is formulated as the infinite sum of discounted rewards. Partially Observable MDP (POMDP). The POMDP is a variant of the MDP in which the state of the environment is only partially visible to the learning agent. What are available are indirect, potentially stochastic observations of the environment state. Value-functions. Almost all reinforcement learning algorithms are based on estimating either state-value or action-value functions. State-value function, Vπ(s), estimates the expected future reward to the agent when starting in state and following the policy π thereafter. Action-value function, Qπ(s,a), estimates the expected future reward to the agent when it performs a given action in a given state and following the policy π thereafter.
A Survey of Reinforcement Learning Based Routing Protocols
3
2.1 RL Algorithms Features As defined in [1], any algorithm that can solve a reinforcement learning problem either defined by a MDP or a POMDP is an RL algorithm. RL algorithms may vary according to multiple dimensions [1]: Bootstrapping Vs Sampling. We say that a RL method bootstraps if it updates estimates of the values of states are based on estimates of the values of successor states. In Contrast, a RL method that relies on sampling learns value functions from experiences having the form of sample sequences of states, actions, and rewards from on-line or simulated interaction with an environment. Model-based Vs Model-free RL algorithms. In RL, a model consists of knowledge of the state transition probability and the reinforcement functions. RL model-free methods learn a policy without learning a model, whereas a model-based method learns a model and use it to derive a policy. Greedy Vs ε -greedy and Soft-max action-selection rules. A very intuitive way to achieve a maximum return is to always choose the action with the highest expected reward. We call this a greedy action-selection rule. However, this rule limits agent exploration of new appearing optimal actions. In effect, the most important issue in all RL methods is how to balance exploration and exploitation. To do so, ε -greedy and Soft-max rules are generally used .An ε-greedy rule selects the best action most of the time, and selects uniformly with a small probability,ε, an action at random. In a softmax rule, the highest selection probability is attributed to the best action whereas all the others are ordered in function of their estimated values. A frequently used softmax rule is the Boltzmann rule that chooses action at time step with / ⁄∑ / probability:e e , where is a positive parameter called the temperature. A high temperature value implies that all actions will have approximately the same selection-probability whereas a low temperature involves a more significant difference in actions selection probabilities. Off-policy Vs On-policy RL algorithms. In an off-policy RL algorithm, the learning agent follows a different policy called “behavior policy” than the one it is actually learning about called “estimation policy”. This is in contrast to an on-policy method that attempts to evaluate and improve the policy that is actually used to make decisions. Phylogenetic Vs Ontogenetic RL algorithms. A classification proposed in [2] divides RL methods into Phylogenetic and Ontogenetic algorithms. In Phylogenetic methods, the RL problem is seen as a black box optimization problem with the specific goal of optimizing a policy for maximal accumulated reward. Intuitively, any optimization method can be applied. Ontogenetic algorithms are based on a mapping between states or states-actions pairs to their corresponding expected rewards. Then, a policy is defined according to the obtained sate-value or action-value functions. 2.2 RL Algorithms In what follows, we only focalize on algorithms that are directly related to the RL-based routing protocols described in this paper.
4
S. Chettibi and S. Chikhi
Q-learning. The Q-learning algorithm [3] is a model-free Off-policy RL-method that belongs to the class of TD (Temporal Difference) methods. TD methods combine sampling and bootstrapping where the learning agent takes a sample of just one step and then bootstraps information. Let us define , , , ́ to be an experience tuple summarizing a single transition in the environment. Here, is the agent state before the transition, is its choice of action, r the immediate reward it receives and ́ the resulting state. The one-step Q-learning version of Q-learning algorithm is depicted on the Fig.1. α: learning rate; γ: discount factor (0 ,
, ,
,
γ
,
1)
,
,
Fig. 1. The Q-learning Algorithm
Monte Carlo methods. MC methods [1] are model-free RL resolution methods based on averaging sample returns. To ensure that well-defined returns are available, MC methods are defined only for episodic tasks. It is only upon the completion of an episode that action-value functions, Q(s,a), and policies are changed. We distinguish two families of MC methods namely: the every-visit and the first-visit MC methods. The former estimates the value of a state-action pair as the average of all returns that have followed visits to the state in which the action was selected, whereas the latter averages only returns following the first time in each episode that the state was visited and the action was selected. In addition, we can find two incarnations of MC methods, namely, on-policy and off-policy MC. The first visit ε-greedy on-policy version is depicted in Fig. 2. ,
,
:
,
;
, ,
For all
: ,
,
,
, ;
,
, 1 |
|
|
|
Fig. 2. The first visit ε-greedy on-policy MC method
;
,
A Survey of Reinforcement Learning Based Routing Protocols
5
Collaborative Reinforcement Learning. CRL [4] extends the conventional RL framework with feedback models for decentralized multi-agent systems. The feedback models include a negative feedback and a collaborative feedback models. The former model decays an agent’s local view of its neighborhood either by constraints in the system or by a decay model. The latter model allows agents to exchange the effectiveness of actions they have learned with one another. In CRL, RL agents collaborate to solve the optimization problem. To do so, this latter is divided into a set of Discrete Optimization Problems (DOPs). The set of actions that a CRL-agent can execute include DOP actions that try to solve the DOP locally, delegation actions that delegate the solution of the DOP to a neighbor and a discovery action that allows agents to find new neighbors. In fact, CRL is a model-based RL technique with the following update rule: ,
́| ,
,
.
́| , ́
(1)
́
́ | , is the Where is a delegation action; , is the MDP termination cost; transition model; ́ is the estimated optimal value function for the next state at ́| , is the estimated connection cost to the next state. agent and Policy search by Gradient. The RL problem can be addressed as a search problem in the space of behaviors where the evaluation of a candidate policy is done by trial through the interaction with the environment. Hence, the gradient algorithm can be used for optimal policy search in RL problem. The first to introduce policy search via gradient for RL was Williams in the REINFORCE algorithm [5]. Generally speaking, the idea behind policy search by gradient is to start with some policy, evaluate it and make an adjustment in the direction of the empirically estimated gradient of the aggregate reward, in order to obtain a local optimum policy [6].
3 Adaptive Network-Routing as a RL Problem A routing protocol for MANETs should be adaptive in face of frequent and unpredictable changes in network topology. Moreover, adaptivity in face of changing traffic loads is very important to avoid congestion areas in the network. Hence, we need to design adaptive routing policies which have the ability to learn directly by interacting with their operational environment. Early proposed routing protocols for MANETs have adopted different mechanisms to enhance adaptivity in face of nodes mobility such as routes maintenance in reactive protocols [7],[8] and periodic or event-based updates in proactive protocols[9],[10]. In addition, using link-stability as a routing-metric enhances adaptivity in presence of poor quality and instable links [11], [12]. Generally, all proposed routing protocols perform well in their experimental settings. Nevertheless, if experimented with different parameters-values or if their conception hypothesis violated, not surprisingly, they will perform very badly. In reality, almost all existing routing protocols for MANETs make very simplistic assumptions about the network characteristics. Namely, a perfect wireless network model is generally considered, where all links in the network are assumed either on or off and where all functioning links are assumed to have the same quality.
6
S. Chettibi and S. Chikhi
In addition, topology is considered to be random which is not always true. Furthermore, some routing protocols functional parameters are simply prefixed thresholds although the fact of their dependence of many network conditions. For example, minimal residual-battery and reputation values are commonly used in energy and security constrained routing schemes, respectively. All the above mentioned factors limit routing protocols adaptivity when encountering varying network conditions in terms of traffic, mobility and links quality. The reinforcement learning has been shown to be an appropriate framework to design adaptive routing policies in fixed networks. For example, the first application of a RL technique to deal with packet routing in network communication was Q-routing [13]. Q-Routing is a distributed version of the conventional Q-learning. In Q-routing, each node makes its routing decisions based on the local routing information represented as a lookup table of Q-values that estimate the quality of the alternative routes. When a node sends a packet to one of its neighbors, it proceeds to update the corresponding entries in its Q-table. This way, as a node routes packets its Q-values gradually incorporate more global information. It was shown that such network-exploration enable nodes to adapt their routing policies to changes in trafficloads. More recently researchers were interested to the application of RL algorithms to achieve adaptive routing in MANETs. The following section describes most major researchers’ contributions dealing with RL-based routing in MANETs.
4 RL-Based Routing Protocols for MANETs To the best of our knowledge, the first application of reinforcement learning to the routing problem in MANETs was in Q-MAP protocol [14] where Q-learning approach was used to find and build the optimal multicast tree. However, authors in [15] claim that the learning in Q-MAP is exploration-free which not only contradicts the learning paradigm but makes the protocol insensitive to topology changes and thus reduces it to a static approach. Therefore, we omit the description of Q-MAP in this survey. 4.1 Mobility Aware Q-Routing In [16], the authors proposed a straightforward adaptation of traditional Q-routing [13] algorithm to the context of ad-hoc mobilized networks1 in order to achieve a traffic-adaptive routing. For ease of referencing, we call this protocol MQ-routing. Similarly to Q-routing, in MQ-routing, each node learns the expected delivery time to destination node via each possible neighbor , , , which is updated as is the time the shown in equation (2),where 0 1 is the learning rate and current packet spent on the queue at node before being sent off at period time t. ,
1
,
min
,
(2)
When a node receives a packet for destination , it sends the packet to the neighbor , .To take care of nodes mobility, with the lowest estimated delivery time 1
In a mobilized ad-hoc network, nodes have control on their own movements.
A Survey of Reinforcement Learning Based Routing Protocols
7
two additional rules are proposed for Q-values updates of neighboring , ∞ when moves out of range; and , 0 when moves nodes: into range.Note that the second update rule is made optimistic to encourage exploration of new coming neighbors. In simulations, the authors have considered a network with 10 sources, 15 mobile nodes and one receiver. The considered movement policy called centroidal is as follows: a node that is holding a connection move to the middle of its connected neighbors, which increases the likelihood of preserving these connections over time. Reported results from MQ-routing comparison with a typical directional routing protocol showed the outperformance of MQ-routing in terms of success rate in function of buffer size. This is due to the fact that the MQ-routing creates alternate paths to the receiver as soon as a path becomes congested. In directional routing, on the other side, certain paths become overloaded with traffic leading to significant packet drop. Even in cases where buffer size is not a direct constraint, the same results hold. In addition, since next hops are chosen among neighbors in receiver’s direction, directional routing is unable to find circuitous paths from sources to the receiver. Under random movement policy, both protocols have marked a degradation of their performances. Particularly, MQ-routing has performed somewhat worse than directional routing. LQ-routing [17], a similar work to MQ-routing, combines Q-routing with DSDV routing protocol [9]. To deal with mobility, the notion of paths lifetime was introduced to reflect paths stability. The proposed routing schema has been shown to outperform DSDV under high traffic loads. 4.2 RL - Based QoS Path Search Strategy To deal with delay-constrained and least-cost routing in MANETs, a combination of the TBP (Ticket-Based Probing) path search strategy [18] with an adaptation of the first visit ONMC method for POMDPs was proposed in [19]. In the original TBP scheme, the amount of flooding is controlled by issuing a limited number of logical tickets M0 at the source node that is computed via a heuristic rule. In fact, M0 is the sum of Y0 (yellow tickets) and G0 (green tickets) used to maximize the chances of finding feasible and low cost paths, respectively. The authors’ contribution in [19] is the use of the first-visit ONMC method to determine M0 value. This latter is chosen . The source node selects an action among a finite set of actions 0, … , , ∆ :1 depending on the current observation belonging to ,1 ∆ , where is the number of discrete end-to-end delay intervals; ∆ is the is the interval on number of discrete end-to-end delay variation intervals; is interval on 0, ∞ (this variable is included to reduce the 0, ∞ and ∆ uncertainty of the actual end-to-end delay).If at least one feasible path is found, then a , is generated. Otherwise, the action is penalized: reward 0 ,
– 0
0
(3)
0
is the immediate reward parameter for a given service-type j. Note Where that the authors have omitted G0 (i.e. G0= 0). Thus, tolerating high cost paths and
8
S. Chettibi and S. Chikhi
maximizing the chance of finding feasible paths. But if multiple feasible paths are discovered, then destination node chooses the least cost path. Once the discovery process completed, destination node returns an acknowledgment message including the new end-to-end delay and its variation. By receiving this message, the source node updates the corresponding entry in its information table. The simulation results have shown that the TBP scheme based on the ONMC method achieves 22.1–58.4% reduction in the average number of search messages in comparison to the flooding-based TBP scheme with a diminution of 0.5–1.7% in success ratio. In addition, the ONMC scheme can reach 13–24.3% higher success ratio than the original TBP scheme but with higher average message overhead. However, as the maximum number of allowable tickets is reduced to a level in which the average message overhead of the ONMC and the original TBP schemes are of the same scale, the ONMC scheme still marks 28% higher success ratio and 7% lower average path cost over the original TBP scheme. 4.3 RL-Based Secure Routing For secure routing, reputation schemes are widely used to identify and avoid malicious nodes. The reputation of a node is function of the number of data packets that have been successfully relayed by that node. Indeed, almost all proposed reputation schemes rely on a fixed-minimum acceptable reputation threshold in the forwarding-nodes selection. However, reputation values vary dynamically in function of traffic load and behavior of nodes themselves. This was the main motivation for authors in [20] to adopt the first visit ONMC RL method to enhance the reputation schema already proposed in [21]. In [20], each mobile node learns a policy for selecting neighboring nodes in a path search. A node’s state-set contains quantized reputation values of its neighbors. If a route search succeed then a reward of +1 is assigned to every node in all successful paths; if no path is found then a reward of 0 is given to all nodes involved in the route discovery. The simulation results have shown that the proposed solution can attain up to 89% and 29% increase in throughput over the reputation only scheme with a fixed reputation threshold of 0.5 for scenarios of a static and a dynamic topology, respectively. 4.4 CRL-Based Routing In [22], a CRL-based reactive routing protocol for MANETs called SAMPLE was proposed. The envisioned optimization goals are to maximize overall network throughput, maximize the ratio of delivered to undelivered packets and minimize the number of transmissions required per packet sent. In SAMPLE, each agent stores the last advertised route cost to a given destination from each of its neighbors in a routing table, but considers this value to decay from the time it is advertised. Hence, routes that are not advertised are progressively degraded and eliminated from consideration for routing decisions. Another source of negative feedback is network congestion that causes routing agent to choose alternate routes. In contrast, stable routes are reinforced by positive feedback. The routing problem in SAMPLE is modeled as an absorbing MDP where a node n state indicates either a packet is in a buffer waiting to be forwarded, has been successfully unicast to a neighbor, or has been delivered at
A Survey of Reinforcement Learning Based Routing Protocols
9
node n . The actions available at different states in the MDP are a packet delivery action, a broadcast action to discover new neighbors, links, and routes; and for each neighboring node, a delegation action (unicast). Concerning delegation actions, the decision of which next hop to take is chosen probabilistically using Boltzmann-action selection rule. Furthermore, SAMPLE also uses a simple greedy heuristic in order to restrict exploration to useful areas of the network by only allowing a node to forward to those neighboring nodes with a value that is less than its function value. To learn new routes, the discovery action is permitted with a certain probability. In SAMPLE, a statistical transition model that favors stable links is considered. It acquires information about the estimated number of packets required for a successful unicast as an indication of links quality. The rewards are set at values -7 and 1 to model the reward when transmission succeeds under a delegation action and fails, respectively. In fact, these values reflect connection costs in IEEE.802.11 MAC protocol. SAMPLE was compared to AODV and DSR protocols, in two different settings. The first one is a random network, whereas the second is a metropolitan area MANET with a set of stable links. The congestion was introduced in both scenarios. Simulation results show that SAMPLE can meet or approach many of its system optimization goals in a changing MANET environment. However, AODV and DSR perform well only when their assumptions of perfect radio links and a random network topology hold. The authors claim that this is because SAMPLE avoids generating a large number of routing packets by learning that not every dropped packet is necessarily a broken link. In addition, the retransmission of failed unicast packets in 802.11 does not change route costs for AODV and DSR, since their route costs are based on a hop-count metric, but in SAMPLE a failed unicast updates the state transition model for the network link in a way that the failing link will not be chosen in the future. Furthermore, in SAMPLE, the collaborative feedback adapts routing agent behavior to favor paths with stable links which is not possible with a discrete model of network links. Note that the on-demand and opportunistic transfer of routing information, in SAMPLE, reduces the amount of the generated controltraffic. The same MDP model used in SAMPLE protocol was applied in SNL-Q routing protocol [23]. Through simulations, the proposed protocol has shown its efficiency in comparison to AODV and DSR protocols in presence of network congestion. 4.5 Routing Policy Search via Stochastic Gradient Descent In [24], the routing problem is mapped into a POMDP where the node state is a vector of its one-hop-neighboring nodes parameters. Those parameters can be about congestion level, selfishness, remaining energy, etc. However, those parameters are usually unknown to the decision-maker node. To deal with this partial observability, a source node derives estimates about the values from past experiences with its neighboring nodes. For this purpose, the principle of WIN-OR-LOSE FAST (WoLF) [25] that consists of using a variable learning rate was adopted. Indeed, two learning rate parameters are used such that the learning rate when losing is higher than it is when losing. This enables to learn quickly while losing and slowly while winning. Furthermore, a stochastic gradient descent based algorithm that allows nodes to learn a near optimal controller was exploited. The controller f, that estimates forwarding
10
S. Chettibi and S. Chikhi
probability via node j is updated following the direction of the gradient of loss function E as follows: f
f
η
∂E ∂Θ
(4)
Where Θ is the parameters vector of node j; η is the variable step-size of the gradient; f is a function differentiable and strict increasing in Θ ; E is a loss error function that measure error between the observed and the estimated probability. When a source node needs to make decision it calculates the value of the controller for all nodes in the set of one hop neighboring toward a destination d, given the current nodes parameters estimates. Then, it selects the greedy action i.e. the node that is most likely to forward the packets with probability 1 ε and a randomly selected 1/t. node, different from the greedy choice, with probability ε where ε In the experiments, only energy and selfishness parameters were considered. Furthermore, a non-linear reward function was used. Simulation results have shown that there is a compromise between success rate and number of alive nodes when considering energy and selfishness alternatively or jointly. When energy was not considered at all, cooperative nodes run out of energy which decrease consequently the success rate and vice versa. Thus, considering selfishness and energy fairly may give the best trades-off. 4.6 RL-Based Energy-Efficient Routing To strike a balance between the contrasting objectives of maximizing the nodes lifetime and minimizing the energy consumption in an adaptive way, authors in [26] have adopted a learning module based on the first-visit ONMC method in their energy–aware algorithm. They modelized the energy-efficient path selecting problem as a MDP, where a node state encompasses information about the residual battery and the energy consumption required to forward a packet. The decision that a source node faces is which path it should select to achieve the best long-term performance. The authors define a finite action space, based on three commonly-used energy-aware routing mechanisms, namely, the minimum energy routing, the max-min routing, and the minimum cost routing. Once the source node selects an action (a path) at a given state, the following cost incurs: c s, a
P
B
B
(5)
Where B is the initial level of battery assumed to be constant for all nodes; x1, x2, x3 are weight factors all, empirically, fixed to 1; B is the battery bottleneck of path and P is energy consumption along the path . Three variants of the proposed protocol were studied, namely : BECRL with A={lb,le,lc},BERL with A={lb,le}, ECRL with A={le,lc}. Where, A denotes the action space and le, lb,lc, denote, respectively, the minimum energy, the max-min residual battery level and the minimum cost paths. These variants were compared to: Low-cost [27], MMBR, MTPR, and CMMBR [28] routing protocols. Simulation results have shown that the three variants of the RL algorithm, exhibit good routing performance in terms of successfully delivered packets ratio over all other methods with least
A Survey of Reinforcement Learning Based Routing Protocols
11
energy consumption. Furthermore, they attain higher number of alive nodes even under high mobility conditions. Particularly, among the considered RL variants, ECRL was the best one. 4.7 Multi-criteria RL - Based Routing In [29], the QOS routing is addressed as a MCDM (Multi-Criteria Decision Making) problem and routes discovery is RL-based. In effect, at each source node , computation of expected criteria vectors , , , where a, d and n respectively correspond to “application class”, “destination node” and “neighbor node”, is done in a Monte Carlo style as follows: , , ,
, , ,
, , ,
(6)
Where is the criteria vector evaluated at destination node and seen as a reward and is a constant learning rate belonging to [0,1). When an exploration packet is sent by source towards destination , router chooses a neighbor randomly. The other routers of a path choose the next hop that best satisfies the QoS requirements among its neighbors. This is done by comparing their expected multi-criteria vectors via a Russian Doll method [29]. Simulation results reported in [29] are beyond the scope of our interest because experiments were focalized on studying the performance of the Russian Doll method in comparison to the traditional MCDM methods.
5 Conclusions and Future Research Directions Routing problem in MANETs can be naturally formulated as a MDP with complete or partial state observability. Nevertheless, reinforcement learning application for routing in MANETs stills a very young research field with only few contributions. Throughout this paper, we have seen that according to if routing is QoS, energy or security constrained the mapping to the RL framework has yielded to various models and to the application of different RL-algorithms. In the light of works described in this paper, we state that nodes, in a MANET, seek to achieve a multi-objectives RL in a partially-observable multi-agent environment. Hence, we believe that the following research areas merit to be investigated: Multi-agent RL. Apart from SAMPLE routing protocol, described works in this paper have made a straightforward adaptation of single-agent RL algorithms. However, well developed studies are done in the area of multi-agent RL that integrates results from single-agent RL, game theory and direct search in the space of behaviors. We could adapt previously proposed multi-agent RL algorithms to the context of MANETs. Multi-objectives RL. Since routing performance-optimization requires balancing, generally , many conflicting goals then routing should be better addressed as a multiobjectives learning problem. Phylogenetic RL. Dealing with the routing problem in MANETs as a problem of policy search is addressed through the application of the gradient method which is
12
S. Chettibi and S. Chikhi
better ranged in the Ontogenetic-family as stated in [2]. More attention must be given for other optimization techniques. Finally, we believe that simulation results reported in this survey were expected and thus less informative. Certainly, comparative studies between different RL-based routing protocols will contribute to a better understanding of RL potentials and limitations when dealing with adaptive routing in MANETs. Acknowledgments. The authors would like to thank Mr. Laboudi for his help in enhancing the quality of this paper.
References 1. Sutton, R., Barto, A.: Reinforcement learning. MIT Press, Cambridge (1998) 2. Togelius, J., Schaul, T., Wierstra, D., Igel, C., Gomez, F., Schmidhuber, J.: Ontogenetic and phylogenetic reinforcement learning. ZeitschriftK unstlicheIntelligenz 3, 30–33 (2009) 3. Watkins, C.J.: Learning with Delayed Rewards. PhD thesis, Psychology Department, University of Cambridge, UK (1989) 4. Dowling, J., Cunningham, R., Harrington, A., Curran, E., Cahill, V.: Emergent consensus in decentralised systems using collaborative reinforcement learning. In: Babaoğlu, Ö., Jelasity, M., Montresor, A., Fetzer, C., Leonardi, S., van Moorsel, A., van Steen, M. (eds.) SELF-STAR 2004. LNCS, vol. 3460, pp. 63–80. Springer, Heidelberg (2005) 5. Williams, R.J.: Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning 8(3), 229–256 (1992) 6. Peshkin, L.: Reinforcement Learning by Policy Search. PhD thesis, Brown University (2001) 7. Johnson, D.B., Maltz, D.A.: Dynamic source routing in ad hoc wireless networks. In: Mobile Computing, ch. 5, pp. 153–181. Kluwer Academic Publishers, Dordrecht (1996) 8. Perkins, C.E., Royer, E.M.: Ad-hoc on-demand distance vector routing. In: WMCSA 1999, New Orleans, pp. 90–100 (1999) 9. Perkins, C.E., Watson, T.J.: Highly dynamic destination sequenced distance vector routing (DSDV) for mobile computers. In: ACM SIGCOMM 1994 Conf. on Communications Architectures, London (1994) 10. Jacquet, P., Muhlethaler, P., Clausen, T., Laouiti, A., Qayyum, A., Viennot, L.: Optimized link state routing protocol for ad hoc networks. In: IEEE INMIC, Pakistan (2001) 11. Toh, C.: A novel distributed routing protocol to support ad-hoc mobile computing. In: IEEE 15th Annual Int. Phoenix Conf., pp. 480–486 (1996) 12. Dube, R., Rais, C., Wang, K., Tripathi, S.: Signal stability based adaptive routing (SSA) for ad hoc mobile networks. IEEE Personal Communication 4(1), 36–45 (1997) 13. Boyan, J.A., Littman, M.L.: Packet routing in dynamically changing networks: A reinforcement learning approach. Advances In Neural Information Processing Systems 6, 671–678 (1994) 14. Sun, R., Tatsumi, S., Zhao, G.: Q-map: A novel multicast routing method in wireless ad hoc networks with multiagent reinforcement learning. In: Proc. of the IEEE Conf. on Comp., Comm., Control and Power Engineering, vol. 1, pp. 667–670 (2002) 15. Förster, A.: Machine learning techniques applied to wireless ad hoc networks: Guide and survey. In: Proc. 3rd Int. Conf. on Intelligent Sensors, Sensor Networks and Information Processing (2007)
A Survey of Reinforcement Learning Based Routing Protocols
13
16. Chang, Y.-H., Ho, T.: Mobilized ad-hoc networks: A reinforcement learning approach. In: ICAC 2004: Proceedings of the First International Conference on Autonomic Computing, pp. 240–247. IEEE Computer Society, USA (2004) 17. Tao, T., Tagashira, S., Fujita, S.: LQ-Routing Protocol for Mobile Ad-Hoc Networks. In: Proceedings of the Fourth Annual ACIS International Conference on Computer and Information Science (2005) 18. Chen, S., Nahrstedt, K.: Distributed quality-of-service routing in ad-hoc networks. IEEE Journal on Selected Areas in Communications 17(8), 1488–1505 (1999) 19. Usaha, W., Barria, J.A.: A reinforcement learning Ticket-Based Probing path discovery scheme for MANETs. Ad Hoc Networks Journal 2, 319–334 (2004) 20. Maneenil, K., Usaha, W.: Preventing malicious nodes in ad hoc networks using reinforcement learning. In: The 2nd International Symposium on Wireless Communication Systems, Italy, pp. 289–292 (2005) 21. Dewan, P., Dasgupta, P., Bhattacharya, A.: On using reputations in ad hoc networks to counter malicious nodes. In: Proceedings of Tenth International Conference on Parallel and Distributed Systems, pp. 665–672 (2004) 22. Dowling, J., Curran, E., Cunningham, R., Cahill, V.: Using feedback in collaborative reinforcement learning to adaptively optimize MANET routing. IEEE Trans. Syst. Man, Cybern. 35, 360–372 (2005) 23. Binbin, Z., Quan, L., Shouling, Z.: Using statistical network link model for routing in ad hoc networks with multi-agent reinforcement learning. In: International Conference on Advanced Computer Control, pp. 462–466 (2010) 24. Nurmi, P.: Reinforcement Learning for Routing in Ad Hoc Networks. In: Proc. 5th Intl. Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks. IEEE Computer Society, Los Alamitos (2007) 25. Bowling, M., Veloso, M.: Rational and convergent learning in stochastic games. In: Proceedings of the 17th International Joint Conference on Artificial Intelligence, pp. 1021–1026. Morgan Kaufmann, San Francisco (2001) 26. Naruephiphat, W., Usaha, W.: Balancing tradeoffs for energy-efficient routing in MANETs based on reinforcement learning. In: The IEEE 67th Vehicular Technology Conference, Singapore (2008) 27. Chang, J.H., Tassiulas, L.: Maximum lifetime routing in wireless sensor networks. IEEE/ACM Transactions on Networking 12(4), 609–619 (2004) 28. Toh, C.K.: Maximum battery life routing to support ubiquitous mobile computing in wireless ad hoc networks. IEEE Communications Magazine 39, 138–147 (2001) 29. Petrowski, A., Aissanou, F., Benyahia, I., Houcke, S.: Multicriteria reinforcement learning based on a Russian Doll method for network routing. In: 5th IEEE International Conference on Intelligent Systems, United Kingdom (2010)
Detection of Denial of Service Attack Due to Selfish Node in MANET by Mobile Agent Debdutta Barman Roy1 and Rituparna Chaki2 1
Calcutta Institute of Engineering and Management 2 West Bangal University of Technology {barmanroy.debdutta,rituchaki}@gmail.com
Abstract. Mobile Adhoc Network (MANET) is highly vulnerable to malicious attacks due to infrastructure less network environment, be deficient in centralized authorization. The fact that security is a critical problem when implementing mobile ad hoc networks (MANETs) is widely acknowledged. One of the different kinds of misbehavior a node may exhibit is selfishness. Routing protocol plays a crucial role for effective communication between mobile nodes and operates on the basic assumption that nodes are fully cooperative. Because of open structure and limited battery-based energy some nodes (i.e. selfish or malicious) may not cooperate correctly. There can be two types of selfish attacks –selfish node attack (saving own resources) and sleep deprivation (exhaust others’ resources. In this paper, we propose a new Mobile Agent Based Intrusion Detection System (IDS). The approach uses a set of Mobile Agent (MA) that can move from one node to another node within a network. This as a whole reduces network bandwidth consumption by moving the computation for data analysis to the location of the intrusion. Besides, it has been established that the proposed method also decreases the computation overhead in each node in the network. Keywords: MANET, Mobile Agent, Selfish Node, IDS.
1 Introduction A Mobile Ad Hoc Network (MANET) is a dynamically changing network without any centralized authorization and has co operative algorithm. This kind of network is well suited for the critical applications in remote places like emergency relief, military operations where no pre-deployed infrastructure exists for communication. Due to the lack of centralized authorization and volatile network topology it is difficult to detect adversary nodes [4, 5], MANETs are highly vulnerable to attacks. Lastly, we can conclude that in a MANET nodes might be battery-powered and might have very limited resources, which may make the use of heavy-weight security solutions undesirable [7, 8, 9, 10 and 11]. This paper deals with the Denial of service attack (DoS) by a selfish node; this is the most common form of attack which decreases the network performance. A selfish node does not supposed to directly attack the other nodes, but is unwilling to spend battery life, CPU cycles, or available network bandwidth to forward packets not of A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 14–23, 2011. © Springer-Verlag Berlin Heidelberg 2011
Detection of Denial of Service Attack Due to Selfish Node in MANET
15
direct interest to it. It expects other nodes to forward packets on its behalf. To save own resources there is a strong motivation for a node to deny packet forwarding to others, while at the same time using the services of other nodes to deliver own data. According to the attacking technique the selfish node can be defined in three different ways [1] SN1: These nodes take participation in the route discovery and route maintenance phases but refuses to forward data packets to save its resources. SN2: These nodes neither participate in the route discovery phase nor in dataforwarding phase. Instead they use their resource only for transmissions of their own packets. SN3: These nodes behave properly if its energy level lies between full energy-level E and certain threshold T1. They behave like node of type SN2 if energy level lies between threshold T1 and another threshold T2 and if energy level falls below T2, they behave like node of type SN1. One immediate effect of node misbehaviors and failures in wireless ad hoc networks is the node isolation problem and network partitioning due to the fact that communications between nodes are completely dependent on routing and forwarding packets [2].
Fig. 1. Node isolation due to selfish neighbors
In Figure.1, suppose node x3 is a selfish node. Here, the node u initiates a RREQ message for the destination node v. The selfish node x3 may be unwilling to broadcast the route request from u. It is also possible for x3 to forward control packets; however, the situation could be worse since u may choose x3 as the next hop and send data to it. Consequently, x3 may reject all data to be forwarded via it, and then communications between u and v cannot proceed. If all the neighbors of u behave as selfish node then u becomes an isolated node in the network.
2 Related Works Several methods proposed to defend these attacks have been studied. These can be classified into three types: reputation based scheme, credit based approach and game theoretic approach [1] [3] [6]. Reputation Based scheme In a reputation based scheme [1] watchdog and path
16
D.B. Roy and R. Chaki
rater approach the IDS overhear neighbors’ packet transmission promiscuously and notify misbehavior to the source node by sending a message. The source node collects the notifications and rates every other node to avoid unreliable nodes in finding a path. The scheme is easier to implement but it depends only on promiscuous listening that may results false identification. CONFIDANT (Cooperation of Nodes, Fairness in Dynamic Ad-hoc Networks), in this scheme the IDS performs task in a distributed ways the monitor node promiscuously observes route protocol behavior as well as packet transmission of neighbor node. The Trust manager sends ALARM messages on detection of misbehavior. The Reputation system: maintains a rating list and a blacklist for other nodes. CORE (Collaborative Reputation) approach, here the source node observes usual packet transmission and the task specific behavior of neighbor nodes and rate the node by using the positive reports from other nodes. The malicious node with bad reputation rate is isolated. But in this approach reputation of node is not changed frequently, thus the nodes temporarily suffering from bad environmental conditions are not punished severely. Credit based scheme Sprite Simple, cheat-proof, credit based system; here the node s send CAS Central Authorized Server) a receipt for every packet they forward, CAS gives credits to nodes according to the receipt. This approach is useful as it is easy to implement but the major problem is scalibility and message overhead. Ad hoc-VCG(Vickery, Clarke and Groves) scheme ,this is a two phase approach in the Route Discovery phase destination node computes needed payments for intermediate nodes and notifies it to the source node or the central bank. In the Data Transmission phase actual payment is performed .This scheme is fully depends on the report of the destination node. Game Theoretic scheme In game theoretic scheme the IDS compares node’s performance against other node based on a repeated game. This scheme is easy to implement but it needs fair comparison among nodes other wise it may falsely identify a node as adversary node.
3 Motivations The initial motivation for our work is to address limitations of current IDS systems by taking advantage of the mobile agent paradigm. Specifically, we address the following limitations of the earlier proposed IDS. False Positive Rate: The IDS reduces the False Positive rate that may arise in Reputation based scheme, which effectively increase the network performance. Scalability: The process scalability of the credit based approach or any centralized approach is much lower. By using Mobile Agent the scalibility may increase that enhance the network performance. Interdependencies: In the Credit based scheme the IDS depends on the report of the destination node that make the network not convenient that require for MANET. Centralized Authorization: Due to centralized authorization of previous IDS the IDS can not perform efficiently. In Mobile Agent based IDS the computation is done in distributed manner that increase the efficiency of the IDS.
Detection of Denial of Service Attack Due to Selfish Node in MANET
17
4 Proposed Work Our objective is to find out the malicious node that performs the DOS by selfish node in network. The assumptions regarding the proposed work are listed below The following assumptions are taken in order to design the proposed algorithm. 1. A node interacts with its 1-hop neighbors directly and with other nodes via intermediate nodes using multi-hop packet forwarding. 2. Every node has a unique id in the network, which is assigned to a new node collaboratively by existing nodes. 3. The source node generates mobile agent after a specific period of time. 4. The mobile agent moves towards forward path created using RREQ and RREP. 5. The agent calculates the packet receive and forward by a node. 6. If the agent discovers a malicious node, instead of moving forward, it sends a report to the source node. Architecture of a Mobile agent based system: From the figure 2, it is observed that the mobile agent performs three tasks. At first the mobile agent (MA) has to collect the raw data from the host machine then it computes the packet delivery ratio (Pdr) after computation it compares the resultant Pdr with the predefined one and then gives responses to the source node accordingly.
Give Response
WƌĞǀŝŽƵƐůLJ ƐƚŽƌĞĚ
EĞƚǁŽƌŬ
WĞƌĨŽƌŵ ŽŵƉƵƚĂƚŝŽŶ
dĂƐŬ ŽĨ Ă ŵŽďŝůĞĂŐĞŶƚ
DŽďŝůĞ
,ŽƐƚ DĂĐŚŝŶĞ
WŚLJƐŝĐĂů ůĂLJĞƌ
Fig. 2. Architecture of proposed Mobile Agent IDS
The Mobile Agent maintains the following table to perform the computation and comparison with threshold value Table 1. Data structure of the Mobile Agent
Source node ID
Destination Node ID
HOP count
THRESOLDPdr
18
D.B. Roy and R. Chaki
The table contains the source node id, destination node id that will be initiated by the source node. The HOP count field in the table denotes number of HOP between source node and destination node. THRESOLDPdr signifies the number of packet drop to be considered for any node in the forward path. The forward path is generated by the AODV routing protocol. The network is modeled based on the de-bruijn graph as follows: Node Sequence: The Node sequence describes a set of nodes where the link among the nodes are created in such a way that when the node n with bit sequence (a0n a1n a2n…. akn) is connected with a node m having a bit sequence (a0m a1ma2m…. akm) where 1W[CH] Then CH=id of ith node This way SRCHS concentrates on selection of stable, reliable cluster head.
4 Performance Analysis The simulation experience conducted for the performance evaluations were implemented in NS2.NS2 is a simulation tool, widely used for wireless network. To determine the efficiency of proposed protocol, we monitored four parameters: no of dominating nodes, no of non dominating nodes, load in the network and the delivery time. Load is computed by total no packets transmitted during simulation period. Figure 1 shows the no of dominating nodes as a function of number of nodes in the network. Figure 2 shows delivery time for the routing protocol as a function of load in the network.
42
D. Sinha and R. Chaki
Fig. 1. Showing no of nodes vs. no of dominating nodes
The figure1 shows, nodes vs dominating nodes graph. Here, it is found that if nodes increase, dominating nodes does not increase rapidly. From 5 to 10 nodes, 35
30
DELIVERY TIME
25
20
15
10
5
0 40
60
80
100
120
140
160
LOAD Fig. 2. Load vs Delivery time
180
200
220
SRCHS – A Stable Reliable Cluster Head Selection Protocol
43
dominating nodes increase with no of nodes in the network. After that, no of dominating nodes decrease with increase of nodes in the network. After from 20th node, it’s value again increase. Then after 22 nodes value of dominating nodes are same with increase of no of nodes in the network. So it is proved that value of dominating nodes’ number does not depend on increase of nodes in the network .After certain number of nodes it will be same. So, number of clusters also does not increase rapidly and overhead does not increase. The figure2 shows load vs delivery time graph .From this figure, it is determined, if load in the network increases delivery time does not increase rapidly. From 160 to 200 delivery time increases with the load in the network. But, from 10 to 160 load in the network, delivery time is stable in network.
5 Conclusion Our proposed algorithm is a cluster based routing protocol for adhoc network. In our method cluster creation is based on distributed manner. In this proposed algorithm cluster head selection is an important part. Most potential node in a cluster is selected as cluster head. In this algorithm, cluster head is selected using the concept of dominating and non dominating nodes. Most secured, stable node is selected as cluster head in a cluster. This way, the proposed algorithm enhances the security as well as stability of cluster. Some potential nodes among dominating nodes are selected as cluster heads. Theses nodes are eligible for intra cluster communication. This algorithm also concerns about the dynamic nature of mobile adhoc network for cluster head maintenance. In the proposed protocol routing is also done quickly. The reason behind this, routing is depended on the address of cluster heads. By failing any node in the route, its CH may make another node to forward packets. This causes error tolerance to be enhanced. If destination node, is not within cluster CH forwards packets to it's gateway node and gateway node forwards packets to cluster head of destination node. This way, intra cluster routing is also done efficiently. The performance of proposed algorithm is evaluated through simulation of network topology. Simulation demonstrates significant improvement in packet delivery ratio with load in the network. Currently, we are in the process of conducting simulation experiments for comparing this proposed protocol with other cluster based routing protocol.
References 1. Schwartz, M., Stern, T.E.: Routing Techniques used in Communication Networks. IEEE Trans. on Communications, 539–552 (April 1980) 2. Ramamoorthy, C.V., Bhide, A., Srivastava, J.: Reliable Clustering Techniques for Large, Mobile Packet Radio Networks. In: Proc. IEEE INFOCOM, pp. 218–226 (May 1987) 3. Zhang, J., Jeong, C.K., Lee, G.Y., Kim, H.J.: Cluster-based Multi-path Routing Algorithm for Multi-hop Wireless Network 4. Krishna, P., Vaidya, N.H., Chatterjee, M., Pradhan, D.K.: Cluster Based Routing Protocol. ACM SIGCOMM Computer Communication Review 27(2), 49–64 (1997)
44
D. Sinha and R. Chaki
5. Gerla, M., Tasai, J.: Multicluster, mobile, multimedia radio network. ACM-Baltzer Journal Wireless Networks 1(3), 255–256 (1997) 6. Jiang, M., Li, J., Tay, Y.C.: Cluster Based Routing Protocol(CBRP) Functional Specification Internet Draft(June 1999) (draft-ieft-manet-cbrp.txt) 7. Zheng, Z.-w., Wu, Z.-h., Lin, H.-z., Zheng, K.-g.: CRAM: An Energy Efficient Routing Algorithm for Wireless Sensor Networks. Computer and Information Sciences, 341–350 (2004) 8. Dhurandher, S.K., Singh, G.V.: Power aware cluster efficient routing in wireless ad hoc networks. In: Pal, A., Kshemkalyani, A.D., Kumar, R., Gupta, A. (eds.) IWDC 2005. LNCS, vol. 3741, pp. 281–286. Springer, Heidelberg (2005) 9. Wang, Y., Ling, T., Yang, X., Zhang, D.: Scalable and Effective Cluster Based Routing Algorithm Using Nodes‘ Location for Mobile Ad Hoc Networks. Information Technology Journal 7(7), 958–971 (2008) 10. Anitha, V.S., Sebastian, M.P.: SCAM: scenario-based clustering algorithm for mobile ad hoc networks. In: Proceedings of the First international conference on COMmunication Systems And NETworks, Bangalore, India, pp. 375–382. IEEE Press, Piscataway (2009); ISBN:978-1-4244-2912-7
Handover Latency Reduction Using Integrated Solution Scheme for Proxy Mobile IPv6 Md. Mahedi Hassan and Kuan Hoong Poo Faculty of Information Technology, Multimedia University, 63100, Cyberjaya, Malaysia {md.mahedi.hassan08,khpoo}@mmu.edu.my
Abstract. The next-generation mobile and wireless communications will be supported by an all-IP based infrastructure which requires an effective mobility management protocol to support ubiquitous network access by providing seamless handover. However, the recent explosion on the usage of mobile devices has also generated several issues in terms of performance and quality of service (QoS). Nowadays, mobile users demand high quality performance, best QoS and seamless connections that support real-time application such as audio and video streaming. This paper aims to study the impact and evaluate the mobility management protocols under micro mobility domain on link-layer and network-layer handover performance. We propose an integrated solution of network-based mobility management framework, based on Proxy Mobile IPv6, to reduce handover latency when mobile host moves to new network during handover on high speed mobility. We conducted simulations and analyze the network performances for mobile host under high speed for the proposed mobility protocols. Keywords: Seamless handover; Handover latency; Mobility protocols; Intradomain; Proxy MIPv6; NS-2.
1 Introduction In recent years, mobile and wireless communications have undergone tremendous changes due to the rapid development in wireless and communication technologies as well as the ever increasing demands by users. Nowadays, mobile end-users are constantly on the go and most of the time, they are moving from one place to another place in rapid pace. As a result, connected mobile devices are also constantly changing their points of attachment to the communication networks, such as Mobile Cellular Networks (MCN), Wireless Local Area Networks (WLAN), Wireless Personal Access Networks (WPAN), and so on. These days, most of the wireless and mobile communication networks are moving towards all IP based. These communication networks are either connected together through the Internet or through private IP core networks. In order to maintain connection, one of the main challenges faced by Mobile Host (MH) is the ability to obtain a new IP address and A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 45–56, 2011. © Springer-Verlag Berlin Heidelberg 2011
46
Md.M. Hassan and K.H. Poo
update its communication partners, while moving amongst these different wireless and mobile networks. In order to meet the above challenge, Internet Engineering Task Force (IETF) [1] designed a new standard solution for Internet mobility officially called – IPv6 mobility support and popularly named as Mobile IPv6 (MIPv6) [2]. MIPv6 is the modified version of MIPv4, that has great practicality and able to provide seamless connectivity to allow a mobile device to maintain established communication sessions whilst roaming in different parts of the Internet. When a MH is handed over from one network to another network, it changes the point of attachment from one access router (AR) to another. This is commonly known as handover which allows MH to establish a new connection with a new subnet. Handover is also defined as the process of changing between two ARs and when ARs’ point of attachment in the network changes. The point of attachment is a BS for cellular network, or an AR for WLAN. Commonly, handover can be handled in the link layer, if both the ARs are involved in the same network domain. Otherwise, a route change in the IP layer possibly will be needed the so-called network layer handover. In this case, Mobile IPv6 is a standard protocol for handling network layer handover. For IP-mobility protocols, the IP handover performance is one of the most important issues that need to be addressed. IP handover occurs when a MH changes its network point of attachment from one base station (BS) to another. Some of the major problems that may occur during handover are handover latency and packet loss which can degrade the performance and reduce quality of service. In a nutshell, handover latency is the time interval between the last data segment received through the previous access point (AP) and first data segment received through the next AP [3]. The major problem arises with handovers is the blackout period when a MH is not able to receive packets, which causes a high number of packet loss and communication disruption. Such long handover latency might disrupt ongoing communication session and some interruptions. If that change is not performed efficiently, end-to-end transmission delay, jitters and packet loss will occur and this will directly impact and disrupt applications perceived quality of services. For example, handovers that might reach hundreds of milliseconds would not be acceptable for delay-sensitive applications like video streaming and network gaming [3]. Currently, there are several mobility protocols which have been proposed in order to alleviate such performance limitations. One of which is the enhanced version of terminal independent Mobile IP (eTIMIP) [4], which is a kind of mobility management protocol. eTIMIP enhances the terminal independent Mobile IP (TIMIP) by reducing the amount of latency in IP layer mobility management messages exchanged between an MH and its peer entities, and the amount of signaling over the global Internet when an MH traverses within a defined local domain. TIMIP [4] is an example of IP based micro-mobility protocol that allows MH with legacy IP stacks to roam within an IP domain and doesn’t require changes to the IP protocol stack of MH in a micro mobility domain. Compared to the above mobility protocols, Proxy Mobile IPv6 (PMIPv6) [5] defines a domain in which the MH can roam without being aware of any layer 3 (L3) movement since it will always receive the same network prefix in the Router Advertisement (RA). The PMIPv6 specification defines a protocol to support Network-based Localized
Handover Latency Reduction Using Integrated Solution Scheme
47
Mobility Management (NETLMM) [5] where the MH is not involved in the signaling. This new approach is motivated by the cost to modify the protocol stack of all devices to support Mobile IP and potentially its extensions and to support handover mechanisms similar to the ones used in 3GPP/3GPP2 cellular networks. We make use of Network Simulator, ns-2 [6] in this paper to simulate, examine and compare the performances of eTIMIP, TIMIP, PMIPv6 as well as our proposed integrated solution of PMIPv6 with MIH and Neighbor Discovery (PMIPv6-MIH) in intra-domain traffic with high speed MH. We compare the handover latency and packet delivery throughput of transmission control protocol (TCP) and user datagram protocol (UDP) for eTIMIP, TIMIP, PMIPv6 and our proposed integrated solution of PMIPv6-MIH in intra-domain traffic. The rest of this paper is structured as follows: Section 2 briefly explain related research works on the mobility protocols. Section 3 explains overview of media independent handover. Section 4 briefly describes the propose solution scheme. Section 5 shows simulation results of UDP and TCP flow under intra-domain traffic. Finally, Section 6 we conclude the paper and provide possible future works.
2 Existing Mobility Protocols For mobility protocols, there are several protocols to reduce handover latency and packet loss, such as the Session Initiation Protocol (SIP) [7] and the Stream Control Transmission Protocol (SCTP) [8]. Both protocols focus on mobility management on an end-to-end basis but they don’t have the potential to achieve short handover latency in network layer. The communication sessions in these protocols are initiated and maintained through servers. The behavior of these protocols is similar to the standard Mobile IP scheme during handovers. However, there are some enhanced Mobile IP schemes that able to reduce the handover latency such as PMIPv6 and CIMS, (Columbia IP Micro-Mobility Suite) [9]. 2.1 Micro Mobility Protocols Micro mobility protocols work within an administrative domain which is to ensure that packets are arriving from the internet and addressed to the MHs that forward to the appropriate wireless AP in an efficient manner. It is also called intra-domain traffic [10]. Under the CIMS (Columbia IP Micro-Mobility Suite) project, several micro mobility protocols have been proposed such as –Handoff-Aware Wireless Access Internet Infrastructure (Hawaii) and Cellular IP (CIP). The CIMS is an extension that offers micro-mobility support. CIMS implements HMIP (Hierarchical Mobile IP) and two micro-mobility protocols for CIP and Hawaii. The CIMS project is mainly focused on intra-domain handover and uses the basic idea of Mobile IP for inter-domain handover. Subsequently, the CIMS project was enhanced by Pedro et. al. [9] which included the original implementation of TIMIP protocol, and the extended version of TIMIP protocol such as eTIMIP as well as the implementation of CIP, HAWAII, and HMIP protocols. The proposed eTIMIP protocol which is a mobility solution protocol that
48
Md.M. Hassan and K.H. Poo
provides both network and terminal independent mobile architectures based on the usage of overlay micro-mobility architecture. 2.2 Enhanced version of Terminal Independent Mobile IP (eTIMIP) The physical network and overlay network are two complementary networks that are organized in the architecture of eTIMIP. Both networks are separated in the mobile routing from the traditional intra-domain routing which also known as fixed routing. Generally, the physical network can have any possible topology, where it is managed by any specialized fixed routing protocol. The overlay network is used to perform the mobile routing, where it selects routers which support the eTIMIP agents, in which will be organized in a logical tree that supports multiple points of attachment to the external of the domain. 2.3 Proxy Mobile IPv6 (PMIPv6) PMIPv6 is designed to provide an effective network-based mobility management protocol for next generation wireless networks that main provides support to a MH in a topologically localized domain. In general terms, PMIPv6 extends MIPv6 signaling messages and reuse the functionality of HA to support mobility for MH without host involvement. In the network, mobility entities are introduced to track the movement of MH, initiate mobility signaling on behalf of MH and setup the routing state required. The core functional entities in PMIPv6 are the Mobile Access Gateway (MAG) and Local Mobility Anchor (LMA). Typically, MAG runs on the AR. The main role of the MAG is to perform the detection of the MH’s movements and initiate mobility-related signaling with the MH’s LMA on behalf of the MH. In addition, the MAG establishes a tunnel with the LMA for forwarding the data packets destined to MH and emulates the MH’s home network on the access network for each MH. On the other hand, LMA is similar to the HA in MIPv6 but it is the HA of a MH in a PMIPv6 domain. The main role of the LMA is to manage the location of a MH while it moves around within a PMIPv6 domain, and it also includes a binding cache entry for each currently registered MH and also allocates a Home Network Prefix (HNP) to a MH. Since the PMIPv6 was only designed to provide local mobility management, it still suffers from a lengthy handover latency and packet loss during the handover process when MH moves to a new network or different technology with a very high speed. Even more, since detecting MHs' detachment and attachment events remains difficult in many wireless networks, increase handover latency and in-fly packets will certainly be dropped at new MAG (n-MAG).
3 Overview of Media Independent Handover The working group of IEEE 802.21 [11] developed a standard specification, called Media Independent Handover (MIH), which defines extensible media access independent mechanisms that facilitates handover optimization between heterogeneous IEEE 802 systems such as handover of IP sessions from one layer 2 (L2) access technology to another. The MIH services introduce various signaling, particularly for handover initiation and preparation and to help enhance the handover performance.
Handover Latency Reduction Using Integrated Solution Scheme
49
Basically, IEEE 802.21 introduces three different types of communications with different associated semantics, the so-called MIH services: Media Independent Event Service (MIES), Media Independent Command Service (MICS) and Media Independent Information Service (MIIS). MIES introduces event services that provide event classification, event filtering and event reporting corresponding to dynamic changes in link characteristics, links status, and link quality. It also helps to notify the MIH users (MIHU) such as PMIPv6 about events happening at the lower layers like link down, link up, link going down, link parameters report and link detected etc and essentially work as L2 triggers. MICS provides the command services that enable the MIH users to manage and control link behavior relevant to handovers and mobility, such as force change or handover of an interface. The commands generally carry the upper layers like L3 decisions to the lower layers like L2 on the local device entity or at the remote entity. There are several examples of MICS commands, such as MIH scan, MIH configure, MIH handover initiate, MIH Handover prepare and MIH handover complete. MIH provides information about the characteristics and services through a MIIS which enables effective handover decisions and system access based on the information about all networks from any single L2 networks. MIIS provides registered MIH users with the knowledgebase of the network and information elements and corresponding query-response mechanisms for the transfer of information. By utilizing these services, the MIH users are able to enhance handover performance such as through informed early decisions and signaling. MIIS are classified into three groups, namely general or access network specific information, Point of Attachment specific information and vendor specific information.
4 Proposed Solution Scheme In response to the PMIPv6 problems mentioned in Section 2, we proposed solution scheme that provides an integrated solution with integrate the analysis of handover latency introduced by PMIPv6 with the seamless handover solution used by MIH as well as the Neighbor Discovery message of IPv6 to reduce handover latency and packet loss on network layer at n-MAG to avoid the on-the-fly packet loss during the handover process. Figure 1 represents the proposed integrated solution of PMIPv6-MIH.
Fig. 1. Proposed Integrated Solution
Fig. 2. Integrated solution architecture of PMIPv6
50
Md.M. Hassan and K.H. Poo
Figure 2 presents the key functionality is provided by Media Independent Handover (MIH) which is communication among the various wireless layers and the IP layer. The working group of IEEE 802.21 introduces a Media Independent Handover Function (MIHF) that is situated in the protocol stack between the wireless access technologies at lower layer and IP at upper layer. It also provides the services to the L3 and L2 through well defined Service Access Points (SAPs) [11]. 4.1 Neighbor Discovery Neighbor Discovery (ND) enables the network discovery and selection process by sending network information to the neighbor MAG before handover that can helps to eliminate the need for MAG to acquire the MH-profile from the policy server/AAA whenever a MH performs handover between two networks in micro mobility domain. It avoids the packet loss of on-the-fly packet which is routed between the LMA and previous MAG (p-MAG). This network information could include information about router discovery, parameter discovery, MH-profile which contains the MH-Identifier, MH home network prefix, LMA address (LMAA), MIH handover messages etc., of nearby network links. 4.2 Analysis of Handover Latency and Assumptions The overall handover latency consists of the L2 and L3 operations. The handover latency is consequent on the processing time involved in each step of handover procedure on each layer. The handover latency (Lseamless) can be expressed as: (1)) where LL3 represents the network layer as example switching latency and LL2 represents link layer as example switching time. On L3, the handover latency is affected by IP connectivity latency. The IP connectivity latency results from the time for movement detection (MD), configure a new CoA (care-of-address), Duplicate Address Detection (DAD) and binding registration. Therefore, L3 can be denoted as follows: (2) where Tmove represents the time required for the MH to receive beacons from n-MAG, after disconnecting from the p-MAG. In order to estimate the movement detection delay, based on the assumptions of mobility management protocols that the times taken for MD are RS and RA messages as follow (3)
Tconf represents the time that taken for new CoA configuration. Treg represents the time elapsed between the sending of the BU from the MH/MAG to the MAP/LMA and the arrival/transmission of the first packet through the n-MAG. Binding registration is the sum of the round trip time between MH/MAG and MAP/LMA and the processing time as follows:
Handover Latency Reduction Using Integrated Solution Scheme
51
(4) TDAD represents the time required to recognize the uniqueness of an IPv6 address. Once the MH discovers a new router and creates a new CoA it tries to find out if the particular address is unique. This process is called DAD and it is a significant part of the whole IPv6 process. As simplification of (2), (3) and (4) equations, it can be expressed as: (5) On L2, MH has to perform three operations during the IEEE 802.11 handover procedure such as scanning (Tscan), authentication (TAAA) and re-association (Tre-ass). Handover latency at L2 can be denoted as follows: (6) Tscan represents the time that taken the MH performs a channel scanning to find the potential APs to associate with. When MH detects link deterioration, it starts scanning on each channel finding the best channel based on the Received Signal Strength Indicator (RSSI) value. TAAA represents the time taken for authentication procedure that depends on the type of authentication in use. The authentication time is round trip time between MH and AP. While Tre-ass represents the time needed for re-association consists of reassociation request and reply message exchange between MH and AP if authentication operation is successful.
Fig. 3. An Analytical Model of Integrated solution of PMIPv6-MIH
The following notations are depicted in Figure 3 for integrated solution of PMIPv6-MIH. • • • • • •
The delay between the MH and AP is tpm, which is required the time for a packet send between the MH and AP through a wireless link. The delay between the AP and n-MAG is tma, which is the time between the AP and the n-MAG connected to the AP. The delay between the n-MAG and LMA is tag. The delay between the LMA and Corresponding Node (CN) is tca. The delay between the n-MAG and CN is tcm, which is the time required for a packet to be sent between the n-MAG and the CN. The delay between the mobility agents and AAA is ta.
52
Md.M. Hassan and K.H. Poo
As shown in figure 3, we proposed integrated solution of PMIPv6 with MIH and ND to reduce handover latency as the time taken for scanning by informing the MH about the channel information of next APs and use ND message of IPv6 to reduce handover delay and packet loss on network layer at n-MAG to avoid the on-the-fly packet loss during the handover process. During the IEEE802.11 handover procedure the MH performs scanning on the certain number of channels to find the potential APs to associate with. By informing the MH about the channel information of next APs can significantly reduce the scanning time. However, the scanning time also depends on the type of scanning is used. There are two types of scanning which are defined as active and passive. In active scan mode, MH sends probe request and receives probe response if any AP is available on certain channel. While in passive scan mode, each MHs listens the channel for possible beacon messages which are periodically generated by APs. The handover delay in active scan mode is usually less than in passive scan mode. The operation of passive scan mode depends on the period of beacon generation interval. Therefore, this can provide better battery saving than active scan mode of operation. As in L2 trigger, the p-MAG has already authenticated the MH and sends the MH's profile which contains MH-Identifier to the n-MAG through the ND message since the MH is already in the PMIPv6 domain and receiving as well as sending information to CN before the handover. Hence, the authentication delay is eliminated during actual handover. Thus, the L2 handover delay can be expressed as: 2
(7)
As the parts of L3 handover delay that should be taken into consideration in PMIPv6. Since we proposed the integrated solution of PMIPv6 with MIH services and ND, the number of handover operations should not be considered for overall handover latency. As a result, L3 handover delay is considered only two things in integrated solution of PMIPv6-MIH in a micro mobility domain. o
o
When MH attaches to the n-MAG and delivers event notification of MIH_Link_up indication, n-MAG sends a PBU message to the LMA for updating the lifetime entry in the binding cache table of the LMA and triggering transmission of buffer data for the MH RA message
Therefore, the overall handover delay at L3 can be expressed as: 3 Based on Analytical model: 2
(8)
Seamless Handover Latency of integrated solution of PMIPv6 with MIH can be expressed as:
Handover Latency Reduction Using Integrated Solution Scheme
53
5 Simulation Experiment and Results In order to examine, evaluate and compare the impact on intra-domain handover performance, simulations were performed to compare and evaluate micro mobility protocols by using the ns-2 [6]. For the simulations, two important performance indicators are measured which are the throughput for packet delivery and handover latency. In order to obtain reasonable results, we measure the performance for micro mobility protocol in intra-domain traffic for both TCP and UDP packet flow. 5.1 Simulation Setup The simulation scenario setup is implemented as a network-based mobility management solution in the simulation of mobility across overlapping wireless access networks in micro mobility domain. The proposed integrated solution scenario setup is the same as the PMIPv6 but further incorporates MIH functionality in the MH and the MAGs. Thus, the simulation setup scenario is as shown in figure 4 below:
Fig. 4. Simulation Scenario Setup of proposed integrated solution of PMIPv6-MIH
In the above simulation scenario, the p-MAG and n-MAG are in separate subnets. The two MAGs have both L2 and L3 capabilities that handles handovers. The router is interconnected to the LMA by a series of agents that are organized in a hierarchical tree structure of point-to-point wired links. The packet flow of CBR and FTP are simulated and transmitted from the CN to the MH using UDP and TCP. The link delay between the CN and the LMA is set at 10ms while the bandwidth is set at 100Mb. The link delay between the LMA and the respective MAGs is set at 1ms. The CBR and FTP packet size is set at 1000 and 1040 bytes while the interval between successive packets is fixed at 0.001 seconds. 5.2 Simulation Results Simulation results for intra-domain traffics are obtained as follows:
54
Md.M. Hassan and K.H. Poo
Fig. 5. Handover Latency of UDP Flow in micro mobility domain
Fig. 7. Throughput (Mbps) of UDP Flow in micro mobility domain
Fig. 6. Handover Latency of TCP Flow in micro mobility domain
Fig. 8. Throughput (Mbps) of TCP Flow in micro mobility domain
In above results, it is observed that UDP and TCP performance of eTIMIP and TIMIP increased the handover latency during the MH moves to new network in micro mobility domain. It also noted from the simulation results that performance of throughput also shown degradation. This is due to the fact that, when MH moves away from one network to another in micro mobility domain with high speed mobility, there are lots of operations to perform between the changes of network, such as configuring new CoA, DAD operation, binding registration and MD. In comparison to PMIPv6, it does not require CoA and DAD as MH is already roaming in the PMIPv6 domain. Once the MH has entered and is roaming inside the PMIPv6 domain, CoA is not relevant since according to the PMIPv6 specification, the MH continues to use the same address configuration. The operation of a DAD is required for a link-local address since address collision is possible between MH, MAG and all MH’s attached to the same MAG. The DAD operation may significantly increase handover latency and is a very time consuming procedure. As DAD requires around one second (or even much than one sec.), PMIPv6 introduce a per-MH prefix model in which every MH is assigned a unique HNP. This approach may guarantee address uniqueness. But still PMIPv6 suffers from a lengthy handover latency and packet loss during the handover process when MH speed is high. To overcome these problems, we proposed integrated solution scheme for PMIPv6 that can send the MHprofile to the n-MAG through ND message before handover on L3 and also reduce the time on L2 scanning by informing the MH about the channel information of next APs using MIH services. Based on the proposed solution scheme, the result of handover latency and throughput are better than other mobility protocols. The reason of reduce handover latency and improve throughput in micro mobility domain as below:
Handover Latency Reduction Using Integrated Solution Scheme
55
¾ The time required to obtain MH profile information can be omitted since nMAG performs this information retrieval prior to MH’s actual attachment. ¾ As the specification of PMIPv6, the time needed to obtain the DAD operation and configure new CoA can also be non-appreciable since n-MAG performs a pre-DAD procedure like assigning a unique HNP during available resource negotiation with p-MAG and the MH continues to use the same address configuration. ¾ The time required to obtain mobility-related signaling massage exchange during pre-registration may not be considered since this negotiation is established before MH attachment. Since the MH is already pre-registered and there is no need to confirm at the n-MAG, therefore the last Proxy Binding Acknowledgement (PBA) message send from the LMA may not be considered.
6 Conclusion In this paper, we conducted simulations to evaluate, compare and examine the mobility protocols under intra-domain approaches. As for performance, we compared performance indicators such as handover latency and throughput for mobility protocols to the proposed integrated solution. Based on our simulation results obtained, the integrated solution of PMIPv6-MIH demonstrates better performance as compared to other mobility protocols. As for the future work, we would like to improve the handover latency, and evaluate the performance of the proposed PMIPv6-MIH on real-time applications e.g. video streaming.
References 1. Johnson, D., Perkins, C., Arkko, J.: IP Mobility Support in IPv6. RFC 3775 (June 2004), http://www.ietf.org/rfc/rfc3775 2. Perkins, C.E.: Mobile Networking Through Mobile IP. IEEE Internet Computing 2(1), 58–69 (2002) 3. Yaakob, N., Anwar, F., Suryady, Z., Abdalla, A.H.: Investigating Mobile Motion Prediction in Supporting Seamless Handover for High Speed Mobile Node. In: International Conference on Computer and Communication Engineering, pp. 1260–1263 (2008) 4. Estrela, P.V., Vazao, T.M., Nunes, M.S.: Design and evaluation of eTIMIP – an overlay micro-mobility architecture based on TIMIP. In: International Conference on Wireless and Mobile Communications (ICWMC 2006), pp. 60–67 (2006) 5. Kong, K., Lee, W., Han, Y., Shin, M., You, H.: Mobility management for all-IP mobile networks: mobile IPv6 vs. proxy mobile IPv6. In: International Conference on Wireless Communications, pp. 36–45 (2008) 6. NS-2 home page, http://www.isi.edu/nsnam/ns 7. Kwon, T.T., Gerla, M., Das, S.: Mobility management for VoIP service: Mobile IP vs. SIP. IEEE Wireless Communications 9(5), 66–75 (2002)
56
Md.M. Hassan and K.H. Poo
8. Jung, J.-W., Kim, Y.-K., Kahng, H.-K.: SCTP mobility highly coupled with mobile IP. In: de Souza, J.N., Dini, P., Lorenz, P. (eds.) ICT 2004. LNCS, vol. 3124, pp. 671–677. Springer, Heidelberg (2004) 9. Columbia IP micro-mobility suite (CIMS), http://tagus.inesc-id.pt/~pestrela/ns2/mobility.html 10. Abdalla Hashim, A.H., Ridzuan, F., Rusli, N.: Evaluation of Handover Latency in IntraDomain Mobility. In: The Fourth World Enformatika Conference (2005) 11. Taniuchi, K., Ohba, Y., Fajardo, V.: IEEE 802.21: Media independent handover. Features, applicability, and realization 47(1), 112–120 (2009)
Modeling and Simulation Analysis of QPSK System with Channel Coding T.P. Surekha1, T. Ananthapadmanabha2, C. Puttamadappa3, and A.P. Suma4 1
Assistant. Professor., Dept. of E&CE, Vidyavardhaka College of Engineering, Mysore and Research Scholar at NIE, Mysore, India
[email protected] 2 Professor, Dept. of E&EE, National Institute of Engineering, Mysore, India And Honorary secretary of IEI, Mysore local center, India
[email protected] 3 Professor and Head, Dept. of E&CE, S.J.B. Institute of Technology, Kengeri, Bangalore north, India
[email protected] 4 Persuing her masters at SJCE, Dept. of Instrumentation and Technology, Mysore
Abstract. The most appropriate modulation and channel coding for a Very Small Aperture Terminal (VSAT) system is Quadrature Phase Shift Keying (QPSK). The Channel can be wire or wireless voice and data applications due to its flexible system architecture. In this paper, a Simulink based QPSK system is simulated to study the characteristic performance analysis of Additive White Gaussian Noise (AWGN) channel. Simulation study helps to visualize eyediagram and Root Raised cosine (RRC) Filter with scatter plot. The error rate is calculated by comparing a transmitted data stream with a receive data stream with the help of delay introduction. Characteristic performance analysis is done by comparing the un-coded data with coded data using two different models. The Bit Error Rate (BER) curve for a Communication system illustrates the relationship between power in the transmitted signal in terms of signal to noise ratio (SNR) and the resulting BER for the system. Keywords: AWGN, BER, Channel modeling, E SNR, VSAT.
b
/ No, QPSK, RRC filter,
1 Introduction
T
he choice of the modulation technique for a given communication system strongly depends on the nature of the characteristics of the medium on which it has to operate. QPSK is the most popular choice of modulation technique for use in satellite communication links carrying digital data. All digital links are designed using specific symbol rate, and specific filters that minimize Inter symbol interferences (ISI). A symbol in a baseband link is a pulse of current or voltage. VSAT stands for very small aperture terminal, used in commercial satellite communication‘s system. The VSAT earth station consists of two basic components, an outdoor unit (OUT) and A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 57–64, 2011. © Springer-Verlag Berlin Heidelberg 2011
58
T.P. Surekha et al.
an Indoor Unit (IDU). IDU consists of QPSK modem equipment. In a Satellite link, Modulation and Channel coding is an important key consideration in determining the efficient and error-free transfer of information over a communication channel. In choosing the most appropriate modulation for a VSAT system, ease of implementation is a major factor, since VSATs are very cost-sensitive. The most common forms of modulation used in VSAT system are Quadrature Phase shift Keying (QPSK). The purpose of this paper is to illustrate some important aspects on analysis and simulations of QPSK system operating over an Additive White Gaussian Noise (AWGN) channel. All the modeling and simulation is carried out using Simulink. In the simulation model, Bit Error rates (BER) of QPSK system versus the Eb /No the digital signal to noise (SNR) are used to evaluate the system performance analysis. The basic description of QPSK system is as shown in Fig.1. It consists of Data source, QPSK Transmitter, Channel, QPSK Receiver and a Data sink.
Data source
QPSK Transmitter
Channel
QPSK Receiver
Data sink
Fig. 1. The basic QPSK system
1.1 Data Source The original binary data to be transmitted is first converted into bit stream. In QPSK system, input signal is frame-based signal representation, by formatting data as an Mby-1 column vector, where M represents the number of samples per frame and element of the vector corresponds to values at the sample time. Thus data signal is a Bernoulli Binary Generator, will generate an output bit sequence. In – frame based processing, the blocks operate on each column (channel) of the incoming data. A multichannel frame based signal is represented by an M-by-N matrix. Usually, the number of possible signals is M= 2n , where n is an integer. 1.2 QPSK Transmitter The QPSK transmitter converts the bits into integers or symbols and applies to baseband modulation and further followed by optional pulse shaping. The result is a pass band signal which can be transmitted over a physical channel. The modulation process can be viewed as a complex modulation scheme using a scatter diagram. The scatter diagram allows us to visualize the real and the imaginary (in-phase and quadrature) component of the complex signal. Pulse shaping is an important consideration in the design of a system. The pulse shape filter must make efficient use of bandwidth and also have limited duration in time. A pulse too wide in time will overlap into adjacent symbol periods and cause inter symbol interference (ISI). This
Modeling and Simulation Analysis of QPSK System with Channel Coding
59
filtering can be performed by using Root-Raised cosine (RRC) filters. The eye diagram allows us to understand the time domain characteristic of a signal and its susceptibility to symbol timing error. 1.3 Channel Coding Channel coding is an important technique to detect and correct errors that occur when messages are transmitted in a digital communication system. Channel coding can take the form of a block code or a convolutional code. Convolutional codes create a trellis structure, a decision tree that traces all possible sequences of codeword’s. To decode this coding scheme, the Viterbi decode is typically implemented. Communication channels introduce noise, fading, interference, and other distortions to the transmitted signals. Several different channels are possible. The one being used here is AWGN channel, It is assumed that while passing electromagnetic waves through air or other mediums, there is an additive noise introduced to the transmission. Thus channel simply adds white Gaussian noise to the signals as shown in Fig.2. White noise
From transmitter
To receiver
Fig. 2. AWGN channel
1.4 QPSK Receiver The receiver is the most complex part in the system. It performs the reverse process of the transmitter. Receiver block takes the output from the channel, and filters out the signals by using RRC filter and demodulates the QPSK signals and finally convert it into either bit converter or viterbi decoder (for coded). 1.5 Data Sink The Error Rate Calculation block compares a transmitted data stream with a receive data stream to calculate the error rate of a system. It also outputs the number of error events that have occurred, and the total number of bits or symbols compared.
2 Implementation 2.1 Methodology Modeling and simulating of coded QPSK system is implemented in this paper. The bits are mapped onto corresponding QPSK symbols using Gray coding, as shown in Fig. 3.
60
T.P. Surekha et al.
Q B=01
A=11
I C= 00
D=10
Fig. 3. Constellation diagram for QPSK system
The implementation of QPSK is of higher order PSK. Writing the symbols in the
Si (t) =
2
Cos 2
2
1
0≤ t ≤ Ts Where
i = 1,2,3,4--------- (1)
This yields the four phases , , , as needed. For the QPSK signal set, the four signals in the set can be expressed in terms of the basis signals as 2
S qpsk (t) = √
1
ф
2
√
1
ф
Where i = 1,2,3,4--------- (2) Ф1 (t) and ф2 (t) are the basis functions defined by ф (t)= 2 T Cos (2πf t) ф (t)=
2
T Sin (2πf t)
0≤ t ≤ Ts
(3)
0≤ t ≤ Ts
(4)
The first basis function is used as the in-phase component of the signal and the second as the quadrature component of the signal. The average probability of bit error in the additive white Gaussian noise (AWGN) channel is obtained as
Pe, QPSK =
Q
2E
N
(5)
Where Es is the energy of the signal given by Es = 2Eb, and No is the noise. 2.2 Simulation Model Simulink, developed by the Math works, is an tool for multi-domain simulation and Model-based Design for dynamic and Communication systems. Communication Block set of Simulink is helpful in simulating the modeling. The base-band simulation model of coded QPSK is as shown in Fig.4.
Modeling and Simulation Analysis of QPSK System with Channel Coding
61
QPSK Specifications: Up-sample Factor = 8 Pulse shaping Filter α =0.25 Group Delay = 4
Fig. 4. The QPSK simulation model
All signal sources in the signal processing and communication‘s can generate frame based data. In this work, the signal is frame based and samples are propagated through a model and multiple samples are processed in batches. Frame – based processing takes advantage of Simulink matrix processing capabilities to reduce overhead. Complex modulation scheme are best viewed using a scatter diagram. The scatter diagram allows us to visualize the real and imaginary (in-phase and quadrature) component of the complex signal. By doing so, the phase and amplitude distortion caused by pulse shaping channel or other impairment is revealed. Thus Fig. 5. shows the Scatter plot of QPSK modulation. An Eye diagram is a convenient way to visualize a shaped signal in the time domain, which indicates that the ‘eye’ is most widely opened, and use that point as the decision point when de-mapping a demodulated signal to recover a digital message as shown in Fig. 6. Using the Root-raised Cosine (RRC) filters at the transmitter, a slight amount of phase and magnitude distortion can be seen at the output of the transmitting filter. To verify that the model was built properly, Error rate Calculation block compares a transmitted data stream with a receive data stream to calculate the error rate of a system. It also outputs the number of error events that have occurred, and the total number of bits or symbols compared. The block can output the error statistics as a variable in the displayed port. 2.3 System Analysis Characterizing the performance of a communication system under noisy conditions is an important part of the design process. Noise, interference, fading, and other types of distortion affecting the transmitted signal can cause incorrect decisions to be made by
62
T.P. Surekha et al.
Fig. 5. Scatter plot of QPSK system
Fig. 6. Eye diagram of QPSK system
the receiver, resulting in bit errors. The ratio of bit errors to received bits is called the bit error rate (BER). The BER curve illustrates the relationship between power in the transmitted signal in terms of signal-to-noise ratio (SNR) and the resulting BER for the system. By analyzing the BER curve for a given system, we can find the minimum SNR that is required to achieve a particular BER. Thus bit error rate is computed by simulating the QPSK system and comparing the input with the resulting output sequence without channel coding as shown in Fig.7 and with channel coding as shown in Fig.8.Which Performs simulation for a range of SNR value results in the BER curve.
Modeling and Simulation Analysis of QPSK System with Channel Coding
Fig. 7. Bit error rate as a function of E b /No for un-coded QPSK curve
Fig. 8. Bit error rate as a function of Eb /No Compared with coded And un-coded results
63
64
T.P. Surekha et al.
3 Results and Conclusion Bit error rate for a given QPSK system is computed and compared by using two different models. First one being un-coded model and second being coded model. One method of computing the bit error rate of a communication system is to simulate the system and compare the input with the resulting output sequences. Characteristic performance analysis of such simulations for a range of SNR values results in the BER curve. The first method being un-coded model, varying E b /No ranges from 1: 8. The second method being coded model, uses E b /No range of 1:5. Thus by using coded model of QPSK system, Fig.8 gives a final comparison of two BER curves. When designing a system with a specified BER of 1e-6, we can simulate transmission of 1e8 bits for each point on the BER curve where the width of confidence interval can be 95% .The purpose of this work is to use QPSK modulation technique in VSAT system for data transmission from one point to the other using wireless concept which is also an example of IEEE 802.16. BER Tool also allows us to visualize plots of multiple BER curves on a single axes. Which is taken as future work.
Acknowledgement The authors are very grateful to the Management of Vidya vardhaka College of Engineering, Mysore, Karnataka India, The National Institute of Engineering, Mysore, Karnataka, India..S J B I T,Kengeri, Bangalore, Karnataka, India. For their constant encouragement, and Motivation during their work.
References [1] Li, X.: Simulink – based Simulation of quadrature Amplitude Modulation (QAM) System. In: Proceedings of the 2008 IAJC – IJME International Conference (2008) [2] Sukla, T., Jain, D., Gautham, S.: Implementation of Digital QPSK modulator by using VHDL/MATLAB. International Journal of Engineering and Technology 2(9) [3] Pratt, T., Bostian, C., Allnutt, J.: Satellite Communication, 2nd edn. John Wiley and Sons, Chichester [4] Rappaport, T.S.: Wireless Communications, Principles and Practice, 2nd edn. Prentice – Hall of India Private Limited [5] Sharma, S.: Wireless and Cellular Communications, 2nd edn. S.K. Kataria and Sons Katson Books [6] Elbert, B., Schiff, M.: Simulating the performance of Communication Links with Satellite Transponders. Application Technology Strategy, Inc., http://www.goggle.com
Combating Sybil Attacks in Vehicular Ad Hoc Networks Khaled Mohamed Rabieh1 and Marianne Amir Azer2 1
Root Certification Authority Department, ITIDA, Egypt
[email protected] 2 School of Communications and Information Technology, Nile University, Egypt Computer and Systems Department, National Telecommunication Institute, Egypt
[email protected] Abstract. Vehicular Ad Hoc Networks (VANETs) are considered as a promising approach for facilitating road safety, traffic management, and infotainment dissemination for drivers and passengers. However, they are subject to an attack that has a severe impact on their security. This attack is called the Sybil attack, and it is considered as one of the most serious attacks to VANETs, and a threat to lives of drivers and passengers. In this paper, we propose a detection scheme for the Sybil attack. The idea is based on public key cryptography and aims to ensure privacy preservation, confidentiality, and nonrepudiation. In addition, we suggest a scalable security and privacy solution using short-lived and authenticated certificates that must be issued from the national certification authority in order to guarantee trust among vehicles. Keywords: Digital envelope, security, Sybil attack, vehicular ad-hoc networks.
1 Introduction Wireless Networks have a wide range of applications that is why they have become an essential part of our daily life. Amongst wireless networks, VANETs have many applications such as managing traffic and providing safety for vehicles. In VANETs, every vehicle communicates with other vehicles and with roadside infrastructures as well. These networks are mainly used for informing vehicles in case of emergencies such as car accidents, urgent breaking or traffic jam. This is done by broadcasting safety messages to warn other vehicles. As those safety messages have a substantial role in these networks, they should be sent from credible transmitters and contain proper and unaltered information. Despite the importance of location and identity privacy in such networks together with the need to protect vehicles from being tracked, it is necessary to identify the message sender in order to prevent repudiation in cases of accidents and crimes. A serious problem arises when a malicious vehicle is to launch a Sybil attack by holding the identities of multiple vehicles and enforcing false data. If benign entities are unable to recognize the Sybil attack, they will believe the false information, and base their decisions on it. Hence, addressing this problem is crucial to practical vehicular network systems. Fig. 1 illustrates the Sybil attack problem in which some or malicious nodes cooperate to deceive the other honest nodes, or in other scenarios of this attack, a node illegitimately claims multiple A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 65–72, 2011. © Springer-Verlag Berlin Heidelberg 2011
66
K.M. Rabieh and M.A. Azer
Fig. 1. Sybil attack in which a Sybil node pretends to have several identities or a group of malicious nodes cooperate together to affect the overall network decision, the bold black lines show the false data and false information between honest and Sybil nodes [2]
identities. A Sybil attack may be prevented by requiring vehicles to include a unique identity in transmitted packets. However, privacy is recognized as one of the most important attributes of a VANET, and cannot be compromised at any time [1]. It follows that such a solution will compromise the privacy of vehicles as an attacker will be able to identify their position based on the packets transmitted. Therefore, Sybil attacks need to be detected while preserving the vehicles’ privacy. In this paper, we propose a solution based on Public Key Infrastructure (PKI) for preventing and detecting Sybil attacks. The remainder of this paper is organized as follows. Section 2 describes the work done in order to solve the Sybil attack’s problem and the pros and cons of each solution. Section 3 describes our detailed proposed solution, and finally we conclude this paper in section 4.
2 Related Work Due to the danger inherent in the Sybil attack; its detection in VANETs has recently attracted much attention. Douceur in [3] introduced the puzzle solution that tests the computing resource of nodes to detect Sybil attack. But in [4], J. Newsome et al. proved that Douceur’s method is not suitable for VANETs because an attacker can easily have a stronger computing ability than the honest node. Newsome et al. improved the method by assuming that each vehicular node can only have one wireless communication module and only occupy one channel resource at a time. But the special radio modules can bring hazards during implementation. One of the most famous solutions was proposed in [4]. It uses the resource testing idea which relies on the assumption that any physical device has only one radio and it can assign each of its n neighbours a different channel to broadcast some messages on and it can then choose a channel randomly to communicate with other vehicles. However, this is subject to the risk of not accurately detecting all Sybil nodes and making all communication channels busy as well. Hubaux[5] proposed another solution as three or more road side units perform distance bounding on a vehicle before computing its location. In [6] public key cryptography was used to solve the security problems in VANETs but it was a general solution not specific to any type of attack. Guette and Bryce [7] suggested a secure
Combating Sybil Attacks in Vehicular Ad Hoc Networks
67
hardware based method built on the trusted platform module (TPM). Secure information and related protocols are stored in shielded locations of the module where any forging or manufacturing of data is impossible, and the platform credentials are trusted by car manufacturers; therefore, the communications between TPMs of the vehicles are protected from the Sybil attack. However, as the TPM is a variation of a certificate, it still needs trusted authorities that can take the responsibility of managing individual vehicles it is also costly because it requires specific hardware. In [8] there exists a solution that depends on Road Side Units (RSUs) broadcasting the digital signatures with timestamp to vehicles in their communication range, however the attacker can impersonate RSU and injects malformed digital signatures with timestamps that can affect the total decision of VANETs. Certified timestamps signed by RSUs are sent to vehicles upon request were suggested in [9], but this solution requires time synchronization between all RSUs, which may be difficult. Also, RSUs are subject to resource exhausting by continuously sending previous time stamps to enforce it to produce aggregated time stamp.
3 Proposed Scheme for Sybil Attack Prevention In this section, we present our proposed scheme that aims to detect and prevent the Sybil attack. We will start by presenting the VANET architecture that we shall be adopting throughout this paper, as well as our assumptions. This will be followed by describing the steps that should be used for the Sybil attack prevention. In the VANET architecture, there are Road-Side Boxes (RSBs) spread along the roads, they represent semi-trusted entities that work in conjunction with the Department of Motor Vehicle (DMV) that is responsible for yearly registration renewal of all vehicles and plays a very critical rule in detecting the Sybil attack in the proposed solution. The DMV will be treated as a standard certificate authority that regularly receives a valid certificate from the national root certificate. Fig. 2 depicts the VANET architecture that we shall be using throughout this document.
Fig. 2. VANET Architecture [1]
68
K.M. Rabieh and M.A. Azer
In vehicular network applications, vehicles are expected to broadcast new events whenever they are detected. In Order to clearly define the notion of events, we need to unambiguously standardize their format. An event consists of the following fields: • • • •
Timestamp Location (x,y) The information describing the event Car ID
Events can be categorized into two types: • Green events like spreading unique packets for each vehicle to attract the attention of surrounding vehicles. This is called beaconing. • Red events like Emergency, accidents, traffic jam. These events are more important and subject to attacks and therefore may lead to disasters and human injury or human death. Our proposed scheme has the following assumptions: • The car’s owner has a smart token or smart card that carries his digital signature. • The car itself has a kind of smart card reader or USB for the use of the token. • Either the smart card or the token is tamper proof, it carries the private key of the driver and it is able of signing events. In order to detect and prevent the Sybil attack, our proposed scheme has the following steps: 1- The driver watches an emergency and needs to inform the DMV for taking an action. 2- The driver describes the event and inserts his own card/token in the reader/USB. 3- The event is structured such that the information describing the event is entered by the driver in addition to the timestamp of the event. The driver signs the dual hash which is based on the dual signature idea used in Secure Electronic Transaction (SET [10]) protocol to guarantee the linkage between the payment order and purchase order. In the proposed solution we replace the payment order and purchase order originally mentioned in the SET protocol by the personal ID stored in the token/smart card and the event data. Both are hashed and the hashes are concatenated, hashed again and then signed by the private key from the token/smart card. Fig. 3 depicts the process of the dual signature. The Personal ID is taken from the smart card, the event is created with a predefined structure and both are signed by the driver’s private key, this is to ensure the nonrepudiation. Similar events with the same signature within specified time frames will be discarded by the RSBs; this is to ensure that Sybil attack is prevented. Another important function added by the Dual Signature is that it links between the person’s ID and the event, therefore; it can be used for legal issues in case of attacks.
Combating Sybil Attacks in Vehicular Ad Hoc Networks
69
Fig. 3. Customized Dual Signature for VANETs
Fig. 4. Creation of Digital Envelope inside the processing centre of the vehicle
4- A Digital envelope is then created as shown Fig.4, it consists of four fields as follows. • Event: Information that describes the emergency or the event. • Personal ID: Unique identifier for each vehicle owner. • Dual Signature: Consisting of the event and the personal ID signed by user’s private Key. • Random Symmetric key that is used to encrypt the whole digital envelope. The digital envelope is encrypted by a random symmetric key because symmetric encryption is much faster than asymmetric one, whereas the symmetric key itself is encrypted by the DMV public key. Hence, only the DMV can decrypt and read out the event. The Digital envelope is created inside the processing centre of the vehicle and then transmitted to the RSB. 5- As the RSB is a semi– trusted device, it only forwards the envelope to the DMV; The DMV decrypts the envelope using the private key, decodes the event, gets the
70
K.M. Rabieh and M.A. Azer
ƌŝǀĞƌŚĂƐ ƐŵĂƌƚĂƌĚ
dŚĞĚƌŝǀĞƌǁĂŶƚƐƚŽŝŶĨŽƌŵ ƚŚĞDsĂďŽƵƚĂĐĞƌƚĂŝŶ ĞǀĞŶƚ
^ƚĂƌƚ
zĞƐ dŚĞĚƌŝǀĞƌĚĞƐĐƌŝďĞƐƚŚĞĞǀĞŶƚĂŶĚ ŝŶƐĞƌƚƐŚŝƐŽǁŶƐŵĂƌƚĐĂƌĚ
EŽ E ƌŝǀĞƌŚĂƐĂ h^dŽŬĞŶ
^LJďŝůĂƚƚĂĐŬŝƐĚĞƚĞĐƚĞĚĂŶĚĐĂƌŝƐ ďůŽĐŬĞĚ͘
zĞƐ
EŽ ŝŐŝƚĂůŶǀĞůŽƉĞŝƐĐƌĞĂƚĞĚĂŶĚŝƐ ƐĞŶƚƚŽƚŚĞZ^
džŝƚ
zĞƐ EŽ /ŶĐŽŶƐŝƐƚĞŶĐLJ ŽĐĐƵƌƌĞĚ͍
dŚĞĚƌŝǀĞƌĚĞƐĐƌŝďĞƐƚŚĞ ĞǀĞŶƚĂŶĚŝŶƐĞƌƚƐŚŝƐ ŽǁŶh^dŽŬĞŶ
Z^ǁŝůůĐĂŶĐĞůƚŚĞ ^ƵƐƉŝĐŝŽƵƐDŽĚĞ
Z^ĨŽƌǁĂƌĚƐƚŚĞĞŶǀĞůŽƉĞƚŽƚŚĞDs
DsǀĂůŝĚĂƚĞƐƚŚĞĐĂƌ/ĂŶĚƌĞƐƉŽŶĚƐƚŽ ƚŚĞZ^
Z^ǁŝůůƉŝĐŬƌĂŶĚŽŵŶƵŵďĞƌŽĨŚĂƐŚŝĚƐŽĨĐĂƌƐ ĂŶĚǁŝůůĐŽŶŶĞĐƚƚŽĂĚĞǀŝĐĞŝŶƚŚĞĐĂƌƚŚĂƚŚĂƐ ĂĐĐĞƐƐƚŽƐŽŵĞǀĂƌŝĂďůĞƐůŝŬĞƚŚĞǀĞůŽĐŝƚLJĂŶĚ ĂĐĐĞůĞƌĂƚŝŽŶ
EŽ Z^ Z^ǁŝůůƚƵƌŶƚŽ^ƵƐƉŝĐŝŽƵƐDŽĚĞ
sĂůŝĚZ/
ŽƵŶƚĞƌŝƐ ŝŶĐƌĞĂƐĞĚ
zĞƐ Z^ƚŚƌĞƐŚŽůĚ sĂůƵĞZĞĂĐŚĞĚ͍
Z^ǁĂŝƚƐ ĨŽƌŽƚŚĞƌĞǀĞŶƚƐ
Fig. 5. Flow chart of the proposed algorithm
vehicle’s ID and compares it with the database to determine the owner of the car. If the owner of the car’s ID is identical to the personal ID sent in the digital envelope, the RSB forwards its response to the DMV to confirm that it is valid, otherwise it will respond to the RSB by invalid. A Threshold value is maintained in all RSBS. If an RSB’s counter has reached this threshold the RSB will turn to suspicious mode. This value is adjustable and it should depend on the region’s vitality to the motion of the city. 6- During the suspicious mode, in order to verify whether an attack (co-operative Sybil attack) has occurred or not, the RSB will pick a random number of hash IDs of cars and will connect to a device in the car that has access to some variables like the velocity and acceleration of the vehicle. If there is consensus of the events, then suspicious mode is cancelled and the RSB is sure that an accident has really occurred. The RSB should now inform the DMV about the accident’s location determined from the received events or from the responses of the vehicle devices. The DMV then broadcasts the emergency’s location in order for other vehicles to take another route.
Combating Sybil Attacks in Vehicular Ad Hoc Networks
71
If there is no consensus, then there is some kind of co-operative Sybil attack. 7- Once an attack is detected, the RSBs block any further messages from the car ID which reports false events. The RSB also reports the result to the DMV that in turns applies penalties to the guilty car or group of cars. 8- A receipt is then sent to the attacker to inform him about the penalty.
4 Conclusions and Future Work In this paper, we proposed a method for detecting and combatting Sybil attacks. This method depends on well- established architecture through distributed RSBs along the road and a centralized DMV which decides whether Sybil nodes exist or not. The mechanism is based on PKIand takes advantages of Dual Signature concept explained in the secure electronic transaction protocol to be used in courts as an evidence of the occurrence of such an attack. Based on PKI, the solution takes advantage of the digital envelope in which a digital signed combination of personal ID, event, and dual signature are encrypted with the DMV public key to be transferred to the DMV. This guarantees both security and privacy preservation of the Vehicle Information and the Personal ID information as well. We plan to simulate our proposed scheme as a first phase in order to evaluate its performance and complexity in different environments. We also plan to investigate through simulations the optimum threshold value for the RSB in order to decide whether an attack has occurred or not. This is a challenge because if the threshold is too low, we risk having lots of false positives; whereas if it is too high false negatives can occur. Moreover, Traditionally, Certificate Revocation Lists (CRLs) were used in conjunction with PKI schemes in order to verify the validity of certificates used within the network. However, we are willing to make use of the Online Certificate Status Protocol (OCSP) [11], by integrating it to our proposed scheme, to guarantee that the used certificates are fresh enough and avoid using already revoked ones. This protocol is much easier and faster than the CRLs, and therefore more convenient to VANETs.
References 1. Zhou, T., Choudhury, R.R., Ning, N., Chakrabarty, K.: Privacy-Preserving Detection of Sybil Attacks in Vehicular Ad Hoc Networks. In: Proceedings of the 2007 Fourth Annual International Conference on Mobile and Ubiquitous Systems Networking&Services, MobiQuitous (2007) 2. Yu, H., Kaminsky, M., Gibbons, P., Flaxman, A.: Defending Against Sybil Attacks via Social Networks. In: Proceedings of the 2006 conference on Applications, Technologies, Architectures, and Protocols for Computer Communication (2008) 3. Douceur, J.: The Sybil Attack in Peer-To-Peer Systems. In: Proceedings of First International Workshop on Peer-to-Peer Systems, March 7-8 (2002) 4. Newsome, J., Shi, E., Song, D., Perrig, A.: The Sybil Attack in Sensor Networks, Analysis & Defences. In: Third International Symposium on Information Processing in Sensor Networks, IPSN 2004 (2004)
72
K.M. Rabieh and M.A. Azer
5. Hubaux, J.P., Capkun, S., Luo, J.: The Security and Privacy of Smart Vehicles. IEEE Security and Privacy Magazine 2(3), 49–55 (2004) 6. Raya, M., Hubaux, J.P.: Securing vehicular ad hoc networks. Journal of Computer Security - Special Issue on Security of Ad-hoc and Sensor Networks (2007) 7. Guette, G., Bryce, C.: Using TPMs to Secure Vehicular Ad-Hoc Networks (VANETs). In: Proceedings of the 2nd IFIP WG 11.2 International Conference On Information Security Theory and Practices: Smart Devices, Convergence and Next Generation Networks (2008) 8. Chen, C., Wang, X., Han, W., Zang, B.: A Robust Detection of the Sybil Attack in Urban VANETs. In: Proceedings of the 2009 29th IEEE International Conference on Distributed Computing Systems Workshops, ICDCSW 2009 (2009) 9. Park, S., Aslam, B., Turgut, D., Zou, C.: Defense Against Sybil Attack in Vehicular Ad Hoc Network-based on Roadside Unit Support. In: Proceedings of the 28th IEEE Conference on Military Communications (2009) 10. Wei, H., Huang, S., Vi, G., Xie, Z.: An Anonymity Improvement Scheme of Secure Electronic Transactions Protocols. In: 3rd International Conference on Advanced Computer Theory and Engineering(ICACTE 2010) (2010) 11. Zhao, X., Wenyan, Z., Shanshan, C.: New Certificate Status Verification Scheme Based on OCSP for Wireless Environment. In: International Forum on Computer ScienceTechnology and Applications (2009) 12. El Zarki, M., Mehrotra, S., Tsudik, G., Venkatasubramanian, N.: Security Issues in A Future Vehicular Network. In: Euro Wireless Conference (2002) 13. Yan, G., Choudhary, G., Weigle, M., Olariu, S.: Providing VANET security through Active Position Detection. In: Proceedings of the Fourth ACM International Workshop On Vehicular Ad Hoc Networks (September 2007)
TCSAP: A New Secure and Robust Modified MANETconf Protocol Abdelhafid Abdelmalek1,2 , Zohra Slimane1 , Mohamed Feham1 , and Abdelmalik Taleb-Ahmed2 1
STIC Laboratory University of Tlemcen Algeria LAMIH Laboratory University of Valenciennes France {a_abdelmalek,m_feham,z_slimani}@mail.univ-tlemcen.dz,
[email protected] 2
Abstract. Different protocols have been developed throughout the last years to achieve automatic IP address allocation in Mobile Ad hoc Networks (MANETs). However, Autoconfiguration security issues are still an open problem. In this paper, a new secure and robust IP Address allocation protocol for standalone MANETs inspired from MANETconf and named TCSAP is specified and evaluated within NS2. The proposed solution is efficient and thwarts all possible attacks associated with dynamic IP address assignment in MANETs.
1
Introduction
In the last decade, large research efforts have been made to address challenges posed by MANETs, These challenges include mainly IP address autoconfiguration, routing, security and QoS issues. In security context, the major part of research up to now was concentrated mainly on trust models and routing security problems. However, the lack of security in previously suggested autoconfiguration schemes can lead to serious attacks in potentially hostile environments, mainly IP spoofing attack, sybil attack, traffic overload DoS attack, exhaustion address space attack, and conflict address attack. This problem was tackled by some few papers [1]-[5]. We have analyzed these proposals and pointed out their weaknesses and shortcomings in [13]; we have identified also the imperative security requirements related to this problem. In the present paper, we propose a new robust and secure stateful IP address allocation protocol for MANETs, by applying a cooperative security scheme to cope with malicious nodes including misbehaving nodes that could be compromised by potential adversaries. The scheme relies on a fully distributed Certification Authority based trust model in conjunction with a threshold signature scheme for issuing and revoking certificates, and ‘On-line Joint IP Address and Public Key Certificate’ ; this solves definitively the problem of some attacks such as IP spoofing and Sybil attacks, unsolved up to now by conventional mechanisms. The remainder of the paper is organized as follows. In section 2, we develop our secure and robust autoconfiguration scheme on the basis of threshold cryptographic tools. Section 3 is devoted A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 73–82, 2011. c Springer-Verlag Berlin Heidelberg 2011
74
A. Abdelmalek et al.
to the design of the basic building blocks of the protocol TCSAP. A security discussion is given in section 4. Section 5 presents our simulation experiments. Finally, section 6 concludes the paper.
2
Threshold Concept Based Autoconfiguration and Trust Model
We adopt in our solution for both schemes: IP Address allocation and trust model, a fully distributed approach based on threshold concept. 2.1
Trust Model
Our trust model is based on a Fully Distributed CA [7] in conjunction with threshold cryptography [8]. In a (k,n) threshold cryptosystem, the network consists of n nodes, each node holds a share of the Network’s private key. The On-line CA service is achieved transparently by a large enough subset of nodes (i.e. a number greater or equal to the threshold k). To implement a threshold cryptosystem in a spontaneous MANETs, we need: 1. To generate randomly and in a distributed manner (without a trusted party) a pair of Network’s private/public keys, to split the Network’s private key among the network and to allow shareholders to verify the correctness of their shares. This is done by a joint verifiable random secret sharing protocol [10] based on Shamir’s secret sharing [9]. 2. To provide for any new joining node with a share of the Network’s private key [7] 3. To provide a threshold digital signature scheme to sign issued, renewed or revoked certificates. With regard to the threshold signature protocol, a variety of discrete log based schemes have been proposed [11] including NybergRuepple or ElGamal- like and Elliptic Curve threshold digital signatures. Note that in our scheme, each node must hold: On one hand, a valid share of the Network’s private key and a pair of private/public keys approved by the On-line CA. 2.2
Autoconfiguration Model
Let us consider a standalone MANET. We develop a stateful autoconfiguration scheme inspired from MANETconf [6]. We distribute the autoconfiguration service to all nodes in such a way that only a threshold number of nodes can collaborate in performing the service functionality. Then, the IP Address for a newly arrived node is assigned by a subset of at least k nodes. Instead of MANETconf scheme in which an affirmative response from all nodes in the network is needed before assigning any available IP address to a newly arrived node, our scheme modifies MANETconf protocol and saves the communication bandwidth
TCSAP: A New Secure and Robust Modified MANETconf Protocol
75
by assigning free IP addresses without asking permission from any other node in the MANET. To achieve this, we divide the Address Space into a fixed number (say M ) of disjoint IP Address Blocks with equal sizes. We define an IP Address Block as a consecutive range of IP addresses. The parameter M is a power of 2. Maintained State Information: At any instant of time, each configured node (see 3.1) must maintain some state information defined hereafter: – Free IP Address Table (FAT ): contains the lowest free address of each IP Address Block. (i-e M values) – Pending IP Address Table (PAT ): contains recently assigned IP addresses which are not registered yet. – Registered IP Address Table (RAT ): Each entry in this table contains any assigned and registered IP address, the corresponding node’s identity, its public key and the On-line joint certificate validity period. A registered node will be removed from the RAT if its certificate has expired. Nodes wishing to maintain their addresses must make a request for maintenance within a time specified before the expiry of their certificates. – Requester Counter (RC ): this counter is maintained for each new node requesting for an autoconfiguration and to which an IP address is assigned but not registered yet. It is incremented for each new request. To prevent the Exhaustion Address Space Attack, the authorized attempts for the Autoconfiguration Service Requesting are limited. A configured node must update his state information in the following situations: (i) Each time it reboots, (ii) Each time it leaves and joins the MANET again, (iii) If it has not been solicited for a long time to perform the Autoconfiguration Service. The node wishing to update its state information must collect redundant data from at least k nodes. IP Address Assignment: A new node will be assigned randomly one of the lowest free addresses contained in the FAT, which means that the IP address is assigned in an increasing order from a randomly chosen IP Address Block. We impose to the new joining node to obtain its IP address from at least k nodes. We use for this purpose the threshold signature described above and the new concept of ‘On-line Joint IP address and Public Key Certificate’. Each allocated IP address in the network is bound to node’s identity by means of this certificate which must be signed by the On-line CA. After having received a signed IP Address, the new node must broadcast a signed registration message to all nodes to be able to participate actively in the network. Any assigned IP address which is not registered yet is removed from the FAT and kept in the PAT, either by the k signer nodes after having assigned this address or by all nodes after having received a registration message for a higher IP address in the same IP Address Block. If the registration message is received the IP address is removed from the PAT and put in the RAT.
76
3
A. Abdelmalek et al.
TCSAP Protocol Details
This section specifies our new protocol TCSAP (Threshold Cryptography based Secure Auto-configuration Protocol) which implements the trust model and the autoconfiguration scheme described above. 3.1
Definition of Node’s States
1. Unconfigured node: any node wishing to join the MANET, and which is not already registered with an ‘On-line Joint IP address and Public Key Certificate’. 2. Configured node: any registered node within the MANET with an ‘On-line Joint IP address and Public Key Certificate’ 3. Node with Configuration in Progress: any unconfigured node which has initiated an autoconfiguration process that is not finished yet. 3.2
MANET Neighbors Discovery
MANET Neighbors Discovery protocol allows a node to discover its one-hop neighbors by broadcasting periodically signed Discovery messages. The signature here is done according to the node’s Off-line public key certificate. The node uses the DiscoveryTimer to detect the presence of its one-hop neighbors. This timer is rescheduled each time a Discovery_Request message is broadcast. Type of Messages – Discovery_Request : this message is used by an Unconfigured node to discover its one-hop neighbors. – Discovery_Welcome: is a reply message to any Discovery_Request message when the responder is in state Configured. MANET Neighbors Discovery Protocol: The MANET Neighbors Discovery protocol is executed automatically by a node on boots/reboots when his state is Unconfigured. The Discovery_Request message must contain the originator’s Off-line public key certificate and its signature. The recipient will response if its state is Configured. It checks the signature; if it is valid it replies by a Discovery_Welcome message containing its IP address, its Off-line public key certificate and the signature, otherwise it discards the message. When the requester receives the Discovery_Welcome message, it concludes that a MANET is already established and has to start the Autoconfiguration Service Requesting. 3.3
Autoconfiguration of a Newly Arrived Node
In the proposed scheme, a new joining node is assigned an IP address by means of the Autoconfiguration Service Requesting. Subsequently it is provided with an ‘On-line Joint IP address and Public Key Certificate’. This is achieved in 4 phases: (1) Closest Servers Research, (2) Requesting the On-line certificate, (3) Threshold signature of the On-line certificate and (4) Registration in MANET.
TCSAP: A New Secure and Robust Modified MANETconf Protocol
77
Type of Messages – Config_Request : this message is used to start an Autoconfiguration Service Requesting after receiving a Discovery_Welcome message. – Config_Reply: this message is used by a node in state Configured as a reply to Config_Request message. – Config_Cert_Request : this message is used to request an ‘On-line Joint IP address and Public Key Certificate’. – Config_Cert_Reply: this message is used by the combiner as a reply message to Config_Cert_Request message. – Config_Advert : this message is sent by the combiner to inform all nodes about the new assigned IP address. – Config_Alert : this message is sent when a malicious node is discovered among the coalition. – Config_Register : this message enables a new configured node to perform a registration within the MANET. Closest Servers Research: Upon receiving Discovery_Welcome messages, the newly arrived node broadcasts a Config_Request message to all nodes in a radius rk calculated from the number of received Discovery_Welcome messages. The algorithm for this procedure using the Abstract Protocol Notation [12] is given in Fig. 1-a. Any Configured node receiving a Config_Request message cheks if the requester is listed in the Black List, or if the signature is not valid. If so, the message is discarded; otherwise it checks the Requester Counter (RC ). If the requester has already reached the limit, it is declared as malicious and the message is discarded. Otherwise, the recipient sends a Config_Reply message including the On-line Certification Authority’s public key, the list of available subnets and the corresponding list of the lowest free Host-ID of each block from its FAT, the received HopLimit, its ‘Off-line Public Key Certificate’, and its signature. If the total received Config_Reply messages within a timeout period determined by the ConfigTimer are less than the threshold k, then the requester increments rk and repeats the process. Otherwise, it starts requesting the On-line certificate. Requesting the On-Line Certificate: Upon receiving Config_Reply messages, the requester starts requesting an On-line certificate. The procedure is summarized in the following steps (the algorithm is given in Fig. 1-b): 1. Step1: It selects among the closest responding nodes a coalition of at least k nodes according to the received HopLimit values appearing in the Config_Reply messages. 2. Step2: It chooses randomly a lowest free Host-ID common to all the members of the selected coalition. 3. Step3: It unicasts to these members a Config_Cert_Request message including the list of the coalition members, the chosen lowest free Host-ID,
78
A. Abdelmalek et al.
its ‘Off-line Public Key Certificate’ and its signature, expecting reception of a Config_Cert_Reply message from the combiner within a timeout period determined by the ConfigCertTimer.
Fig. 1. APN Algorithm: (a) Research of closest servers, (b) Procedure of coalition selection and certification requesting
Threshold Signature of the On-Line Certificate: Each member in the coalition checks the validity of the Config_Cert_Request message, looks in its CRL and Black List tables if no member of the coalition is malicious nor his public key is revoked. If this holds, then each member makes its partial threshold signature for the requested ‘On-line Joint IP address and Public Key Certificate’. The combiner of the partial signatures replies to the requester by a Config_Cert_Reply message, and informs all nodes by a Config_Advert message that an IP address has been attributed to the node in question. Then, all nodes increment its Requester Counter (RC ) and delete this address from the FAT and save it in the PAT. Hence, a new coming node will not have the possibility of choosing this address. If a malicious node has been discovered among the coalition members, a Config_Alert message is sent to the honest members of the coalition and to the new joining node. Subsequently, the requester performs a new coalition selection while excluding the malicious nodes. Registration in MANET: To proceed to registration in MANET, the newly arrived node broadcasts to all nodes a Config_Register message using the sitelocal scope all-nodes multicast address FF05::1 as destination address. This message must include the new node’s ‘On-line Joint IP address and Public Key
TCSAP: A New Secure and Robust Modified MANETconf Protocol
79
Certificate’ and the signature of the whole IPv6 packet. This request must be processed by each node without any acknowledgement.
4
Security Discussion
In this work, we have adopted a threshold cryptographic approach (n≥2k-1) to achieve for our scheme security and robustness in the presence of (k-1) faults. Unlike the previous approaches, which rely on a single node either in providing autoconfiguration service, in our scheme the service is initiated and provided by at least k arbitrary honest nodes. Consequently, we avoid any single point of failure or trusted party. Moreover, our scheme is totally distributed over the whole network, and a new joining node does not need any particular distribution of its neighbors to be initiated with network and security parameters. Hence, the service availability is guaranteed ubiquitously. The mechanism of mutual authentication with Off-line certificates allows the servers to authenticate the requester, that is only legitimate nodes can take part in the network, but also the requester to authenticate the servers to prevent Man-In-the-Middle Attack. However, malicious nodes may be present among the servers selected by the requester. For this reason, Config_Alert messages are used to prevent malicious nodes from providing or disturbing the autoconfiguration service. Threshold signature verification should also be used to isolate misbehaving nodes that are not yet in Black List. The requester may also be malicious, the Requester Counter (RC ) and the registration mechanism can efficiently thwart Exhaustion Address Space and Sybil Attacks, the only possible ones in this case. The Traffic overload DoS Attack is prevented by the maximum authorized HopLimit (less than the threshold k) used in Config_Request messages. The mechanism of assigning an IP address by a coalition instead of a single entity solves the problem of Conflict Address Attack present in both stateful and stateless earlier approaches. The concept of ‘On-line Joint IP address and Public Key Certificate’ we introduced in our scheme represents, in the other hand, an effective mechanism to thwart IP Spoofing Attack. A malicious node which wants to spoof either an unused IP address or an already assigned IP address must hold an ‘On-line Joint IP address and Public Key Certificate’ in which its public key is bind to the spoofed IP address. Hence, instead of the limitations of the various solution approaches analyzed in [13], none of the attacks quoted in section 1 appear to break our proposal.
5
Simulation Experiments
Simulation experiments were performed using the network simulator NS-2 [14] with CMU mobility extensions to evaluate the performance of our protocol in terms of configuration latency and communication overhead. The configuration latency metric represents the average delay for a new joining node to obtain an ‘On-line Joint IP address and Public Key Certificate’. This includes all possible
80
A. Abdelmalek et al.
delays caused by the messages exchanges, timeouts and cryptographic primitives. The communication overhead metric represents the number of control packets transmitted during the autoconfiguration process. The protocol TCSAP is implemented within NS-2 using C++, by creating a new agent TCSAP. The cryptographic primitives were simply modeled by delays. We used the results published by [15] for a 1.83 GHz Dual Core Intel processor, under Windows Vista (32 bits mode). We considered the RSA-2048 algorithm and the ECDSA233 algorithm respectively for ordinary signature and threshold signature. 5.1
Simulation Scenarios and Parameters
The random waypoint mobility model was used. The simulation time was set to 120 seconds. We used the AODV routing protocol. Each data point represents an average value of five runs with the same settings, but different randomly generated topology and mobility scenarios. The following sets of simulation were performed. a) Varying network density: We study here the effect of the network density on latency and communication overhead. The area of the network was set to 1000m*1000m, for the 15, 25, 50, 75 and 100 node population, ensuring respectively 15, 25, 50, 75, and 100 nodes/km2 for the network density. The simulations were performed for different values of threshold k. No motion was applied in this scenario. b) Varying network mobility: we examine the protocol efficiency when the mobility of nodes increases. A network area of 1000m*1000m with 50 nodes is simulated for different values of threshold k. We vary the maximum node speed from 0 to 50 m/s; pause time is set to 0, according to the following command (example for 20 m/s node speed): Setdest –v2 –n 50 –s 1 –m 20 –M 20 –t 120 –P 1 –p 0 –x 1000 –y 1000. 5.2
Simulation Results
a) Latency: 1. Impact of network density: Figure (Fig. 3-a) shows an increase in latency when the network density is low (below 25 nodes/km2), in particular for the high threshold values. The mean value of latency is less than one second. The minimum was observed at (25 nodes/km2). But again, from this point latency increase linear with respect to density. Latency increase also with respect to the threshold parameter. 2. Impact of mobility: It was observed that node mobility has no significant effect on latency (Fig. 3-b). This was because the simulated speeds were lower than 50m/s, and that the mean latency is less than 1 second, for such delay a node movement does not exceed 50m, and this in most time does not break links. In some particular cases, the mobility may have positive/ negative impact on latency.
TCSAP: A New Secure and Robust Modified MANETconf Protocol
81
Fig. 2. Configuration Latency: (a) vs Network Density , (b) vs Node Mobility
Fig. 3. Communication Overhead: (a) vs Network Density, (b) vs Node Mobility
b) Communication Overhead: 1. Impact of network density: In networks with high density, there are more nodes in the neighbourhood of the new joining node, and all reply to its autoconfiguration service requesting, leading to a higher number of messages exchange. For this raison, we observe in (Fig. 4-a) an increasing in overhead when density increases. Note that this will increase also latency. 2. Impact of mobility: For the same reasons provided above, the node mobility has no significant effect on overhead. (Fig. 4-b)
6
Conclusion
The TCSAP protocol proposed in this paper achieves IPv6 stateful dynamic configuration for MANETs. Our solution provides both security and robustness
82
A. Abdelmalek et al.
and overcomes all the limitations of the previously proposed approaches while still ensuring the timely IP address allocation. Furthermore, instead of others approaches which use separate mechanisms for network parameters configuration and security parameters configuration, our scheme achieves the same purpose at once, which make it efficient in terms of latency and communication overhead as shown by NS2 simulation results.
References 1. Buiati, F., Puttini, R.S., de Sousa Jr., R.T.: A Secure Autoconfiguration Protocol for MANET Nodes. In: Nikolaidis, I., Barbeau, M., An, H.-C. (eds.) ADHOC-NOW 2004. LNCS, vol. 3158, pp. 108–121. Springer, Heidelberg (2004) 2. Cavalli, A., Orset, J.M.: Secure hosts autoconfiguration in mobile ad hoc networks. In: ICDCSW 2004, pp. 809–814 (2004) 3. Wang, P., Reeves, D.S., Ning, P.: Secure Address Autoconfiguration for Mobile Ad Hoc Networks. In: MOBIQUITOUS 2005, pp. 519–522 (2005) 4. Hu, S., Mitchell, C.J.: Improving IP Address Autoconfiguration Security in MANETs Using Trust Modelling. In: Jia, X., Wu, J., He, Y. (eds.) MSN 2005. LNCS, vol. 3794, pp. 83–92. Springer, Heidelberg (2005) 5. Langer, A., Kühnert, T.: Security issues in Address Autoconfiguration Protocols: An improved version of the Optimized Dynamic Address Configuration Protocol. In: archiv.tu-chemnitz.de (2007) 6. Nesargi, S., Prakash, R.: MANETconf: Configuration of Hosts in a Mobile Ad Hoc Network. In: IEEE INFOCOM (June 2002) 7. Kong, J., Zerfos, P., Luo, H., Lu, S., Zhang, L.: Providing Robust and Ubiquitous Security Support for MANET. In: IEEE International Conference on Network Protocols, pp. 251–260 (November 2001) 8. Di Crescenzo, G., Arce, G., Ge, R.: Threshold Cryptography in Mobile Ad Hoc Networks. In: Blundo, C., Cimato, S. (eds.) SCN 2004. LNCS, vol. 3352, pp. 91–104. Springer, Heidelberg (2005) 9. Shamir, A.: How to Share a Secret. Communications of the ACM 22(11), 612–613 (1979) 10. Pedersen, T.P.: A threshold cryptosystem without a trusted party. In: Davies, D.W. (ed.) EUROCRYPT 1991. LNCS, vol. 547, pp. 522–526. Springer, Heidelberg (1991) 11. Hwang, M., Chang, T.: Threshold Signatures: Current Status and Key Issues. International Journal of Network Security 1(3), 123–137 (2005) 12. Gouda, M.G.: Elements of Network Protocol Design. John Wiley and Sons, Chichester (1998) 13. Abdelmalek, A., Feham, M., Taleb-Ahmed, A.: On Recent Security Enhancements to Autoconfiguration Protocols for MANETs: Real Threats and Requirements. IJCSNS 9(4), 401–407 (2009) 14. The Network Simulator manual, The NS2 homepage, http://www.isi.edu/nsnam/ns 15. Speed Comparison of Popular Crypto Algorithms, http://www.cryptopp.com
Highly Resilient Communication Using Affine Planes for Key Predistribution and Reed Muller Codes for Connectivity in Wireless Sensor Network Samiran Bag1 , Amrita Saha2 , and Pinaki Sarkar3 1
3
Applied Statistics Unit, Indian Statistical Institute, Kolkata-700108, India samiran
[email protected] 2 CSE Department, IIT Bombay, Mumbai-400076, India
[email protected] Department of Mathematics, Jadavpur University, Kolkata-700032, India
[email protected] Abstract. Wireless Sensor Networks (WSN) consist of low powered and resource constrained sensor nodes which are left unattended for long duration of time. Hence it is very challenging to design and implement cost effective security protocols for such networks. Thus symmetric key cryptographic techniques are preferred over public key techniques for communication in such scenarios. Prior to deployment, keys are usually predistributed into the nodes and this problem has been well studied. Highlighting that connectivity and communication are two separate aspects of a WSN, we propose a secure connectivity model using Reed Muller codes. The model is then utilized to securely establish communication keys and exchange messages in a WSN designed on the basis of a scheme that uses affine planes for key predistribution. Novel combination of both the ideas yields highly resilient communication model with full connectivity between nodes. Keywords: Security, Connectivity, Communication, Reed-Muller Codes, Affine Planes.
1 Introduction Wireless sensor networks consist of tiny sensor nodes that have very limited battery power, less amount of storage, low computational power and they are scattered in large numbers over a vast region. The sensors communicate between each other and with the base station via radio frequencies. These networks are used in civilian purposes like smoke detection, wild fire detection, seismic activity monitoring, ocean temperature monitoring, salinity monitoring of sea water. Besides they have large application in military purposes, for instance monitoring enemy movements. Clearly, the nodes deal with very sensitive data and can communicate within a special range called Radio Frequency range. Since sensors are deployed unattended over the target area this makes them physically insecure and prone to adversarial attacks. Thus arises the need of secure communication model in WSN to circumvent these attacks. ¨ A. Ozcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 83–94, 2011. c Springer-Verlag Berlin Heidelberg 2011
84
S. Bag, A. Saha, and P. Sarkar
A secure communication model makes use of (low cost) cryptographic primitives. Existing schemes like Kerberos [11] & public key cryptography [6] are not suitable to this kind of resource constrained system due to inherent cost associated to them. Key predistribution is a method to preload cryptographic keys in sensor nodesn before they are deployed in the target area. It is a symmetric key approach, where two communicating nodes share a common secret key. The message encrypted decrypted using the same secret key. Thus both the sender and receiver nodes must be preloaded with the same key. So prior to deployment every node has to be preloaded with a set of keys called its key ring or key chain. A centralized authority called Base Station or Key Distribution Server (KDS) preloads the key ring of every node from a pool (aka key pool) of keys meant for the entire network. Immediately after deployment shared keys are to be established between nodes before actual communication. This phase is called shared key discovery. In absence of common (shared) keys between two sensors a path-key need to be established between them (aka path key establishment). 1.1 Related Work Key predistribution in sensor networks was first considered by Eschenaur and Gligor [5]. In their work ever key is associated with an unique key identifier. Keys are randomly drawn from the key pool to form the Key rings of the sensors. Key establishment is also random. Such method of key predistributuion is probabilistic in the sense that both key distribution and etablishment is done randomly. Many such probabilistic key predistribution schemes have been well studied and presented in a survey report publisehed in 2005 by C¸ampete and Yenner [2]. Shared key establishment and Path key discovery can become very difficult task for above probabilistic approaches. Lee and Stinson proposed two schemes [7,8] where they have adopted combinatorial techniques for predistribuion and later establishment of keys. Their works also suggests that both shared key establishment and path key discovery can be better achieved by the suggested deterministic approach. Chakrabarti et al. [3] proposed a hybrid key predistribution scheme by merging the blocks in combinatorial designs. They randomly selected blocks from transversal design proposed by Lee and Stinson [7,8] & merged them to form the sensor nodes. Though this technique increase the key ring sizes per node, it improves the resilience & communication probability of the network. Ruj & Roy [9,10] used several combinatorial designs & codes like Partially balanced incomplete block designs (PBIBD), transversal design & Reed-Solomon codes to predistribute keys. 1.2 Our Contribution Very recently, Bag and Ruj [1] have utilized finite affine geometry to propose a deterministic key predistribution scheme. In this paper we discuss enhancement of resiliency of their scheme. Their scheme uses finite affine plane over Zq , where q is a prime. For this we observe that communication and connectivity are two separate aspects of a WSN. Then apply Reed Muller Codes to model the connectivity aspect so as to make it secure by using suitable cryptosystems. To the best of our knowledge, this novel idea of separating connectivity from communication and then applying a secure model to
Highly Resilient Communication Using Affine Planes for Key Predistribution
85
the connectivity aspect of a WSN been proposed for the first time by Sarkar et al. in [12]. Combination of both the schemes results in a highly resilient key predistribution scheme for a WSN providing full connectivity amongst the nodes. 1.3 Basic Notions Before explicitly explaining the various aspect of our design, we require some basic notions like communication, connectivity, the respective key and communication radius which have been stated in [12, section II]. Throughout the paper we shall use the term “Uncompromised nodes” to mean nodes that are not compromised. The words “communication” and “connectivity/connection” are sometimes abbreviated to com. and con. respectively. The terms “communication model/scheme” and “key predistribution model/scheme” will mean the same.
2 Communication Model Our design is based on a scheme by Bag & Ruj [1]. In their scheme the authors used finite affine plane over Zq where q is a prime number. Affine plane over Zq contains as many as q 2 points and are usually denoted by AG(2, q). The entire key space is split 2 into 4 parts, each part containing q4 points and from each part the ith point is assigned 2
to the ith node. Thus there are a total of q4 nodes, each containing precisely 4 points. The lines through all 4 points of a node represent the set of keys in that particular node. As demonstrated in [1, section VI] there can be 4q − 2 to 4q + 1 keys belonging to any node . The lines through any two points of two distinct nodes serve as the identifier of a common keys between the nodes. The authors showed in [1, section VI] that there can be 1 to 16 common keys between a pair of nodes. Suppose 2 nodes with id i and j want to establish their common keys. They do so by finding lines through any two points belonging to them as follows: The points are distributed among the nodes in such a fashion that the node’s ids reveal the points they contain. Thus on receiving the id of node j, node i gets to know the points in node j. So it can find one line passing through any of its 4 points and any of the points of node j. Similarly if node j uses the same algorithm as node i it will end up finding the same line as node i. As these lines represents the ids of the shared keys between the nodes, the nodes can communicate with thus established common keys.
3 Weakness: Motivation of Our Work We observe a weakness in the aforesaid key predistribution scheme. Here the node ids reveal the points inside a particular node. Let us say node i and node j want to establish their keys securely. An adversary, say Alice can tap the radio frequency channel and come to know the unencrypted node ids passing through them. She can then find the key ids of the shared keys between the sensors in a manner similar to the computation done by the nodes. This clearly implies that selective node attack is quite feasible.
86
S. Bag, A. Saha, and P. Sarkar
These points are again contained in a number of nodes of the sensor network. She can capture one of them and get to know the actual keys. Combined with the knowledge of node ids, she can use these keys to actually affect the com. amongst other nodes. To counter this problem, we first differentiate the two aspects communication and connectivity of a WSN. Then like in [12], apply Reed Muller Codes to suitably model the connectivity aspect. The construction of the model is presented in the following section. The model can be made secure by using suitable cryptosystems. As shall be later established the combination of the two ideas results in a highly resilient key predistribution scheme for WSN providing full connectivity amongst nodes with virtually same communication overhead.
4 Proposed Connectivity Model Reed Muller codes will be utilized to structure the connectivity aspect of the WSN. These codes have been elaborately described in [4] and necessary notational changes have been highlighted by Sarkar et al. in [12, section IV]. We follow similar procedure as described in [12, section IV] baring some modification to be illustrated now. First our model will always have three tiers with the “Base Station” or “KDS” in the 1st or topmost tier. The second tier will consist of q4 newly introduced cluster heads (CHs). Amongst these q4 will be assigned q many nodes in the 3rd and the last level. 2
Whereas l = q4 − q q4 nodes has to be under the remaining 1 CH in the last level. 2
Thus our model needs an extra q4 many CHs and can support q4 ordinary nodes (at the last level). It is evident that current connectivity model is heterogeneous in nature, i.e., has different number of nodes in its various clusters. This along with the fact that exactly three tiers are required for our connectivity model distinguishes our design from the original design of Sarkar et al. in [12, section IV]. To build up the cluster between the various tiers of the connectivity model, we shall make use of first order Reed Muller codes. For connectivity of 1st and 2nd levels, we employ a m complete graph where m = q4 . We consider Z2 [x1 , x2 , . . . , x q4 ] in much the same manner as the authors of [12] had considered Z2 [x1 ,qx2 , . . . , xm ]. Like in [12], the monomials xi will represent the bit pattern of length 2 4 having 2i−1 1’s followed by 2i−1 0’s where 1 ≤ i ≤ q4 . A sample connectivity pattern for a cluster containing KDS and 3 CHs can be represented by the following matrix ⎡ ⎤ KDS 1 1 1 1 1 1 1 1 ⎢ CH1 1 0 1 0 1 0 1 0 ⎥ ⎢ ⎥ ⎣ CH2 1 1 0 0 1 1 0 0 ⎦ CH3 1 1 1 1 0 0 0 0 Matrices like the above one are used for construction of Reed Muller codes. This particular matrix has been referred to as R(1; 3) in [4]. Here 1 means the degree of the monomials is ‘1’ and 3 stands for the number of variables. The significance of the entries 1 and 0 in the above matrix (R(1; 3)) is the presence and absence of a connectivity link at that row and column position respectively. Thus for
Highly Resilient Communication Using Affine Planes for Key Predistribution
87
connectivity of two any entities (KDS/CHs/nodes), both of them should have a 1 in the same column for at least one column. Each column is assigned a separate connectivity key immaterial of them using the same radio frequency channel. The connectivity pattern between of each of the clusters of the 2nd and 3rd level is meant to be a 2 complete graph having m = q variables (nodes) in the matrix. Thus we look at Z2 [x1 , x2 , . . . , xq ] as was similarly done in [12, section IV, subsection B] Connectivity matrix for a cluster having 1 CH and 3 nodes is as follows: ⎡ ⎤ CH 1 0 0 1 0 1 1 1 ⎢ N1 1 0 1 0 1 0 1 0 ⎥ ⎢ ⎥ ⎣ N2 1 1 0 0 1 1 0 0 ⎦ N3 1 1 1 1 0 0 0 0 The construction of the second matrix from the first can be found in [12, Section IV, Subsection B]. Here KDS is not present in the inter-nodal links. There is a broadcast channel and a provision for external only for KDS. In the present case instead of 3, we look at q or l many nodes. Here again wherever there is 1, connectivity link is present. Figure 1 give an lively example with q = 11. There are 11 = 3 CHs in 2nd tier. 4 112 rd This model can support 4 = 30 sensors in the 3 & last level. Out of these 42 sensors, 11 ∗ 2 = 22 will be under 2 CHs and only 30 − 22 = 8 under the remaining CH of 2nd level.
5 Deployment There can be various methods for node deployment. We discuss one of them here as an example. At the time of deployment, we shall drop the CHs along with the nodes of its cluster. Clearly instead of totally random deployment, we are deploying in small groups where exact position of nodes may still be unknown. Thus we adopt a kind of group-wise-random or locally-random deployment technique. This ensures that all the clusters are formed according to the model. However in an unlikely event of some nodes falling out of position, we adopt the following key re-scheduling technique. Assume some node of one cluster A falls into another cluster B. In such a case, CH of cluster B broadcasts the node id or I.P. address of the misplaced node amongst all the CHs to find out the actual cluster where it should have been placed. On seeing the I.P. address or node id of this node, the CHs respond whether or not the misplaced node belongs to their cluster. Since this node was supposed to be in cluster A, its CH is the only who responds with ’YES’. Using the secure link between CH of cluster A and cluster B, the connectivity key corresponding to this sensor and CH of cluster A is transmitted to the CH of cluster B. This key is used to set up a secure connectivity link between the CH of cluster B and the misplaced. Depending on the requirements and practical hazards, CH of cluster B decides on the exact connectivity for this misplaced node in its cluster. Clearly a redistribution of connectivity keys may be required. In case this is not possible, still the node remains connected to the network but all communication will involve CH of B. It is clear that in this scenario, there is a process of node addition in cluster B and node deletion at cluster A. These processes have been described in [12] We would like to remark that instead of interconnectivity (clique connectivity) of
88
S. Bag, A. Saha, and P. Sarkar
Fig. 1. Network structure for q = 11 having q4 = 4 CHs in 2nd & N = 30 nodes in 3rd tier
sensor at the base level, one may desire to have just the connection with the CHs. This will enable better security, make (connectivity) key distribution easier and also reduce the importance of simple nodes at the bottommost level. In such a case the 2nd tier CHs may have to be powerful to ensure security.
6 Communication Key Establishment We now describe how one can utilize the secure connectivity model for communication key establishment. As mentioned earlier node ids can be used for this purpose. Every node encrypts its node id using the connectivity key that it shares with its CH and sends the encrypted node id to its CH. On receiving these encrypted ids, the CHs decrypts them and circulates them securely amongst themselves using the connectivity keys of one another. For each incoming node ids, the CHs immediately look up the preloaded ”node-key assigning matrix” (see section 2) for the key ids of the corresponding node. Once the key ids are obtained, common keys are immediately traced and informed back to the node via the same secure channels. Clearly when the nodes send their ids we utilize the connectivity model of last two tiers. Whereas when the node ids are being circulated at the CH level, we use the connectivity keys corresponding to 1st and 2nd level. Surely, if required one can make use of
Highly Resilient Communication Using Affine Planes for Key Predistribution
89
different cryptosystems for various clusters of 2nd & 3rd tiers and certainly for KDS-CH tier (i.e. 1st & 2nd tier) of our connectivity model.
7 Message Sending Protocol Suppose a message has to be sent from node Ni to node Nj for some fixed 1 ≤ i = 2 j ≤ q4 . Then the following protocol is to be executed. Choose one common communication key between Ni and Nj according to [1, section V]. Call it µij . Ni encrypts the message with this key µi,j . if Ni and Nj share a connectivity key then The message encrypted with com. key is again encrypted with the shared con. key and send directly to node Nj . Nj decrypts the outer encryption done using the con. key common to both the nodes. else node Ni uses the con. key that it shares with its Cluster Head and send the doubly encrypted message to its CH. if node Nj lies in the same cluster then After decrypting with Ni ’s con. key and encrypting with Nj ’s con. key, the common CH directly send it to node Nj . Nj decrypts outer encryption done using the con. key that it shares with the (common) CH. else the doubly encrypted message from Ni is decrypted using Ni ’s con. key at the CH of Ni . It is re-encrypted at CH of Ni using the con. key shared with Cluster Head of Nj . Send the doubly encrypted message to the CH of Nj . Cluster Head of Nj then decrypts it with the con. key shared with the cluster head of Ni . CH of Nj encrypts it using the shared con. key with Nj . Send the doubly encrypted message to Nj . Nj will first decrypt the outer encryption done using the con. key of its CH (not Ni ’s). Nj decrypts outer encryption done using the con. key common to both the nodes. Nj decrypts outer encryption done using the con. key common to both the nodes. end if end if Finally Nj uses the chosen common com. key µi,j shared with Ni to decrypt and read the message.
90
S. Bag, A. Saha, and P. Sarkar
8 Communication Probability and Overhead The probability of direct communication of any given pair of nodes is defined as the communication probability of the network. Since the connectivity model is a path connected graph & communication model assures direct communication between every pair of nodes, we conclude that the communication probability of the proposed scheme is 1. However there has to be some trade offs in regards to communication overhead. n many extra connectivity keys have to be stored per node to ensure clique connectivity in every cluster. In the event of nodes getting overloaded, we can alternatively assign only one extra key meant for connection with its CH. It automatically implies every communication between nodes of the last leyer passes through the CHs of 2nd tier. So these CHs must be much powerful units to enable efficient communication. Analyzing resiliency in way similar to [12] assures significant improvements.
9 Resilience A hypothetical intrusion (i.e. attack) detection mechanism informs the KDS, CHs & subsequently the nodes about compromise of any node(s) as and when it occurs. For capture of a node X1 , connectivity keys sacrificed are its broadcast key, keys between X1 & remaining nodes in its cluster and the exclusive key shared by X1 & its CH. Based on this information the concerned nodes and CH delete all the (above) connectivity keys ensuring that the captured node gets thoroughly delinked from the network. This deletion process has been elaborately described in [12, section V, subsection B]. In fact the beauty of this process is that after deletion of required connectivity links due to capture of some node(s), the other nodes in that cluster remains connected in much the same way as they would without the compromised node(s). Remark: Noted that at any stage the communication keys are not known to the CH. Thus for affecting the resiliency of the network, some nodes have to be captured. Introduction of a secure connectivity model enables doubly encryption of message while transmitting. The second encryption involves connectivity of the nodes & CHs. Nodes contain only the con. keys concerned to itself. Connectivity keys of all nodes in a cluster can only be found in CH of that particular cluster (not even in other CHs or KDS). This automatically implies to affect the communication of any node in the network, its CH must be captured. Thus while calculating the effect of the system when some nodes are captured, we must ensure some CHs are also captured. In practice capturing a CH is quite infeasible. 9.1 Analysis of V(s,t) and E(s,t) Define V (s, t) to be the proportion of nodes disconnected when s nodes of 3rd and t CHs of 2nd tier are compromised. Now let us assume that b nodes gets disconnected when all the c CH of 2nd layer are captured. Thus clearly: V (s, c) =
b N −s
Highly Resilient Communication Using Affine Planes for Key Predistribution
91
Since t CH at 2nd tier are captured, only t out of c clusters should get affected. Assuming that the nodes gets disconnected evenly over the entire network, we conclude: V (s, t) =
bt (N − s)c
E(s, t) measures the ratio of links broken when s nodes of 3rd of t CHs at 2nd tier are compromised. Denote the initial number of links in the network by tot links and the number of broken links case by lbrk . Then like in the above case for capture s nodes and all the c CHs of 2nd tier, we get: E(s, c) = 1 −
lbrk tot links
As only t CH at 2nd tier are compromised & assuming the keys are uniformly distributed under the CHs, we conclude: t lbrk E(s, t) = [1 − ] c tot links Note: The assumed distribution of keys under the CHs is uniform. This is not guaranteed fact. However our simulation results suggest that the assumption is reasonable.
10 Scalability: Addition of Node Connectivity model in [12] allows any number of nodes to be added in the network, whereas the communication model of Bag and Ruj [1] is not flexible in this regard. However we propose alternative tricks allowing extra nodes to come in and communicate with pre-existing nodes. In our 1st suggestion the 2nd tier CHs are required to act as trusted authorities (TAs) temporarily upon deployment of any extra node. These CHs then re-organize the clusters, distribute fresh connectivity keys to these nodes and preexisting nodes. Thus the new node get connected to the network. These connectivity keys are to be used for communication purpose also. Though this method seems quite reasonable for practical applications, however one may look to avoid this method as online key redistribution is required here. Alternatively if we know the number of additional nodes to be deployed, then we can pre-load the 2nd tier CH with that many extra con. keys. The extra nodes are to carry only one of these keys meant for connection as well as communication with the appropriate CH. Thus although clique connectivity is not achieved here but still is model is surely scalable. On top of this, if we want clique connectivity for the clusters where these extra nodes join, one has to ensure the number of extra nodes per cluster is less than q. In such a case we can also preload extra q keys per node. (Our aim is to restrict the key ring to O(q)). Under such circumstance, any incoming node should be loaded with the same (extra) keys of the the old nodes along with keys meant for the CH and other new nodes. In this section by key(s) we meant connectivity key(s) only.
92
S. Bag, A. Saha, and P. Sarkar
11 Simulation Results Experimental results tabulated in Table 1 confimed our analysis of V (s, t) and E(s, t) discussed earlier in section 9.1. s & t denotes the assumed number of ordinary sensors and CHs captured respectively. “BR Exp”. is used as an abbreviation for Bag and Ruj’s experimental results as presented in [1]. Appreciable improvements in resiliency can be observed when our experimental (“Exp”) values are compared with those of Bag and Ruj [1] as is clearly visible in Table 1. Table 1. Simulation and comparative results for V (s, t) & E(s, t) q 59 59 89 89 89
N 870 870 1980 1980 1980
s 5 10 11 15 20
t Our Exp. V (s, t) BR Exp. V (s, t) Our Exp. E(s, t) RR Exp. E(s, t) 1 0.000380 0.0057 0.00458 0.068958 2 0.001531 0.01149 0.02094 0.157406 2 0.000472 0.0055 0.00788 0.090639 3 0.000979 0.00757 0.01812 0.139159 4 0.001752 0.0101 0.03687 0.212303
12 Conclusion First one observes that connectivity and communication can be treated as two separate aspects of a WSN. A key predistribution scheme based on affine planes and providing full node-to-node connectivity is then chosen. Now after necessary modifications to the novel secure connectivity model suggested in [12],we apply it to the chosen key predistribution scheme to obtain a highly resilient communication model providing full connectivity amongst nodes. Experimental results presented in section 11 not only confirm this fact but also exhibit the amount of improvement in resilience as compared the original key predistribution scheme proposed by Bag and Ruj in [1]. It is worth noticing that any two given pair of nodes of the resultant system can communicate between one another without their message been exposed to any other node. As has been elaborately explained in section 7, if these two nodes are in ‘radio frequency range’ of each other (and share a connectivity key), doubly encrypted messages can be exchanged directly. In case they are not in each other’s ‘radio frequency range’ or don’t have any common connectivity key, they are supposed to communicate through their CHs. However these CHs can not decrypt the encryption done with communication key shared by the nodes. However the communication model chosen by [12] didn’t provide full connectivity, hence the resultant system didn’t have full connectivity. Choosing a well connected key predistribution scheme settles this issue. Other than this, they didn’t indicate any particular deployment strategy. Thus how exactly the connectivity model was achieved in the target area was not clear. Section 5 has been devoted to address the deployment issue. From the discussion in section 5, it is clear that no physical movement of a node is required as long as there is some CH in its ‘radio frequency range’ after deployment. Considering the hazards of deployment of nodes in a target area of WSN, this observation can be pretty useful to set up a network.
Highly Resilient Communication Using Affine Planes for Key Predistribution
93
13 Future Work Several future research directions stems out of our current work. Though the chosen key predistribution scheme provides direct node-to-node communication, each node 2 has 4q − 2 to 4q + 1 where the size of the network is q4 keys and shares 1 or 16 keys with any other node. These may prove dangerous when some nodes gets captured. Thus we must seek a scheme having lesser keys per node having O(1) keys shared between any pair of nodes. Then one can perhaps apply the connectivity model in a suitable way to get promising results. Repeated enciphering and deciphering has been suggested at each CH in between two communicating nodes of different clusters. Certainly some communication cost will be reduced if one develops a system avoiding this. In this regard, it may be fascinating to see if one can apply other Mathematical tools.
Acknowledgement Firstly we want to express our gratitude to University Grants Commission of India for financially supporting the doctoral program of Mr. Pinaki Sarkar. This work is meant to be a part of the doctoral thesis of Mr. Pinaki Sarkar. We would like to thank Dr. Goutam Paul of Jadavpur University, Kolkata and Mr. Sumit Kumar Pandey of Indian Statistical Institute, Kolkata for discussing the paper and critically analyzing it. A special word of appreciation goes to Dr. Brijesh Kumar Rai of Indian Institute of Technology, Bombay for his constant motivation and active participation in preparation of the paper.
References 1. Bag., S., Ruj, S.: Key Distribution in Wireless Sensor Networks using Finite Affine Plane. In: AINA (2011) 2. C ¸ amtepe, S.A., Yener, B.: Key distribution mechanisms for wireless sensor networks:A survey 2005. Technical Report. In: TR-05-07 Rensselaer Polytechnic Institute, Computer Science Department (March 2005) 3. Chakrabarti, D., Maitra, S., Roy, B.: A key pre-distribution scheme for wireless sensor networks: merging blocks in combinatorial design. International Journal of Information Security 5(2), 105–114 (2006) 4. Cooke, B.: Reed Muller Error Correcting Codes. In: MIT Undergraduate J. of Mathematics, MIT Press, Cambridge (1999) 5. Eschenauer, L., Gligor, V.D.: A key-management scheme for distributed sensor networks. In: ACM Conference on Computer and Communications Security, pp. 41–47 (2002) 6. Gura, N., Patel, A., Wander, A., Eberle, H., Shantz, S.C.: Comparing Elliptic Curve Cryptography and RSA on 8-bit CPUs. In: Joye, M., Quisquater, J.-J. (eds.) CHES 2004. LNCS, vol. 3156, pp. 119–132. Springer, Heidelberg (2004) 7. Lee, J.Y., Stinson, D.R.: Deterministic key predistribution schemes for distributed sensor networks. In: Selected Areas in Cryptography. ser. Lecture Notes in Computer Scienc, pp. 294–307. Springer, Heidelberg (2004) 8. Lee, J.Y., Stinson, D.R.: A combinatorial approach to key predistribution for distributed sensor networks. In: IEEE Wireless Communications and Networking Conference, WCNC 2005, New Orleans, LA, USA (2005)
94
S. Bag, A. Saha, and P. Sarkar
9. Ruj, S., Roy, B.: Key predistribution using partially balanced designs in wireless sensor networks. In: Stojmenovic, I., Thulasiram, R.K., Yang, L.T., Jia, W., Guo, M., de Mello, R.F. (eds.) ISPA 2007. LNCS, vol. 4742, pp. 431–445. Springer, Heidelberg (2007) 10. Ruj, S., Roy, B.: Revisiting key predistribution using transversal designs for a grid-based deployment scheme. In: International Journal of Distributed Sensor Networks IJDSN 2009, vol. 5(6), pp. 660–674 (2009) 11. Steiner, J.G., Neuman, B.C., Schiller, J.I.: Kerberos: An authentication service for open network systems. USENIX Winter, 191–202 (1988) 12. Sarkar., P., Saha, A., Chowdhury, M.U.: Secure Connectivity Model in Wireless Sensor Networks Using First Order Reed-Muller Codes. In: MASS 2010, pp. 507–512 (2010)
A Cyclic-Translation-Based Grid-Quadtree Index for Continuous Range Queries over Moving Objects Hao Chen, Guangcun Luo, Aiguo Chen, Ke Qin, and Caihui Qu Department of Computer Science, University of Electronic Science and Technology, 611731 Chengdu, China {chenhao,gcluo,agchen,qinke,qucaihui}@uestc.edu.cn
Abstract. To speed up the processing of continuous range queries over moving objects data streams, various query indexing techniques have been proposed. We present a cyclic-translation-based Grid-Quadtree query index and use the methods of query location translation and index tuning to optimize the indexing of continuous range queries. We study the performance of our indexing approach and compare it with a prior CES-based indexing approach. Keywords: Continuous range queries, moving objects, data streams, cyclic translation, query indexing.
1
Introduction
With the continuous development of sensor networks and mobile devices, location-based services(LBSs) such as navigation and information services, emergency services, and object tracking services, are developed in many application systems. In recent years, large amount of research effort are paid to locationdependent query processing [1], which is a fundamental building block of LBSs. Continuous range query is an important type of location-dependent queries over spatio-temporal data streams, which must be reevaluated continuously until it is canceled by the user, and therefore is a useful tool for monitoring frequently changing location of moving objects. There have been extensive researches on evaluating a large number of continuous range queries over moving objects data streams [2],[3],[4],[5],[6]. Continuous range query can be classified into ”location report” and ”exist report” queries, depending on whether it concerns about obtaining the current locations of moving objects or just the set of moving objects that satisfy the query conditions. In this paper, we focus on both ”location report” and ”exist report” queries. Query indexing is a common way to speed up the processing of continuous range queries, since a brute-force approach which evaluates each of all queries for each incoming data tuple is inefficient. To optimize query processing further, a few kinds of virtual constructs (V Cs) for building query indexes were proposed, including virtual construct rectangles (V CRs) [4] and containmentencoded squares (CES) [5]. Particularly, K.-L.Wu [5] adopted a CES-based ¨ A. Ozcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 95–109, 2011. c Springer-Verlag Berlin Heidelberg 2011
96
H. Chen et al.
indexing technology, in which each range query is decomposed into quadtree blocks and the query ID is inserted into the ID lists associated with these decomposed quadtree blocks. By this way, locating the set of range queries that cover a given data point becomes simple as it’s quite easy to locate the set of quadtree blocks that cover the point, and then the total query evaluation time is cut down. However, the shortcoming of quadtree decomposition is the sensitivity of its storage requirements to the position of range queries. Different positions of range queries greatly influence the cost of the quadtree [7],[8],[9]. A square window of 2d × 2d may be decomposed into just one quadtree block or as many as 3(2d+1 − d) − 5 blocks [10], by placing the window at different positions in a grid area. Therefore, it’s highly desired to locate an optimal position translation for range queries so that the total number of quadtree blocks decomposed by quadtree can be minimized. What’s more, in the situation of skewed query distribution, the query evaluation of the CES-based indexing is not efficient, since the query index is implemented with pointer arrays of virtual constructs and constant times of accessing to the index are needed whenever a location update of moving object is received for query processing. In this paper, we present a Cyclic-Translation-based Grid-Quadtree for range query indexing. Based on this index structure, we designed an improved search algorithm to find out optimal translation of range queries for quadtree presentation. In this way, the number of decomposed quadtree blocks is minimized and quadtree nodes for range queries are much fewer, and the storage cost of the index is decreased a lot. What’s more, we tune the division level of Grid-Quadtree (a parameter of the index) to decrease total storage cost of the index further, with less query processing time. We conduct experiments to show the effectiveness of our approach, and compare it with the CES-based indexing method [5]. The results reveal that our approach outperforms the CES-based indexing in terms of storage cost and query evaluation time.
2
Related Work
Our work is mainly related to research domains of data stream processing, location-based systems, and spatio-temporal database. In the domain of data stream processing, a lot of data stream management systems (DSM S) have been designed for the continual evaluation of queries over potentially infinite data streams, including Aurora [13], Borealis [14], P IP ES [15], ST REAM [16], N iagaraCQ [17], T elegraphCQ [18]. The majority of these works focuses mainly on processing continuous queries over traditional data streams, and the spatial and temporal properties of data streams are overlooked. In recent years, a few DSM Ss over spatio-temporal data stream have been prototyped (e.g., P LACE [19], CAP E [20]), and there have been some researches on continuous queries over spatio-temporal streams (e.g., GP AC [20], SOLE [6]). GP AC [20] is designed to deal only with the execution of a single continuous query, while SOLE [6] is designed for the evaluation of concurrent continuous spatio-temporal queries. However, SOLE only supports exist report query,
A Grid-Quadtree Index for Continuous Range Queries over Moving Objects
97
without concerning about location report query. Moreover, the simple grid structure in SOLE is not optimal for large scale queries. In some location-based systems, users are usually interested in the changes of the locations rather than the details of the locations. The location-change events are useful to automatically trigger or stop necessary actions. A query processing framework named BM Q-Processor is proposed in [22] for large-scale bordercrossing event monitoring applications where query results are reported only when moving objects cross the borders of the query ranges. BM Q-Processor uses a novel index for Border Monitoring Query (BM Q) which has similar semantics as ”exist report” range query, and exploits the locality of data streams to achieve both quality search performance and efficient memory consumption. However, BM Q-Processor does not have special mechanisms to support ”location report” queries, and also couldn’t be extended to support irregular border-monitoring range query. In the domain of spatio-temporal database, different index structures for supporting fast spatio-temporal query have been proposed .e.g, B-tree, Quadtree, R-tree, Grid, KD-trie. An overview of spatio-temporal access methods is given by LV Nguyen-Dinh [23]. Index structures for moving objects can be classified according to a variety of criteria. From temporal consideration, there are indexing methods for indexing the past, current and future positions of moving objects. What’s more, specific indexing mechanisms have been proposed for objects moving in road networks, for objects with predefined trajectories, or for broadcast environments. Most of these works study snapshot queries and focus mainly on materializing incoming spatio-temporal data in disk-based index structures. Some other works focus on main-memory indexing and query processing over moving objects (M OV IES [8]). In the context of spatio-temporal data streams, only in-memory solutions are allowed. Therefore, in this paper, we implemented our index structure and query processing as a query operator based on our prototype spatio-temporal DSM S.
3 3.1
Preliminaries and System Model Background Assumptions
Grid of Monitoring Area. Consider a monitoring area with it’s bounding rectangle of size Lw × Lh . We construct a global grid A of 2N × 2N such that the bounding rectangle of monitoring area abuts the southwest corner of A and 2N −1 < max(Lw , Lh ) ≤ 2N . Given the global grid A, the corresponding region quadtree can recursively subdivide A into unit cells by N levels (or times) of quad-splitting. For any splitting level N − d, the size of resulting quadtree blocks(quadblock for short) is 2d × 2d . We call d the dimension of a quadblock and denote by B(x, y, d) a quadblock with dimension d and lower left corner at(x,y). Moving Objects Data Streams. Each tuple of a data stream of moving objects locations has the form τ (OID, x, y, t), where OID is the unique identifier of an moving object, x and y are its horizontal and vertical coordinates, which define
98
H. Chen et al.
the current position of moving object. Object locations can be anywhere in the monitoring area. t is the time stamp of object position. Continuous Range Query. A continuous range query Q can be represented as (QID, R, f lag),where QID is the query identifier, R is the query range specified as rectangle with vertices falling on the crossing point of grid lines, and flag is used to identify two types of range query(1 for location report query and 0 for exist report query ), deciding either to output the current locations of moving objects, or only the OIDs of objects, with respective output of stream τ (OID, x, y, t) or τ (OID, ±, t). Here, ± indicates whether this tuple is a positive or negative update. We use R(x, y, w, h) to denote a query range with width w, height h, and bottom-left corner at(x,y). We denote by RS(= {R1 , R2 , . . . , RM }) a set of query ranges which could be intersected or disjoint. 3.2
The Framework of Query Indexing
We use a combined data structure of grid and quadtree for range queries, which is called Grid-Quadtree in this paper. The monitoring area of size Lw × Lh is w treated as a conceptual Grid where the Grid contains L × 2LKh quadblocks 2K and each quadblock is treated as a region quadtree containing 2K × 2K cells. We call K the division level of the Grid-Quadtree. For two extreme cases, the Grid-quadtree is a quadtree when K is equal to 0, and a uniform grid when K is equal to N . Fig. 1 shows an example of a global grid area 24 × 24 with K = 2, and illustrates the overall design of the Grid-Quadtree index. The Grid-Quadtree is implemented with a pointer array of size 42 ,where each element corresponds to a quadblock spatial area of 22 × 22 . Each element within the pointer array stores a pointer to a dynamically maintained region quadtree corresponding to the square spatial area. There are two types of nodes in quadtrees. The internal node correspond to a quadblock (with dimension larger than 0 and not larger than K), and contains one pointer to a QID list of range queries and another four pointers to four children nodes of SW, NW, SE, NE direction. The leaf node corresponds to grid cells at the lowest level of quadtree and contains only a pointer to a QID list. With this index,each range query R in RS is decomposed into quadtree blocks, and the query ID is inserted into the QID lists of quadtree nodes corresponding to those blocks. The storage cost of the index is decided by three components: the number of decomposed quadblocks (or the number of QIDs in QID lists), the size of pointer array, and the number of quadtree node. The query efficiency of this index is decided by the depth of quadtree. To decrease the index storage cost and improve query efficiency, we optimize the indexing presentation of range queries by two methods. Firstly, we cyclically translate query ranges at location to cut down the number of decomposed quadblocks to minimum since the storage requirement of quadtree decomposition is sensitive to the position of ranges. Secondly, we compute the right division level K of the index structure, to minimize the total storage cost of the index and get a better query efficiency at the same time. Regarding the real time requirement of object tracking
A Grid-Quadtree Index for Continuous Range Queries over Moving Objects
99
*ULG4XDGWUHH
S
S
S
3RLQWHU$UUD\
,QWHUQDO1RGH 4,'/LVW 375
4XDG%ORFN
8QLW&HOO
TLG
S
S
S
TLG
TLG
4,'/LVW375 TLG
TLG
S
/HDI1RGH TLG
Fig. 1. A grid Area of 24 × 24 and the Grid-Quadtree data structure
applications, the processing of range queries in our work is realized as a customized query operator in DSM S. The query processing is data-driven, triggered by the underlying active data sources in a push-based fashion. On arrival of a batch of new object-location stream tuples, each tuple is used to search the query index to find all related range queries that contain the object location. Then, for those location report queries, the tuple with form τ (OID, x, y, t) is pushed into output data streams associated with the identified location report queries. For those exist report queries, more computing steps are needed to decide whether or not to output tuple of form τ (OID, ±, t), and where to output negative or positive tuple.
4 4.1
Tunable Cyclic-Translation-Based Grid-Quadtree Indexing Cyclic Translation of Region Queries
Here, we optimize the indexing of RS by cyclically translating all range queries at locations, to minimize the number of decomposed quadtree blocks and decrease the index storage cost. If we translate RS in all possible locations in global
100
H. Chen et al.
grid and compare the number of decomposed quadblocks, the time complexity of searching optimal translation is high. To improve the search algorithm, we propose a scheme based on the following definition and lemmas. Definition 1(Cyclic Translation of Query Range). For a range R(x, y, w, h) in global area of 2N × 2N , we translate(move) R with magnitude Xtm and Ytm in eastern and northern direction respectively. A range translation is called a cyclic one if the final position of an unit cell located originally at (x’, y’) in this range, after translation, can be derived from the following function: CT (x , y , Xtm , Ytm ) = ((x + Xtm ) mod 2N , (y + Ytm ) mod 2N ). With the cyclic translation of range queries, we can construct a region quadtree of size 2N × 2N to decompose all translated range queries without the need of expanding the quadtree to the size of 2N +1 × 2N +1 . From the property of region quadtree and cyclic translation, we present the following result without proof. Lemma 1. Translating a range embedded in 2N × 2N grid cyclically by 2d grid cells in any direction, where d < N , does not change the number of decomposed quadblocks with a size less than 2d × 2d . From lemma 1, we can derive the minimal searching space for finding the optimal position of a range query in quadtree decomposition. Lemma 2. An optimal position for decomposing a range R into minimal quadtree blocks could be gotten by translating R(x, y, w, h) by less than 2d units to the east and 2d units to the north, where d=min(logw ,logh ). Proof. Since the largest quadblocks possibly decomposed from R is of size 2d × 2d ,we can have the same number of decomposed blocks when we translate R with magnitudes of 2d+1 .It’s easy to see that all blocks of size 2d × 2d in R are always arranged in a line. Therefore, translating R with magnitudes of 2d won’t change the number of blocks of size 2d × 2d , the same as those blocks no more than 2d−1 × 2d−1. We thus have the proof. From lemma 1 and 2, we know that the sample solution space of optimal translation for RS is of size 2D × 2D , where D=max(d1 ,d2 ,. . . ,dM ), dm is the dimension of each range Rm in RS, and we can derive a basic search algorithm for finding the optimal position of minimizing decomposed quadtree blocks. The algorithm is constructed by a sequence of translations to RS, with magnitudes from 20 to 2D−1 in four different directions: no movement, north, east, and northeast. Since the translations with larger magnitudes would not affect the number of quadtree blocks with smaller magnitudes, we adopt the strip-splitting-based decomposing method [11] to recursively strip-split each range in RS while evaluating the sum of stripped quadtree blocks. Specifically, at ith level of translation, the search process proceeds by translating RS four times: by 0 cell (no movement), by 2i cell to the north, by 2i cell to the east, and by 2i cell to the northeast. After each translation, each range of RS is stripped off some slices of size c × 2i × 2i where c is a positive integer. Then, we recursively do the translation at next level
A Grid-Quadtree Index for Continuous Range Queries over Moving Objects
101
on these remaining new ranges. Recursion halts when all ranges are stripped to null. Thus, the translation magnitudes for the optimal location in two directions, i.e. Xtm and Ytm , can be easily acquired by the above process in a sequence of binary digits. To simplify the searching process, we use a pruning mechanism of ”candidate” to cut down unnecessary paths of recursions. For each translation, instead of performing translations in all next levels, we calculate the total number of quadtree blocks stripped so far, which is computed from the bottom level of strip-splitting to the level that being processed, and then compare it with the value of the current candidate which is stored in the least block counter. If the total number of blocks obtained so far is larger, the recursions of further levels are cut down.In order to obtain a candidate close to the optimal solution, we used an iterative process described in Algorithm 1 to quickly derive the first candidate. Here, we assume that M (x, y) = x mod y.
Algorithm 1. FindInitialCandidate. Require: RS(={R1 , R2 , . . . , RM }); /*a set of M query ranges*/ Ensure: least(the number of quadblocks derived from decomposition at candidate location); 1: least = 0; 2: for (d = 0; d < D; d + +) do 3: for (k = 0; k ≤ 3; k + +) do 4: RSk ← RS; Ck = Ck + RangesSplittingShrink(RSk , k); 5: end for 6: K = argmink (Ck ); least = least + CK ; RS ← RSK ; 7: end for 8: if RS = φ then least = least + RangesSplitShrink(RS, 0); 9: 10: f unction : RangesSplitShrink(RS, k) 11: i = k/2; j = M (k, 2); 12: for each Rm (xm , ym , wm , hm ) in RS do 13: xm = xm + i; ym = ym + j; 14: if M (xm , 2) = 0 then C = C + hm ; xm = xm + 1; wm = wm − 1; 15: if M (ym , 2) = 0 then C = C + wm ; ym = ym + 1; hm = hm − 1; 16: if M (xm + wm , 2) = 0 then C = C + hm ; wm = wm − 1; 17: if M (ym + hm , 2) = 0 then C = C + wm ; hm = hm − 1; 18: xm = xm /2; ym = ym /2; wm = wm /2; hm = hm /2; 19: if wm = 0 or hm = 0 then delete Rm from RS; 20: end for 21: return C;
Algorithm 1 takes RS as input, and conducts a sequence of translations with magnitudes from 20 to 2D−1 in four different directions. For each iteration, four moves in different directions are performed, some slices are strip-splitted away from each range, the number of quadtree blocks are computed, and then the
102
H. Chen et al.
translated RS with the smallest number of quadtree blocks is selected for the next iteration. After the strip-splitting of each iteration, we can shrink these remaining ranges since the same number of blocks is maintained while making the granularity coarser [12]. After at most D iterations, the initial candidate is identified to establish the bound. Obviously, the above algorithm takes O(D) time to find the initial candidate. We know from [12] that, if a range is a magic rectangle (it’s width w and height h are of the form 2i − 1 and 2j − 1, the number of its decomposed blocks is independent of the position of the anchor. Therefore, we can get rid of magic ranges from RS beforehand. The integrated pseudocode of locating the optimal translation for range queries is described in Algorithm 2. The pruning process is accelerated by calculating the initial candidate and excluding all the search paths rooted by a configuration whose number of decomposed quadtree blocks so far is already greater than that of current candidate. The worst time complexity of this improved algorithm is O(4D ).
Algorithm 2. LocatingOptimalTranslation. Require: RS(={R1 , R2 , . . . , RM }); /*a set of M query ranges*/ Ensure: Xtm , Ytm (translation magnitudes); 1: d = 0; BN[D + 1] = {0,. . . ,0};Xm = Ym = 0; 2: least = F indInitialCandidate(RS); 3: OptimalSearch(RS, Xm , Ym , least, BN, d); 4: 5: f unction : OptimalSearch(RS, Xm , Ym , least, BN, d) 6: for (k = 0; k ≤ 3; k + +) do 7: i = k/2; j = M (k, 2); RS ← RS; BN ← BN ; 8: BN [d] = RangesSplitShrink(RS , k); d d 9: Xm =d Xm + i × 2 ; Ym = Ym + j × 2 ; 10: if i=0 BN [i] ≤ least then 11: if (d < D − 1) then 12: OptimalSearch(RS , Xm , Ym , least, BN , d + 1); 13: else 14: if RS = φ then BN [D] = RangesSplitShrink(RS , 0); 15: BNtotal = D i=0 BN [i]; 16: if BNtotal < least then least = BNtotal ; Xtm = Xm ; Ytm = Ym ; 17: end if 18: end if 19: end for
4.2
Deciding the Division Level of Grid-Quadtree
To cut down the total storage cost of query index, we present a method to decide the right division level of the Grid-Quadtree. The total storage cost of the index consists of three components: a)the array of pointers to root nodes of quadtrees;
A Grid-Quadtree Index for Continuous Range Queries over Moving Objects
103
b)the nodes of quadtrees, and c)the QID lists, one for each quadblock defined. We denote respectively the number of quadblocks and quadtree nodes at different level by B[N ] and N [N ], which can be computed and derived by Algorithm 2. Given a division level K, the total cost of the index can be calculated as follows: w CSa = cp × L × 2LKh ; 2K CSb = cleaf × N [0] + cinternal ×
K i=1
N [i];
K N−K−1 CSc = cqid × ( i=0 B[i] + 4N −K−i × i=0 B[N − i])); CStotal = CSa + CSb + CSc ; where CSa , CSb ,CSc present the storage cost of the pointers array, quadtree nodes and query ID lists, respectively. The cp , cleaf , cinteral , and cqid present the size of a pointer, a quadtree leaf node, a quadtree internal node, and an query ID, respectively. From the above formula, we can get the right K which makes the index storage minimized as follows: K = argmink (CStotal ), k ∈ (0, 1, 2, .., N ) After choosing the right value of K, we can re-calculate the translation magnitudes of range queries as follows: Xtm = M (Xtm , 2K ); Ytm = M (Ytm , 2K ) 4.3
Cyclic Decomposition and Indexing Building of Range Queries
After finding the optimal translation and deciding division level K, we execute the location translation to query ranges actually and use a strip-splitting-based algorithm modified from [11] to strip quadblocks off translated query ranges. For those stripped quadblock, the corresponding query ID is inserted into the QID list of quadtree node associated with those quadblocks. To do this, the grid partition ID or the corresponding subscript of pointer array of grid for the quadblock must be decided at first for index building. To confine the size of w pointer array with L × 2LKh ,we compute the partition ID of a quadblock as 2K follows: w 0 if 2tK = L Lw 2K . P ID = f (y) × 2K + f (x), f (t) = t 2K otherwise Once the partition ID or subscript of pointer array is inferred, we get the right quadtree and could do the corresponding operation of index building or insertion. Z-order is used to identify quadblocks of same dimensions and help to search the place of corresponding tree node. The z-order of a grid cell with lower left corner coordinates (x, y) can be denoted as Z(x, y, 0) and computed by interleaving the bits of x and y. It should be pointed out that, if the anchor coordinates of any decomposed quadblock are beyond the scope of global grid, we cyclically translate it’s anchor coordinates (x,y) by computing M (x, 2N ) and M (y, 2N ) respectively and calculate the z-order of new coordinates at first, and then do the index insertion with the help of it’s z-order. We denote by b(x, y, i) a quadblock of dimension i with lower left corner coordinates (x, y), and its z-order can be
104
H. Chen et al.
represented as Z(x, y, i) = Z(x, y, 0)/4i . It’s corresponding tree node is the (Z(x, y, 0)/4i−1 − 4 × Z(x, y, 0)/4i )th child of parent node.
5
Query Processing of Continuous Regions Queries
The processing of query is realized as a customized query operator in (DSM S). To support exist report or cross-border monitoring query which has tuple output form of τ (OID, ±, t), a hash table is maintained, which uses OID as the index key and stores in each entry the last updated location of the moving object. The query search algorithm supporting both location report and exist report queries is presented below in Algorithm 3. Whenever receiving location updates of moving objects, the z-order of the quadblocks of all level containing the object locations are computed, and a query search is executed to find the queries covering these blocks. We assume that v.QL denotes the QID list associated with node v, T Q(q) denotes the output data stream of a query q, and P A[] denote the pointer array of global grid. Algorithm 3 works as follows. For each tuple τ (OID, x, y, t), the coordinates (x,y) are cyclically translated using the offsets(Xtm ,Ytm ). Then, the algorithm calculate the z-order and partition ID of two cells covering object’s current location and lastly updated location, respectively. With partition ID and z-order of these two cells, we could get the pointer to root node of corresponding quadtree and find all tree nodes storing the QIDs of covering queries, by taking advantage of the containment encoding of z-order. For those location report queries, we output τ (OID, x, y, t) into each data stream queue T Q(q) for each query contained in the QID lists associated with the quadblocks that cover new location. For those exist report queries, a top-down search of two paths along with one or two quad-trees is executed. At each level of quadtree search, we compare the two pointers to tree nodes associated with object’s current and old locations, to check whether the object has moved into a different quadblock. If the object move into a different quadblock, we need to compare QID lists corresponding to the new quadblock and it’s related child quadblocks with QID lists corresponding to the old quadblock and related child quadblocks, to compute two differential sets of queries which identify whether a moving object enter into or go out some query ranges.Then, we insert τ (OID, +, t) into the T L(q)s for queries in ”enter into” differential query sets and insert τ (OID, −, t) into the T Q(q)s for queries in ”go out” differential query sets, respectively.
6
Experimental Results
We conduct a few experiments to evaluate and compare our solution(we call it CT GQ-based indexing) with CES-based indexing. We assume a monitoring area with the same size as a grid of 210 × 210 .A total number of |Q| continual range queries were registered in the query index. The width and height of query ranges were randomly chosen between 23 and 28 . The bottom-left corners of range queries were distributed according to an α − β rules as in [5]: α fraction
A Grid-Quadtree Index for Continuous Range Queries over Moving Objects
105
Algorithm 3. ContinuousRangeQueryEvaluation. Require: Tuples(a batch of moving objects location data); Xtm , Ytm (optimal translation magnitude at x-axes and y-axes); Ensure: tuples of output data streams τ (OID, x, y, t), τ (OID, +, t), τ (OID, −, t); 1: for each tuple τ (OID, x, y, t) do 2: x = x + Xtm ; y = y + Ytm ; i = K; x = M (x , 2N ); y = M (y , 2N ); 3: Znew = Z(x , y , 0); Zold = Z(xold , yold , 0); Lw 4: P ID = f (y ) × 2K + f (x ); p ← P A[P ID]; w 5: P IDold = f (yold ) × L + f (xold ); pold ← P A[P IDold ]; 2K 6: while p = pold do 7: v1=node pointed by p; v2=node pointed by pold ; 8: output τ (OID, x, y, t) into T Q(q), for any q.QID ∈ v1.QL and q.f lag = ”1”; 9: m = Znew /4i−1 − 4 × Znew /4i ; n = Zold /4i−1 − 4 × Zold /4i ; i = i − 1; 10: follow p to the mth child of v1;follow pold to the nth child of v2; 11: end while 12: i1 = i; i2 = i; 13: while p = φ do 14: v1=node pointed by p; 15: output τ (OID, x, y, t) into T Q(q), for any q.QID ∈ v1.QL and q.f lag = ”1”; 16: insert QID into QSETnew for any q.QID ∈ v1.QL and q.f lag = ”0”; 17: m = Znew /4i−1 − 4 × Znew /4i ; i1 = i1 − 1;follow p to the mth child of v1; 18: end while 19: while pold = φ do 20: v2=node pointed by pold ; 21: insert QID into QSETold for any q.QID ∈ v2.QL and q.f lag = ”0”; 22: n = Zold /4i−1 − 4 × Zold /4i ; i2 = i2 − 1;follow pold to the nth child of v2; 23: end while 24: output τ (OID, +, t) to T Q(q), for q.QID ∈ QSETnew − QSETnew ∩ QSETold ; 25: output τ (OID, −, t) to T Q(q), for q.QID ∈ QSETold − QSETnew ∩ QSETold ; 26: end for
of the bottom-left corners were located within β fraction of the monitoring area, where β = 1 − α. A total of |O| moving objects are tracked at the same time and object locations are up-dated 10 times each minute. We conducted our simulations over a platform of DSM S. 6.1
Comparison of Index Storage Cost
Firstly, query regions were uniformly distributed(α = 0.5). We varied |Q| from 200 to 2000. Fig. 2a shows the number of decomposed blocks and tree nodes(V Cs for CES) with different |Q|. Under all cases, our method gets much less elements in QID lists and tree nodes(V Cs) than CES-based method. For the second experiment, |Q|=1000. The division level K is derived from formula of section 4.2 for both approaches. We varied α from 0.5 to 0.9 (0.9 represent more skewed distribution). Fig. 2b shows the total storage cost of both indices. It is clear
106
H. Chen et al.
that our index outperforms CES-based index greatly in skewed distribution, as more significant as the query positions become more skewed. The reason is that much more decomposed blocks are overlapped and much less tree nodes are constructed for the index in skewed distribution of query regions. 5
x 10
1800
10
1600
8
6
4
CES−QIDs CTGQ−QIDs CES−VCs CTGQ−tree nodes
2
0 200
400
600
800
1000
1200
1400
1600
1800
2000
Total storage cost of index(KB)
Number of query IDs and tree nodes
12
1400
1200
1000
800
CES−total cost CTGQ−total cost 600 0.55
0.6
Number of range queries (a)
0.65
0.7
0.75
0.8
0.85
0.9
0.95
Query distribution with different value of a (b)
Fig. 2. (a)number of QIDs and quadtree nodes(VCs) with different |Q|; (b)total storage cost in different distribution
6.2
Comparison of Query Evaluation Time
For the first experiment, |Q|=1000, α=0.5. The division level K is derived from formula of section 4.2 for both approach. We varied |O| from 4000 to 40000. Fig. 3a shows the impact of |Q| on the query evaluation time. Our approach outperforms CES-based indexing in query evaluation time. This is because the average depth of Traversing tree nodes during an index search is much less for our approach, with the decrease of total number of tree nodes. Such a performance advantage becomes prominent as the number of moving objects increases. For the second experiment, |Q|=1000,|O|=20000. We varied from 0.55 to 0.95. Fig. 3b shows the impact of query distribution on the query evaluation time. Under all cases, our index outperforms CES-based index in skewed distribution, as more significant as the query positions become more skewed. It’s clear that query evaluation efficiency for our approach is high when objects move in quad block regions without covering of queries. This is because no tree nodes are constructed in the index, corresponding to these regions. 6.3
Impact of K on Index Storage Cost and Query Evaluation Time
In this experiment, |Q|=1000, |O|=20000, α=0.5. We varied K from 1 to 10. Fig. 4a and 4b show the impact of K on the index storage cost and query evaluation time respectively. As K increases, the total storage cost of the index decreases first, and then increases after reaching the lowest point. Query evaluation time always increases as K increases. This is because the depth of trees is increased
A Grid-Quadtree Index for Continuous Range Queries over Moving Objects 6
107
2.7
CES−based indexing CTGQ−based indexing
CES−based indexing CTGQ−based indexing
2.6
Query evaluation time(seconds)
Query evaluation time(seconds)
5
4
3
2
2.5
2.4
2.3
2.2
2.1
1 2
0 0
0.5
1
1.5
2
2.5
3
3.5
4
Number of moving objects (|O|) (a)
x 10
1.9 0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
Query distribution with different value of a (b)
4
Fig. 3. (a)query evaluation time over different |O| at uniform distribution; (b)query evaluation time over different query distribution
with the increase of K and then the number of tree nodes traversed during an index search are increased too. In this example, the K with least storage cost is 5. Actually, we could choose the division level of the index with a value smaller than 5, such as K=3 or 4, to further decrease the query evaluation time with the cost of modest increase of storage cost.
16000
2.3
Query Evaluation Time
Total storage cost 2.2
Query evaluation time(seconds)
Total storage cost of the index(KB)
14000
12000
10000
8000
6000
2.1 2 1.9 1.8 1.7 1.6 1.5
4000 1.4 2000
1
2
3
4
5
6
7
The division level of index (K) (a)
8
9
10
1.3
1
2
3
4
5
6
7
8
9
10
The division level of index (K) (b)
Fig. 4. The impact of K on (a) total index storage. (b) query evaluation time
7
Conclusions
We have presented a Cyclic-Translation-based Grid-Quadtree data structure for indexing continuous range queries, and designed related search algorithm to find out optimal translation of range queries for quadtree presentation. experiments show that our approach outperforms the CES-based indexing method.
108
H. Chen et al.
References 1. Ilarri, S., Mena, E., Illarramendi, A.: Location-dependent query processing:Where we are and where we are heading. ACM Computing Surveys 42(3), 1–73 (2010) 2. Kalashnikov, D.V., Prabhakar, S., Aref, W.G., Hambrusch, S.E.: Efficient Evaluation of Continuous Range Queries on Moving Objects. In: Proc. Int’l Conf. Database and Expert Systems Applications (2002) 3. Wu, K.-L., Chen, S.-K., Yu, P.S.: Processing Continual Range Queries over Moving Objects Using VCR-Based Query Indexes. In: Proc. IEEE Int’l Conf. Mobile and Ubiquitous Systems: Networking and Services (August 2004) 4. Wu, K.-L., Chen, S.-K., Yu, P.S.: Efficient Processing of Continual Range Queries for Location-Aware Mobile Services. Information Systems Frontiers 5(4-5), 435–448 (2005) 5. Wu, K.-L., Chen, S.-K., Yu, P.S.: Incremental processing of continual range queries over moving objects. IEEE Trans. Knowl. Data Eng. 8(11), 1560–1575 (2006) 6. Mokbel, M., Aref, W.: SOLE: scalable on-line execution of continuous queries on spatio-temporal data streams. VLDB J., 971–995 (2008) 7. Li, M., Grosky, W., Jain, R.: Normalized Quadtrees with Respect to Translations. Computer Graphics and Image Processing 20, 72–81 (1982) 8. Chen, S.-K.: An exact closed-form formula for d-dimensional quadtree decomposition of arbitrary hyperrectangles. IEEE Trans. on Knowledge and Data Engineering 18(6), 784–798 (2006) 9. Chen, P.-M.: A quadtree normalization scheme based on cyclic translations. Pattern Recognition 30(12), 2053–2064 (1997) 10. Dyer, C.R.: The Space Efficiency of Quadtrees. Computer Graphics and Image Processing 19(4), 335–348 (1982) 11. Tsai, Y.-H., Chung, K.-L., Chen, W.-Y.: A Strip-Splitting-Based Optimal Algorithm for Decomposing a Query Window Into Maximal Quadtree Blocks. IEEE Trans. Knowledge and Data Eng. 16(4), 519–523 (2004) 12. Faloutsos, Jagadish, H.V., Manolopoulos, Y.: Analysis of n-Dimensional Quadtree Decomposition of Arbitrary Rectangles. IEEE Trans. Knowledge and Data Eng. 9(3), 373–383 (1997) 13. Abadi, D.J., Carney, D., Cetintemel, U., Cherniack, M., Convey, C., Lee, S., Stonebraker, M., Tatbul, N., Zdonik, S.: Aurora: A New Model and Architecture for Data Stream Management. The VLDB J. 12(2), 120–139 (2003) 14. Abadi, D.J., Ahmad, Y., Balazinska, M., Cherniack, M., Hwang, J.H., Lindner, W., Maskey, A.S., Rasin, A., Ryvkina, E., Tatbul, N., Xing, Y., Zdonik, S.: The Design of the Borealis Stream Processing Engine. In: Proceedings of the 2nd Biennial Conference on Innovative Data Systems Re-search(CIDR), pp. 277–289 (2005) 15. Kramer, J., Seeger, B.: Semantics and Implementation of Continuous Sliding Window Queries over Data Streams. ACM TODS 34(1) ( April 2009) 16. Arasu, A., Babu, S., Widom, J.: The CQL Continuous Query Lan-guage: Semantic Foundations and Query Execution. The VLDB J. 15(2), 121–142 (2006) 17. Chen, J., DeWitt, D., Tian, F., Wang, Y.: NiagaraCQ: A Scalable Continuous Query System for Internet Databases. In: Proc. ACM SIGMOD (2000) 18. Chandrasekaran, S., Franklin, M.J.: Streaming Queries over Streaming Data. In: Proc. 28th Int’l Conf. Very Large Data Bases, VLDB (2002) 19. Mokbel, M.F., Aref, W.G.: PLACE: A Scalable Location-aware Database Server for Spatio-temporal Data Streams. Data Engineering Bulletin 28(3) (2005)
A Grid-Quadtree Index for Continuous Range Queries over Moving Objects
109
20. Rundensteiner, E.A., Ding, L., Sutherland, T.M., Zhu, Y., Piel-Ech, B., And Mehta, N.: CAPE: Continuous query engine with het-erogeneous-grained adaptivity. In: Proceedings of the 30th International Con-ference on Very Large Data Bases (VLDB 2004), pp. 1353–1356. Morgan Kaufmann, San Francisco (2004) 21. Mokbel, M.F., Aref, W.G.: GPAC: Generic and progressive processing of mobile queries over mobile data. In: Proceedings of the International Conference onMobile DataManagement, MDM (2005) 22. Lee, J., Kang, S., Lee, Y., Lee, S., Song, J.: BMQ-Processor: A High-Performance Border-Crossing Event Detection Framework for Large-scale Monitoring Applications. IEEE Transactions on Knowledge and Data Engineering (TKDE) 21(2) (February 2009) 23. Nguyen-Dinh, L.V., Aref, W.G., Mokbel, M.F.: Spatio-Temporal Access Methods: Part 2 (2003 - 2010). Bulletin of the Technical Committee on Data Engineering 33(2) (June 2010) 24. Dittrich, J., Blunschi, L., Salles, M.A.V.: Indexing moving objects using short-lived throwaway indexes. In: SSTD 2005, pp. 189–207 (2009) 25. Orenstein, J.A., Merrett, T.H.: A Class of Data Structures for Associative Searching. In: Proc. ACM Symp. Principles of Database Systems (April 1984)
Two-Stage Clustering with k-Means Algorithm Raied Salman*, Vojislav Kecman, Qi Li, Robert Strack, and Erick Test Virginia Commonwealth University, Computer Science Department, 601 West Main Street Richmond, VA 23284-3068
[email protected] Abstract. -means has recently been recognized as one of the best algorithms for clustering unsupervised data. Since the -means depends mainly on distance calculation between all data points and the centers then the cost will be high when the size of the dataset is big (for example more than 500MG points). We suggested a two stage algorithm to reduce the cost of calculation for huge datasets. The first stage is fast calculation depending on small portion of the data to produce the best location of the centers. The second stage is the slow calculation in which the initial centers are taken from the first stage. The fast and slow stages are representing the movement of the centers. In the slow stage the whole dataset can be used to get the exact location of the centers. The cost of the calculation of the fast stage is very low due to the small size of the data chosen. The cost of the calculation of the slow stage is also small due to the low number of iterations. Keywords: Data Mining, Clustering, -means algorithm, Distance Calculation.
1 Introduction No theoretical research work available on the running time was required for the means to achieve its goals as mentioned by [1]. They researched the worst-case running time scenario as superpolynomial by improving the lower bound from Ω iterations to 2Ω √ . [9] has developed another method to reduce the number of iterations but it was not as fine-tuned as [1]. On the other hand [4] have proved that the number of iterations required by -means is much less than the number of points. Moreover, [5] were unable to bound the running time of -means, but they proved that for every reclassified point one iteration is required. Then after convergence will be guaranteed.
Δ
iterations the
A group of researchers worked on choosing the best centers to avoid the problems of Means of either obtaining the non-optimal solutions or empty clusters generations. [3] worked on modifying the -means to avoid the empty clusters. They moved the center of every cluster into new locations to ensure that there will be no empty clusters. The comparison between their modified -means and the original -means show that the number of iterations is higher with the modified -means method. In case of the numerical *
Corresponding author.
A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 110–122, 2011. © Springer-Verlag Berlin Heidelberg 2011
Two-Stage Clustering with k-Means Algorithm
111
examples which produce empty clusters, the proposed method cannot be compared with any other method since there is no modified -means algorithm available to avoid the empty clusters. [6] on the other hand developed a procedure in which the centers have to pass a refinement stage to generate good starting points. [7] used genetically guided means where the possibility of empty clusters will be treated in the mutation stage. Another method of center initializing based on values of attributes of the dataset is proposed by [8]. The later proposed method creates a complex procedure which leads to be computationally expensive. [2] on the other hand, developed a method to avoid unnecessary distance calculations by applying the triangle inequality in two different ways, and by keeping track of lower and upper bounds for distances between points and centers. This method is effective when the dimension is more than 1000 and also when the clusters are more than 20. They claimed that their method is many times faster than normal means method. In their method the number of distance calculations is instead of where is the number of points and are the number of clusters and the number of iterations respectively. [9] In contrast, Hodgson used different triangular equality to achieve the goal, in which they reduced the number of distance calculations.
2 Theoretical Background and the Proposed Method Simple modifications of -means clustering method have been proposed. The theoretical background of the proposed method is described below: The main idea behind -means is to calculate the distance between the data point and the centers using the following formula: /
,
Where the Euclidean distances between the data point initial centers are . The points in one cluster are defined as:
(1) at the cluster
and the
for 1,2, … , regarded as one cluster and is the total number of points in that cluster. The chosen randomly either from the dataset or arbitrarily. In our method we have used the random selection of the centers from the dataset to avoid wasting one more calculation (iteration). Any -means clustering method depends on the number of clusters set at the beginning. There is no guarantee that the centers will move or converge to the mean points of the average of the cluster. This is one of the drawbacks of -means. Also there is no guarantee that the convergence will happen to the local mean. is the set of clusters to minimize the criteria . ; so that Assume that converges to (the cluster centers): , where where
,
,…,
;
,…, min |
(2) |
is the probability distribution over the Euclidean space.
(3)
112
R. Salman et al.
If the represents the entire dataset then the objective is to find a subset of . such that We assume that the data with one center is a stationary random sequence satisfying the following cumulative distribution sequence: ,
,…,
,
,…,
,
,…,
,
,…,
(4)
then the above sequence has one mean: (5) The process of clustering is equivalent to minimizing the Within-Cluster Sum of Squares for the, so called, fast stage: min
(6)
and for the so called, slow stage, as follows: min
(7)
where are the centers of the clusters which are equals to the centers of the previous stage. The within cluster sum of squares is divided into two parts corresponding to the fast and the slow stages of the clustering: ,
,
(8)
The centers of the slow stage start with
3 The Results The complexity of the -means is where is the number of clusters, is the number of iteration required to get to the stopping criteria and is the input patterns. For example if the data size is 1000 points, 4 clusters and it require 20 iterations to get the optimal locations of the centers. Then, 80,000 is the time complexity. The time complexity in the proposed method has two folds, first is time complexity of the fast stage of clustering: where is the number of data for the fast stage and is the iterations during the fast stage only. The second part of the time complexity is calculated where is the number of according to the slow stage of clustering:
Two-Stage Clustering with k-Means Algorithm
iterations during the slow stage. Assume that complexity is: =
7200
100 and 8000 =
113
3 then the total time 15200
This will represent a reduction in the calculation time for the clustering of more than 5 times. However, if the data is bigger than the previous figure then the time reduction will be higher. For example if the data is 1Million the reduction will be approximately 10 times. This is quite well illustrated in the following diagram:
Fig. 1. Complexity measure of the k-means and the modified k-means with 100 samples
The and are the fast iterations and the slow iterations of the modified means. Respectively. The black graph in Fig. 1 is the time complexity of the normal -means. Other graphs represent the complexity of the modified -means. Therefore the higher the value of the more the graphs will approach the normal -means the properties. From the above graph it can be concluded that the lower values of less time required to achieve total clustering. The more iterations, for the fast stage, the faster the algorithm works. However, the catch here is we cannot go very low with as the time of the clustering will approach the normal -means. In other words the clustering procedure will produce blank clusters. The proper percentage would be 10% - 20%. The set up of the parameters of the red graph of the above diagram has a complexity of almost 5 times less than the normal -means clustering. In the case of 500, the complexity using higher number of data for the fast stage clustering, results will be skewed upwards as shown below:
114
R. Salman et al.
Fig. 2. Complexity measure of the k-means and the modified k-means with 500 samples
The set up of the parameters of the red graph of the above diagram has complexity less than 2 times than the normal -means clustering. This indicates that the more the data chosen for the fast stage of clustering the less advantages of this method.
4 Numerical Examples Two examples presented here to validate the proposed method. 1- A data set with 800 samples and 2-dimension (3 clusters) is used. The following figures show the movement of one of the centers and the two stage clustering. 2- From Figs. 3 and 4 it is very clear that the approach of the red line (slow stage coordinate of one center) is very smooth comparing with the other fast stage coordinate movements. The first value of the red (slow) graph is the same as the last value of the blue (fast) graph. The number of iterations is higher than is required but this is only for clarification. The number of iterations required for the fast stage will of course be higher than the slow stage scheme. Moreover, as you can see from the above graph, the coordinates have not been changed a lot. This means that the -means algorithm does not need to run many times since we reached the correct accuracy. Another presentation of the effectiveness of the method is the movements of the three centers as shown in figures 5-8.:
The following algorithm describes briefly the proposed procedure:
Two-Stage Clustering with k-Means Algorithm
Algorithm. Input: , , , , with clusters Output: % of from randomly Select While For 1 Calculate the modified distance , Find minimum of Assign the cluster number to point End for Calculate End while Calculate the average of the calculated clusters to find new centers Use the whole dataset While For 1 Calculate the modified distance , Find minimum of Assign the cluster number to point End for Calculate End while
Fig. 3. Fast and Slow stages of the movement of one coordinate during the clustering
115
116
R. Salman et al.
Fig. 4. Fast and Slow stages of the movement of the second coordinate during the clustering
Fig. 5. Three center movement during the fast stage clustering
Two-Stage Clustering with k-Means Algorithm
117
A more detailed description is shown in the following figures in which the fast stage shows the squares and the slow stage shows the diamond symbol:
Fig. 6. Fast and slow stages of the first cluster center movements
Fig. 7. Fast and slow stages of the second cluster center movements
118
R. Salman et al.
Fig. 8. Fast and slow stages of the third cluster center movements
As can be seen from the above diagrams, that the centers have moved many steps during the fast stage, this has been achieved in fast response. The diamond shapes shows the slow stage of iteration. The number of iterations of the slow stage is much less than the fast stage. Also the movements of the centers are very small. In this case the required calculation would be reduced from many steps to only couple of full step (for all dataset). This of course will save some time and reduce expenses. To be more specific about the proposed method Table 1 shows the creation of clusters in different iterations for three dimensional data. Table 1. Distribution of points and centers during the fast and the slow stages of clustering Iter No.
1 Fast 2
Clusters Old Centers
New Centers
C1
8
4
4
4.867
3.267
1.567
C2
4
4
4
6.16
2.85
4.68
C1
4.867
3.267
1.567
4.867
3.267
1.567
C2
6.16
2.85
4.68
6.16
2.85
4.68
Points in Clusters 30,38,44 53,58,72 86,88,93 113,114 138,145 30,38,44 53,58,72 86,88,93 113,114 138,145
Points
15
Two-Stage Clustering with k-Means Algorithm
119
Table 1. (continued) C1
4.867
3.267
1.567
5.015
3.318
1.636
C2
6.16
2.85
4.68
6.323
2.901
4.987
C1
5.015
3.318
1.636
5.06
3.226
1.897
C2
6.323
2.901
4.987
6.396
2.933
5.071
C1
5.06
3.226
1.897
5.083
3.205
1.956
1
2
Slow
3
C2
6.396
2.933
5.071
6.409
2.942
5.1
C1
5.083
3.205
1.956
5.083
3.205
1.956
C2
6.409
2.942
5.1
6.409
2.942
5.1
4
1-50,58 61,82,94 99 51-57,5960 62-81 83,93 95-98 100-150 1-50,54,58 60-61,70 80-82,90 94,99,107 51-53,5557 59,62-69 71-79 83-89 91-93 95-98 100-106 108-150 1-50,54,58 60-61 63 65,70,8082 90,94 99,107 51-53,5557 59,62,64 66-69,7179 83-89,9193 95-98 100-106 108-150 1-50,54,58 6061,63,65 70,8082,90 94,99,107 51-53,5557 59,62 64,66-69 71-79 83-89 91-93,9598 100-106 108-150
150
120
R. Salman et al.
The two stages are clearly indicating the formation of clusters at each stage. The centers of the slow stage are the same as the centers of the end of the fast stage. The fast stage has taken only 10% of the whole dataset. Although the number of iterations of the slow stage is 4, in bigger dataset this can be reduced by increasing the iterations of the fast stage.
5 Speed Up Analysis The speed of the normal -means is shown in blue while the speed of the modified means is shown in red. Two different computers were used of 32bit and 64bit Operating Systems. Regardless, of the speed of the computer used the validation of the modified -means always consistent as indicated by Fig. 9. The data used for the fast stage clustering is only 10%of the whole data which is randomly selected. The dataset used in this example is “Synthetic” which is 100,000 samples with 10 dimensions. The speed of the modified -means is almost twice the speed of normal -means. This is due to the fact that 2-stage -means clustering uses less full data iterations. The speed up is very clear in the high accuracy when the required µ is 10 or less, where µ is the stopping criteria or the required accuracy. This is always important when you try to find good clustering results.
Fig. 9. Comparison in the speed of the modified -means and normal -means with different computers
The speed up of the modified -means comparing with the normal -means is varying according to the accuracy. For the lower range of accuracy the speed up of
Two-Stage Clustering with k-Means Algorithm
121
clustering is ranges from (1-9) times. This would reduced for the higher accuracy for example from 10 to 10 . Figure 9 shows clearly that the speed up is settled for the higher accuracy within 2 times. On the other hand the range of the random data selected to archive the fast clustering is also fluctuating. The best range is between 10%-20%. In the normal situation we require a good accuracy for the clustering to archive the full clustering to all data. This would be between 10%-20% of the data as shown in Table 2. and accuracy between 10 to 10 Table 2. Speed up of clustering with the modified k-means using different dataset sample
Accuracy
percentage of the total data 10%
15%
20%
30%
40%
10
1.9
1.8
1.8
1.7
1.5
10
3.8
3.5
3.4
3
2.5
10
4.7
8.9
3.1
7
4.3
10
1
1.7
1.1
3
8.5
10
2.9
1.6
2.2
2.1
2.4
10
2
1.9
2.6
2.3
2.4
10
2
1.4
2.4
2.3
1.6
10
2
1.4
2.4
2.3
1.6
10
2
1.4
2.4
2.3
1.6
10
2
1.4
2.4
2.3
1.6
The proper range of the sample data is between 10%-20%. Carrying out the required time for running the normal -means and the modified -means for 9 different data samples shows that the best range is 10%-20% to get less time in the calculation of the two algorithms as shown in Table 3.
Percentage of the data
Table 3. Calculation time for normal kmeans and modified kmeans
10% 20% 30% 40% 50% 60% 70% 80% 90%
Fast + Slow k-means (sec) 4.81 9.93 14.94 19.95 25.17 30.45 36.62 42.01 47.66
Normal k-means (sec) 14.3 14.3 14.3 14.3 14.3 14.3 14.3 14.3 14.3
122
R. Salman et al.
6 Conclusion A simple proposal for achieving high speed of -means clustering for ultra dataset has been presented in this paper. The idea has two folds. The first is the fast calculation of the new centers of the -means clustering method. A small part of the data will be used in this stage to get the final destination of the centers. This of course will be achieved in high speed. The second part is the slow stage in which the -means will start from well positioned centers. This stage may take couple of iteration to achieve the final clustering. The whole dataset will be used for the second stage. In normal -means algorithm if the initial centers are exactly located at the means of the clusters of the data, then the algorithm requires only one step to assign the individual clusters to each data point. In our modified -means we are trying to get to that stage of moving any initial centers to a location which is either the locations of the means or near it. The big the difference between these locations will decide on how many times the normal -means required to run to assign all data to their clusters. Our algorithm will move the centers fast to the locations which are near the means. Future work is required to find out the effect of different locations of the clusters on the speed up.
References 1. Arhter, D., Vassilvitskii, S.: How Slow is the kMeans Method? In: SCG 2006, Sedona, Arizona, USA (2006) 2. Elkan, C.: Using the Triangle Inequality to Accelerate K –Means. In: Proceedings of the Twentieth International Conference on Machine Learning (ICML-2003), Washington DC (2003) 3. Pakhira, M.K.: A Modified k-means Algorithm to Avoid Empty Clusters. International Journal of Recent Trends in Engineering 1(1) (May 2009) 4. Dude, R.O., Hart, P.E., Stork, D.G.: Pattern Classification. Wiley-Interscience Publication, Hoboken (2000) 5. Har-Peled, S., Sadri, B.: How fast is the k-means method? Algorithmica 41(3), 185–202 (2005) 6. Bradley, P.S., Fayyad, U.M.: Refining Initial Points for Kmeans Clustering. Technical Report of Microsoft Research Center. Redmond,California, USA (1998) 7. Wu, F.X.: Genetic weighted k-means algorithm for clustering large-scale gene expression data. BMC Bioinformatics 9 (2008) 8. Khan, S.S., Ahmed, A.: Cluster center initialization for Kmeans algorithm. Pattern Recognition Letters 25(11), 1293–1302 (2004) 9. Hodgson, M.E.: Reducing computational requirements of the minimum-distance classifier. Remote Sensing of Environments 25, 117–128 (1988)
An Energy and Delay-Aware Routing Protocol for Mobile Ad-Hoc Networks Jihen Drira Rekik, Leïla Baccouche, and Henda Ben Ghezala RIADI-GDL laboratory, ENSI National school of computer sciences Manouba University, 2010 Manouba, Tunisia
Abstract. A mobile ad-hoc network (MANET) is an autonomous system of mobile nodes which are free to move randomly thus forming a temporary network. Typical applications of MANET are in disaster recovery operations which have to respect time constraint needs. However, it is very difficult to guarantee any quality of service to a real-time flow in such network because it must take into account the specificities of these networks. This paper introduces the Energy Delay aware based on Dynamic Source Routing, ED-DSR. ED-DSR efficiently utilizes the network resources such as the node energy and the node load in order to balance traffic load. It ensures both timeliness and energy efficiency by avoiding low-power node and busy node. Simulation results, using NS-2 simulator, show that the protocol prolongs the network lifetime (up to 66%), increases the volume of packets delivered while meeting the data flows real-time constraints and shortens the end-to-end delay. Keywords: efficient energy, mobile ad-hoc network, quality of service, realtime packet, routing protocol.
1 Introduction A Mobile Ad-hoc NETwork (MANET) has several advantages such as their autonomic and infrastructure-less properties. Mobile nodes can move and access data randomly at anytime and anywhere. There is no need for fixed infrastructure; it can be easily deployed anywhere and anytime. Mobile nodes in MANET such as PDA, laptops or Smartphone are connected by wireless links and each node acts as a host and router in the network. They are characterized by their reduced memory, storage, power and computing capabilities. Mobile nodes are classified into two groups: Small mobile hosts (SMH) which has a reduced memory, storage, power and computing capabilities and large mobile hosts (LMH) equipped with more storage, power, communication and computing facilities than the SMHs. Mobile ad-hoc networks have become increasingly popular due to their autonomic and infrastructure-less properties of dynamically self-organizing, self-configuring, self-adapting. However, a number of challenges like resource constraints, dynamic network topology are posed. MANETs cover a large range of applications from military operations, natural disaster and search-and-rescue operation where common wired infrastructures are not directly reachable to provide communication due to limited provision of this facility A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 123–134, 2011. © Springer-Verlag Berlin Heidelberg 2011
124
J. Drira Rekik, L. Baccouche, and H. Ben Ghezala
in those settlements. We focus especially on real-time applications where a number of them, including defense applications, have to respect time constraint in order to update positions of wounded soldiers or enemies or find medical assistance. Currently, most MANET research has focused on routing and connectivity issues [1] [5] in order to cope with the dynamism of such networks. Just solving the problem of connectivity is not sufficient for using MANETs. Since MANET comprises of resource limited nodes, these nodes demand utilization of data, services and cooperation of other nodes to perform tasks on time. How to guarantee the data exchange without exceeding its real-time constraints or wasting resources? What are the appropriate metrics combining timeliness on the one hand and on the other hand choosing the correct route without depleting resources? Based on dynamic source routing (DSR) [3], we introduce the Energy and Delay-aware Dynamic Source Routing protocol (ED-DSR) for MANET. ED-DSR is a routing protocol which uses information from the physical layer and the MAC layer in choosing routes, focusing on the energy efficiency, delay guarantee and the overall network performance. Simulation results show that ED-DSR outperforms the traditional routing protocol, DSR, in providing longer network lifetime and lower energy consumption per bit of information delivered. In addition, it minimizes the end-to-end delay and upgrades the rate of packets delivered. The rest of the paper is organized as follows: in the second section, we present the related work in QoS routing protocols. In the next section, we describe the proposed Energy Delay-aware Dynamic Source Routing (ED-DSR) protocol. Detailed analysis in performance difference is performed in next sections.
2 QoS Routing Protocols: An Overview The performance of the ad-hoc mobile network highly depends on the lifetime of mobile hosts. The network partition may lead to interruptions in communications, as in such conditions mobile nodes need to deliver their packets through intermediate nodes in order to intend destinations. Therefore, the lifetime of intermediate mobile nodes should be prolonged by conserving energy either at each node and for each connection request, too. In MANET, the mobile nodes are power limited and require energy for computing as well as routing the packets. Moreover applications in this environment are time-critical which require their flows to be executed not only correctly but also within their deadlines. In the literature lot of QoS aware routing protocols have been proposed [4], [6], [8], [9]and [10]. QoS routing is to compute routes for traffic flows with QoS requirements. To determine a route, QoS routing considers QoS requirements of the flow and resources availability, too. 2.1 Energy-Aware Multipath Routing Protocol, EMRP EMRP is an energy-aware multipath source routing protocol derived from Dynamic Source Routing (DSR) [9]. It makes changes in the phases of Route Reply, Route Selection and Route Maintenance according to DSR. EMRP utilizes the energy and queuing information to select better routes. In route response, each intermediate node will stamp its current status in the RREP packet. Finally, the routing agent at the source node will collect the RREP. In routes selection, EMRP chooses the working
An Energy and Delay-Aware Routing Protocol for Mobile Ad-Hoc Networks
125
set of routes from all available routes according to the following rules. First of all, EMRP calculates the cost of each available route according to the following equation: ∑
.
(1)
Where is the cost of the route and , are the costs of node i considering the energy and queue length respectively. and are the costing factors which normalize and . A route is selected based on minimum values is a function depending of the distance and remaining energy of node i of . depends on the queue length along the current route. and the next-hop node. Critics: The above solution proposes multipath routing protocol. It provides routes that reduce the intermediate mobile nodes power consumption, alleviating the network partitioning problem caused by the energy exhaustion of these nodes. However, the exhaustible energy battery is not the only indicator for route selection and a power control scheme. The number of packets in each node’s queue, along the route, doesn’t reflect the local processing time. In fact, each packet has its proper execution time which varies. Thus, the packet handling will inevitably suffer a longer delay and therefore the energy exhaustion of these nodes; while there are other nodes with less energy but where their queues require less time to be treated. The route selection should be done according to energy and more queuing information, in terms of queue length and local processing time of each previous flow, too. 2.2 Real-Time Dynamic Source Routing, RT-DSR RT-DSR is based on the expiration delay to deadline [10]. It makes changes in the phases of routes discovery and reply. In routes discovery, a route request RREQ is broadcasted with the expiration delay to deadline. The route request is accepted for a new flow only if the new packet can reach the destination before the expiration delay. 0.
(2)
Where is the remaining time of the expiration delay to deadline, for the traffic k, received from the node (i-1). is the local processing time of any message; is the transmission time between two neighboring nodes in the worst case remaining times. The delay of each real-time flow in the queue, already admitted, shouldn’t be altered by the newest one. ,1
,
0.
(3)
Where res is the number of real-time flows already admitted in the node. In routes reply, each intermediate node reserves the resources, saves the remaining time to deadline and sends the confirmation to the next node until reaching the source. Critics: RT-DSR purpose is to reserve resources in order to meet the deadlines but it must also take care that the resources are exhaustible. Indeed, choosing the same route to transfer all packets of real-time data through the reserved route may exhaust the energy of these nodes leading to the network partitioning problem. Moreover, the route selection criteria should consider that in MANET there are other traffics generated and they could take some joint nodes. The rules, under which packets are
126
J. Drira Rekik, L. Baccouche, and H. Ben Ghezala
assigned to route, should improve the system performance in terms of real-time constraint and energy efficiency too.
3 The Proposed Energy Delay-Dynamic Source Routing Protocol The proposed routing protocol considers packet deadline (real-time constraint), energy of the forwarding nodes and load at intermediate nodes to deliver real-time traffic. Each packet should be transmitted from the source to the destination within the deadline. The basic working of our proposed protocol is as follows. Each node, before starting the transmission of real-time data, selects a suitable route between the source and the destination. The selected route should satisfy delay requirements, preserve energy consumption and avoid overloaded nodes. Energy delay-dynamic source routing, ED-DSR, protocol is based on DSR. In DSR, the routes are stored in without any constraint on quality of services. The delay requirement is not considered to ensure that packets will reach their destinations before the deadlines. Furthermore, DSR doesn’t contribute to reduce the power consumption of mobile nodes. However, DSR is an on demand protocol ensuring the freshness of constructed route which is more suitable for the real-time flow. Therefore, we opt to DSR as based protocol in our work. DSR discovers a route between two nodes, only when required which reduce the number of packets control. DSR is simple and flexible [2] which facilitates the implementation of our extension. Also, a route response packet sent back to the source can be used to incorporate real-time and energy constraints. The choice of the suitable route to transfer the real time data in ED-DSR is conditioned by three factors: the residual energy of nodes belonging to the route, the delay requirements of the realtime flow and the load of the node’s queue. Routes discovery: When transmitting a new real-time data, the source node checks its route cache first to see whether there are available routes to the destination node. If routes are available, the protocol selects the suitable route according to the rules, which will be presented in next sub-section. Otherwise, the source node goes into the route discovery phase to request available route. Routes reply: When a destination node receives a RREQ, it returns back a Route REPly (RREP) packet to the source node. Different from DSR, in ED-DSR, while an RREP packet is being sent back to the source node, each intermediate mobile node will stamp its current status in the RREP packet. Finally, at the source node, the routing agent collects the RREP. This status information is shown in Table 1, in which i is the index for the mobile nodes. Table 1. Information fields of RREP packets Information fields
Contents Distance to this node provided by the physical layer Current length of queue, provided by the network layer. Current remaining energy of this node, provided by the physical layer.
An Energy and Delay-Aware Routing Protocol for Mobile Ad-Hoc Networks
127
ED-DSR calculates the cost of each available route according to the equation (4): ∑
.
(4)
Where is the cost of the route and , are the costs of node i considering the energy and queue length respectively. , and are the factors , and . is calculated as follows: which normalize .
(5)
is a function depending of the distance and remaining energy of node i. More remaining energy and shorter distance indicate less . is given below: 1
.
(6)
is the queue length at node i. equation is calculated in the same where manner as [9]. It is relative to the queue length along the current route. If there are more packets in the queues along the route, the transmission will increases rapidly with . inevitably suffer a longer delay. .
(7)
where is the queue length at node i, is the local processing time of any is the transmission time between two neighboring nodes in the message in node i; worst case remaining times and is the number of hops. depends on the queue length and the local processing time of each packet along the current route. Each packet should verify if it can reach the destination before the expiration delay (10). Otherwise, the node discards the route. ∑ Where
.
(8)
is the worst case execution time for the packet k.
Routes selection: In ED-DSR, the source node waits a certain period of time to collect RREP messages from the destination nodes along various routes. Among selected routes, the source node selects one based on minimum value of .
4 The Simulation Model We have used the Network Simulator, NS-2 in our simulations. NS-2 is an objectoriented, event driven simulator. It is suitable for designing new protocols, comparing different protocols and traffic evaluations. 4.1 Simulation Environment We simulated a MANET with 10-100 nodes in a 1500m×500m. With a rectangle area, longer distances between the nodes are possible than in a quadratic area, i.e. packets are sent over more hops.
128
J. Drira Rekik, L. Baccouche, and H. Ben Ghezala
Each node is equipped with an IEEE 802.11 wireless interface in a priority queue of size 50 that drops packets at the queue end in case of overflow. A traffic load between pair of source-destination (SMH-LMH) is generated by varying the number of packets per second on the constant bit rate - CBR. Each packet is 512bytes in size. We defined two groups of mobile nodes according to their resource capacity SMH and LMH. At the beginning of simulation, SMH nodes started with a starting energy of 50 joules and LMH with 100 joules. Since we did not address the problem of consumed energy in idle state, we have only considered energy consumed in transmission and reception modes. As values, we have utilized 1.4 W for transmission mode and 1 W for reception mode. The mobile nodes move around the simulation area based on the RWP mobility model, with a maximum speed of 2 m/s and a pause time of 10 seconds for SMH, which model a soldier mobility pattern and speeds of up to 20 m/s for LMH, which corresponds more to vehicular movements. All results reported here are the averages for at least 5 simulation runs. Each simulation runs for 1000 s. During each run, we assume that the node 0 wants to send real-time traffic to last node with an expiration delay equals to 15 seconds (firm realtime flow) and 25 seconds for higher expiration delay (soft real-time flow). Then, we observe the behavior of the nodes. 4.2 Performance Criteria Five important performance metrics are evaluated. They are used to compare the performance of the routing protocols in the simulation: - Real-time packet delivery in time ratio: the ratio of the real-time data packet that are delivered in time to the destination to those generated by CBR sources. - Real-time packet delivery ratio: the ratio of the real-time data packets delivered to the destination to those generated by CBR sources. - Mean end-to-end delay: the mean end-to-end delay is the time of generation of a packet by the source up to data packets delivered to destination. - Network lifetime The network lifetime corresponds to the first time when a node has depleted its battery power. - Energy consumption per bit delivery is obtained by dividing the sum of the energy consumption of the network by the number of successfully delivered bits.
5 Results and Discussions Several simulations are performed using NS-2 network simulator and using parameters shown in table 2. NS-2 generates a trace files analyzed using a statistical tools developed in AWK. The performance study concerns two versions of routing protocol DSR: DSR which refers to the classic DSR protocol [3] and ED-DSR which refers to our QoS protocol for two expiration delays 15s and 25s, which reflect respectively firm and soft real-time streams.
An Energy and Delay-Aware Routing Protocol for Mobile Ad-Hoc Networks
129
5.1 Impact of Traffic Load We propose here to study the impact of traffic load between pair of source-destination (SMH-LMH) by varying the number of packets per second on the CBR streams. The following figures show performance evaluation of DSR and ED-DSR protocols related to {5, 9, 10, 12, 15, 20} p/s on the CBR streams for 50 mobile nodes. 5.1.1 Network Performance The network performance is evaluated with three metrics, namely, the rate of realtime packets that are delivered in-time (where the deadline constraint is respected), the rate of real-time packets delivered and the end-to-end delay. Real-time packet delivery: Firstly, we observe and compare the variation of the ratio of all delivered packets regardless of compliance with the real-time constraints and the ratio of delivered packets in time, which respect the real-time constraint, while the data rate of the CBR flow is increased.
Packet delivery (%packets)
DSR D=15s
ED-DSR D=15s
ED-DSR D=25s
120 100 80 60 40 20 0 5
9
10
12
15
20
Packet rate of CBR (packets/sec streams)
Fig. 1. Real-time packets delivery ratio
Real-time packet delivery intime (%packets)
DSR D=15s
ED-DSR D=15s
ED-DSR D=25s
100 80 60 40 20 0 5
9
10
12
15
Packet rate of CBR (packets/sec streams)
Fig. 2. Real-time packets delivery in time ratio
20
130
J. Drira Rekik, L. Baccouche, and H. Ben Ghezala
The figure 1 proves that DSR provides better performance. However, DSR’s packet deliver ratio includes all packets that have reached the source node and where the deadline is not guaranteed for all packets received, as shown in figure 2. The EDDSR packet delivery ratio reflects the packets that have respected their real-time QoS constraint and will be handled in time. In both figures 1 and 2, the ratio of real-time packet delivery at source node is same. In fact, with ED-DSR, the real-time packets that have expired their deadlines are discarded by the intermediate nodes. In fact, each intermediate node verifies if the route response packet RREP respects or not the real-time constraint before reaching the source node. Thus, the MANET will avoid the network overloading with packets that have expired their deadline in order to reduce energy consumption and alleviate network load through intermediate nodes. However, with DSR, the real-time constraint is not guaranteed especially as the packet rate value increases. With firm real-time constraint, where D=15s, we note that packet delivery ratio in time decreases but stills stationary and better than DSR. The ratio of the packets sent within the compliance of its real-time constraint is over 50%. ED-DSR offers best performance for delivering real-time packets in time with soft real-time constraint. End-to-end delay guarantee: Another commonly used metric is the end-to-end delay. It is used to evaluate the network performance. As shown in figure 3, for low traffic approximate to 5packets/sec, the packet end-to-end delay results experienced by both protocols are comparable. It implies that the delay is respected when the communication load is low. When the communication load increases, a number of packets are dropped, the route discovery is restarted and the packet delay increases with DSR. It indicates that packet delay is sensitive to the communication load and is basically dominated by queue delay. However, with ED-DSR, the end-to-end delay stills low. The network overloading is avoided by discarding the packets that have expired their deadline and thus alleviate the load of mobile node queue.
End-to-end packet delay(s)
DSR D=15s
ED-DSR D=15s
ED-DSR D=25s
5 4 3 2 1 0 5
9
10
12
15
20
Packet rate of CBR (packets/sec streams)
Fig. 3. End-to-end packet delay
Our proposed protocol selects different routes depending on the cost function, thereby balancing the traffic in the network. This balancing helps to avoid overloaded
An Energy and Delay-Aware Routing Protocol for Mobile Ad-Hoc Networks
131
intermediate nodes in the network and reduces the delay for packets. Thus for high network traffic (up to 9packets/sec), our protocol gives much improved performance. We, also, note that the end-to-end delay is better for lower deadline (D=15s, firm real-time constraint). In fact, ED-DSR selects routes which reduce the transmission delay in order to respect the deadline. 5.1.2 Energy Efficiency In this section, we focus especially on the impact of our proposed protocol ED-DSR on energy efficiency guarantees. Two metrics, namely, the network lifetime and the energy dissipation are used. We study the impact of traffic load between pair of source-destination (SMH-LMH) by varying the number of packets per second on the CBR connection for 50 mobile nodes. Network lifetime: Firstly, we observe the variation of network lifetime while the data rate of the CBR flow is increased. Figure 4 shows the simulation results on small mobile host lifetime comparing ED-DSR and DSR under various traffic loads, while the data rate of the CBR flow are increased.
Lifetime of SMH(%)
DSR D=15s
ED-DSR D=15s
ED-DSR D=25s
100 80 60 40 20 0 5
9
10
12
15
20
Packet rate of CBR (packets/sec streams)
Fig. 4. The network lifetime of SMH nodes
We can see that networks running ED-DSR live longer than those running DSR, especially for high network traffic (up to 9packets/sec). As evident by the graph, our ED-DSR is little bit as efficient as DSR with low connection rate and much better in high traffic load. By avoiding the network overloading with packets that have expired their deadlines and selecting routes that minimize energy cost, ED-DSR alleviates network load and reduces energy consumption, too. DSR network lifetime was low in approximately all cases in comparison to EDDSR since DSR generates typically more routing overhead than ED-DSR. Energy dissipation: Figure 5 demonstrates the average energy consumption per bit delivery reflecting the global energy consumption in the network.
132
J. Drira Rekik, L. Baccouche, and H. Ben Ghezala
We see that ED-DSR outperforms DSR under different traffic loads, which is mainly due to the benefit of power control in the MAC layer. The excess packets inevitably introduce more collisions to the network, wasting more energy. ED-DSR chooses alternative routes, avoiding the heavily burdened nodes, thus alleviating the explosion in average energy consumption.
Average energy consumption (J)
DSR D=15s
ED-DSR D=15s
ED-DSR D=25s
1.5 1 0.5 0 5
9
10
12
15
20
Packet rate of CBR (packets/sec streams)
Fig. 5. The average energy consumption per bit
ED-DSR average energy consumption was lower than DSR average energy consumption under all packet rate conditions (over 9packets/sec) because ED-DSR selects path that minimize cost function. Changing expiration delay for different packet rate had not a significant impact on average energy consumption of ED-DSR. 5.2 Impact of Network Density In this section, we study the impact of network density on the ad-hoc routing protocols performance. This criteria is simulated varying the number of network nodes between 10 to 100: {10, 20, 30, 50, 70, 100} with 10 packets per second on the CBR streams. We focus especially on the impact of our proposed protocol ED-DSR on energy efficiency guarantees. Network lifetime: We observe the variation of network lifetime while the number of nodes is increased. Figure 6 shows the simulation results on small mobile host lifetime comparing ED-DSR and DSR. SMH, lifetime diminution according to node density augmentation is justified by the increase of generated routing overhead. Although the generated routing overhead had also increased in DSR, but this did not lead to an augmentation of its network lifetime. Nevertheless, DSR network lifetime was low in approximately all cases in comparison to ED-DSR since DSR generates typically more routing overhead than ED-DSR. In fact, in route selection, our proposal algorithm utilizes the network
An Energy and Delay-Aware Routing Protocol for Mobile Ad-Hoc Networks
Lifetime of SMH(%)
DSR D=15s
ED-DSR D=15s
133
ED-DSR D=25s
100 80 60 40 20 0 10
20
30
50
70
100
Number of nodes
Fig. 6. The network lifetime of SMH nodes
resources in terms of energy and load in order to balance traffic load. It ensures energy efficiency, up to 66%, by avoiding low-power node and busy node. Energy dissipation: Figure 7 gives an idea about the global average energy consumption in the network for both protocols DSR and ED-DSR. Increasing node density leads to an augmentation of collisions risk (consequently to more retransmission attempts) and to a growth in number of exchanged control packets. All those factors cause more energy dissipation for both protocols. ED-DSR average energy consumption is lower than DSR average energy consumption under all density conditions because ED-DSR selects path that minimizes cost function. Thus, its global energy consumption remains lower than DSR one.
Average energy consumption (J)
DSR D=15s
ED-DSR D=15s
ED-DSR D=25s
1.5 1 0.5 0 10
20
30
50
70
100
Number of nodes
Fig. 7. Average energy consumption per bit
Changing expiration delay for different node densities has not a significant impact on average energy consumption of ED-DSR.
134
J. Drira Rekik, L. Baccouche, and H. Ben Ghezala
6 Conclusions and Future Work There is a growing demand that mobile nodes should provide QoS to mobile users since portable devices become popular and applications require real-time services. In this paper, an energy delay-aware routing protocol for mobile ad hoc networks is proposed. ED-DSR is a routing protocol which allows the packets of real-time flows to be routed from the sender to the receiver before the expiration delay to deadline. In addition, the route selection is done according to energy consumption and queue load of intermediate nodes, too. Cost function is defined based on residual energy, queue length, processing and transmission time of intermediate nodes. The route is selected based on minimum value of cost function. Simulation results prove the performance of our proposal routing protocol. They indicate that ED-DSR prolongs network lifetime and achieves lower energy dissipation per bit of data delivery, higher volume of packets delivered and lower endto-end delay. In the future, we plan to integrate weighting factor to the cost function and study their effects on the system in order to appropriate QoS service to user needs: which factor to privilege (energy, delay or node load).
References 1. Das, S., Castañeda, R., Yan, J.: Simulation-Based Performance Evaluation of Routing Protocols for Mobile Ad-hoc Networks. Mobile Networks and Applications 5(3), 179–189 (2000) 2. Hu, Y.C., Johnson, D.B.: Implicit Source Routes for On-Demand Ad Hoc Network Routing. In: ACM MobiHoc (2001) 3. David, B.J., David, A., Maltz, B.J.: DSR: The Dynamic Source Routing Protocol for Multi-Hop Wireless Ad-hoc Networks. In: Perkins, C.E. (ed.) Ad-hoc Networking, ch. 5, pp. 139–172. Addison-Wesley, Reading (2001) 4. Frikha, M., Maamer, M.: Implementation and simulation of OLSR protocol with QoS in Ad-hoc Networks. In: Proc. of the 2nd International Symposium on Communications, Control and Signal Processing (2006) 5. Huang, J., Chen, M., Peng, W.: Exploring Group Mobility for Replica Data Allocation in a Mobile Environment. In: Proc. Twelfth international conference on Information and knowledge management, pp. 161–168 (2003) 6. Kuo, C., Pang, A., Chan, S.: Dynamic Routing with Security Considerations. IEEE Transactions on Parallel and Distributed Systems 20(1), 48–58 (2009) 7. Maleki, M., Pedram, M.: Power-Aware On-Demand Routing Protocols for Mobile Ad-hoc Networks. In: Piguet, C. (ed.) Low Power Electronics Design, The CRC Press, Boca Raton (2004) 8. Mbarushimana, C., Shahrabi, A.: Congestion Avoidance Routing Protocol for QoS-Aware MANETs. In: Proc. of International Wireless Communications and Mobile Computing Conference, pp. 129–134 (2008) 9. Meng, L., Lin, Z., Victor, L., Xiuming, S.: An Energy-Aware Multipath Routing Protocol for Mobile Ad Hoc Networks. In: Proc of Sigcomm Asia Workshop, Beijing, China, pp. 166–174 (2005) 10. Ouni, S., Bokri, J., Kamoun, F.: DSR based Routing Algorithm with Delay Guarantee for Ad-hoc Networks. JNW 4(5), 359–369 (2009)
Energy-Aware Transmission Scheme for Wireless Sensor Networks Abdullahi Ibrahim Abdu and Muhammed Salamah Computer Engineering Department Eastern Mediterranean University KKTC, Mersin 10, TURKEY
[email protected],
[email protected] Abstract. In this paper, we proposed a technique to extend the network lifetime of a wireless sensor network, whereby each sensor node decides whether to transmit a packet or not and with what range to transmit the packet. A sensor node makes this decisions based on its own energy resource and the information contained in each packet. The information content in each packet is determined through a system of rules describing prospective events in the sensed environment, and how important such events are. While the most important packets are propagated by virtually all sensor nodes and with different transmission ranges depending on their battery life, low importance packets are propagated by only sensor nodes that have high energy reserves and with greater transmission ranges due to high reserves. The result show that by adjusting the transmission ranges based on energy reserves, a considerable increase of lifetime is achieved. Keywords: Energy-aware; Wireless sensor networks; Transmission range adjustment; Priority balancing.
1 Introduction A wireless sensor network (WSN) typically consist of a number of small, inexpensive, locally powered sensor nodes that communicate detected events wirelessly through multi-hop routing [1]. Typically, a sensor node is a tiny device that includes three basic components: a sensing subsystem for data acquisition from the physical surrounding environment, a processing subsystem for local data processing and storage, and a wireless communication subsystem for data transmission. In addition, a power source supplies the energy needed by the device to perform the programmed task. This power source often consists of a battery with a limited energy budget. In addition, it could be impossible or inconvenient to recharge the battery, because nodes may be deployed in a hostile or unpractical environment [2]. WSNs are being used in a wide variety of critical applications such as military, health-care applications [3], health care [4], environmental monitoring [5], and defense [6]. A key research area is concerned with overcoming the limited network lifetime inherent in the small, locally powered sensor nodes [1]. To improve this limited network life time, new and modified routing algorithms were proposed. A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 135–144, 2011. © Springer-Verlag Berlin Heidelberg 2011
136
A.I. Abdu and M. Salamah
In this paper, we proposed a technique to extend the network lifetime of a wireless sensor network; referred to as IRT or IDEALS|RMR|TRA (Information manageD Energy aware Algorithm for sensor networks with Rule Managed Reporting and Transmission Range Adjustments). The extension in the network lifetime is achieved at the possible sacrifice of low importance packets and adjustment of transmission ranges based on nodes energy resource. One big advantage of this technique is that, nodes do not have to transmit packets with their maximum transmission ranges all the time. They take into consideration their energy resource and adjust their transmission range accordingly. Nodes also maintain full connectivity by making their minimum transmission range cover at least one sensor node. The rest of this paper is organized as follows: Section 2 presents related work. Section 3 presents our proposed research. Section 4 gives the performance analysis (simulation results and discussions). Section 5 provides conclusion and future work.
2 Related Work The energy management technique IDEALS|RMR [7] extends the lifetime of a wireless sensor network, whereby a node with high energy reserve act for the good of the network by forwarding all packets that come to it and by generating its own packets. However, a node with low energy reserve acts selfishly by only generating or forwarding packets with high information content. In addition, IDEALS|RMR uses a single-fixed transmission range for each sensor node regardless of whether its energy resource is high or low and this causes redundancy in energy consumption as lots of areas are covered by several sensors. Authors in [8] developed a power saving technique by combining two methods: scheduling sensors to alternate between active and sleep mode method, and adjusting sensors sensing range method. They combined both methods by dynamic management of node duty cycles in a high target density environment. In their approach, any sensor schedules its sensing ranges from 0 to its maximum range, where 0 corresponds to sleep mode. Adding the system of information control proposed in this paper could significantly save energy. Authors in [10] try to deal with the problem of energy holes (unbalance distribution of communication loads) in a wireless sensor network. This means that, energy of nodes in a hole will be exhausted sooner than nodes in other region. As, energy holes are the key factors that affects the life time of wireless sensor network, they proposed an improved corona model with levels for analyzing sensors with adjustable transmission ranges in a WSN with circular multi-hop deployment. Additionally, the authors proposed two algorithms for assigning the transmission ranges of sensors in each corona for different node distributions. These two algorithms reduce the searching complexity as well as provide results approximated as optimal solution. M. Busse et al. [11] maximize the energy efficiency by proposing two forwarding techniques termed single-link and multi-link energy efficient forwarding. Single-link forwarding sends a packet to only one forwarding node; multi-link forwarding exploits the broadcast characteristics of the wireless medium. This leads to efficiency because if one node doesn’t receive a packet, another node will receive the packet and
Energy-Aware Transmission Scheme for Wireless Sensor Networks
137
performs the forwarding process. There is however a tradeoff of delivery ratios against energy costs. Based on extensive survey on Wireless sensor networks so far, we believe that energy management information control of [7] coupled with transmission range adjustment in [8] was never considered, hence in this paper, we find it worth considering. Our proposed IRT system extends the network lifetime for packets with high information content by losing packets of less important just as IDEALS|RMR does. However, transmission range of a sensor is adjustable in IRT, thus allows the sensor more choice to reduce its energy consumption, hence prolong the lifetime of WSN.
3 The Proposed IRT Scheme Researchers are continuously developing new and modifying existing energy management techniques in order to improve the network life time of WSNs. This is because radio communication is often the most power-intensive operation in a WSN device. For that reason, we modified the IDEALS|RMR energy management scheme to improve the network life time. The main contributions of this study are: -
Coupling IDEALS|RMR with transmission range adjustment (TRA). Performing a detailed analysis by simulating IRT, IDEALS|RMR and tradition case to prove that IRT is the most energy efficient technique.
The operation of IRT can be seen in figure 1. When a sensor senses a data, it passes the data to the controller, which sends a value (e.g. temperature) to RMR (Rule Management Reporting) unit. RMR is a technique which determines if an event worth reporting has occurred, and how important such an event is. The value is received by the Rule Compliance Testing. This rule compliant testing’s responsibility is to determine if an event worth reporting has occurred. It does that by checking the sensed value against the rules in the Rule Database (getting history information about the previously sensed values), at the same time, updates the history with the current information of packets and sensed value. Rules may be fulfilled or not, any rules which are fulfilled are passed to the Message Priority Allocation to determine how important the content of the packet is. It does that by assigning message priorities (MP) to each fulfilled rule. In this work, five different MPs are used (MP1-MP5). MP1 related to most important packet. (For example, temperature is higher than normal and requires urgent attention). In the contrary, MP5 relates to the least important packet (for example, a normal temperature packet, to indicate everything is fine). Any number of predefined rules can be entered by the designer, and describing different events that can be detected in the sensed environment, examples of possible rules are 1. 2. 3.
Threshold rules (report when the sensed value crosses a preset value). Differential rules (report when the change in the sensed value is larger or smaller than a preset value). Routine rules (report is a packet of that importance or higher has not been sent for a preset period) [7].
138
A.I. Abdu and M. Salamah
Fig. 1. The proposed IRT system diagram
Afterwards, the MP obtained from message priority allocation is passed to the IDEALS unit. Its responsibility is to decide if the node should transmit a packet or not, and it’s done by Priority Balancing. The node’s energy resource is characterized by power priority allocation, which assigns a power priority (PP) based on the state of the battery. The highest power priority is PP5, and it’s allocated to a node with the highest energy reserve, while the lowest power priority is PP1 and it’s allocated to the node with the lowest energy reserve. When priority balancing receives MP and PP, it compares them and if PP ≥ MP, then a packet will be transmitted. Finally, when a node decides to transmit a packet, PP is passed to transmission range adjustment (TRA) unit. Its purpose is to decide with what range a sensor node will transmit a packet, which is done by Suitable Transmission Range. Suitable transmission range gets power priority (PP) from power priority allocation and coordinates from reachable sensors. These reachable sensors are the entire sensors in the maximum transmission range of a sending sensor node. Now, based on the value of the power priority and the coordinates of the sensors in the maximum transmission range of the sending sensor node, a suitable transmission range is determined and passed to the controller to successfully transmit the packet with the new range. In this work, five different TRs are used (TR1-TR5), where TR1 is the minimum transmission range and TR5 is the maximum transmission range. There is a one-to-one mapping of the power priority to the transmission ranges. When a packet is ready to be transmitted, meaning that PP≥MP, the transmission range will be mapped to the value of PP. The priority allocation, balancing process and transmission range adjustment can be seen in figure 2. For example, when we have a full battery PP5, we will transmit all packets regardless of their message priority MP1 to MP5 with the maximum transmission range TR5. However, if our
Energy-Aware Transmission Scheme for Wireless Sensor Networks
139
ȋȌȋȌȋȌ
ͷ Ǥ Ͷ
ͷ
Ͷ
͵
͵
ͷ Ǥ Ͷ
͵
ʹ
ʹ
ͳ Ǥ
ͳ
ʹ
ͳ Ǥ
Ͳ
Fig. 2. Priority Balancing with Transmission Range
battery decreases to the minimum PP1, we will have the chance to transmit only packets with the highest message priority MP1 and with the lowest transmission range TR1. Priority balancing and transmission ranges can also affect routing process-if a node’s residual energy level does not warrant sending a packet of a certain priority, it will not participate in routing. Data deemed not to be significant enough (considering the state of the network) can be dropped at a number of stages: event generation (if a change in the data does not trigger a rule, an event will not be generated), local priority balancing (if the PP<MP, the packet will not be created from the generated event), and routing (if no route exists across the network where PP≥MP due to loops, the packet will not reach its destination).
4 Performance Analysis We performed our simulation using C programming, where we compared the data sets of tradition case, IDEALS|RMR, and IRT. Repetitive simulations were perfumed for IDEALS|RMR, and IRT to verify our simulation results provided in figure 5. IRT system is not suited to an application where all data are equally important. 4.1 Simulation Setup Initially, all sensor nodes have the same initial energy of 100 Joules [7], the equation for energy required to transmit a packet (1), where Eelec[J] is the energy required for the circuitry to transmit or receive a single bit, Eamp[J] is the energy required for the transmit amplifier to transmit a single bit a distance of one meter, d is the separation
140
A.I. Abdu and M. Salamah
distance in meters and l is the length of the packet 1000bits. As it can be seen in figure 4, we used 20 sensors nodes, distributed them randomly in a 70×70 meters area and each node has the same maximum transmission range of 20 meters. The coordinate of the sensors is saved in a file and given as input to the program. Therefore, the distances to each of the sensors in the maximum transmission range of a sending sensor is determined using the distance formula (2). Other simulation parameters are listed in table 1 as shown below. Table 1. Simulation parameters
Simulation area
70×70 meters dimension
Number of nodes
20 nodes
Packet length
1000 bits
Initial node energy
100 Joules
Simulated Node Id
node-08
Minimum transmission range
13.038405 meters
Maximum transmission range
20 meters
Simulated node Coordinate
(x = 38 , y = 37)
Etx(l,d) = Eelecl + Eampld2 . d
x
x
y
y
(1) .
(2)
First, the user has to create a file, and provide the coordinates of sensors in it; the program then outputs (sending node id, its coordinates, the node id’s in its maximum transmission range and their distance to the sending node), for all sensor coordinates provided. One sensor is now chosen for the simulation as the remaining sensors are assumed to be identical. The chosen sensors id and the distance to its closest sensor are given as inputs, the sensor senses data and our IRT algorithm is perfumed as illustrated in figure 3. Since the maximum transmission range (TR5) is fixed for every sensor, five different transmission ranges can be calculated by considering the minimum transmission range (TR1) as the distance to the closest sensor in the sending sensor’s maximum transmission range. So the ranges between TR1 to TR5 are calculated by successively adding the ΔTR = (TR5-TR1)/4. That is, TR(i) = TR(i-1) + ΔTR, for i=2,3,4. For example, adding ΔTR to TR1 gives TR2, and so on. The reason we took TR1 as the distance to the closest sensor in the sending sensor’s maximum transmission range is because it covers at least one sensor so that in the worst case (PP1), we have full connectivity (packets can be delivered to the sink node). All nodes except the sink node (final destination of packets), performs multi-hop routing of packets by using flooding algorithm. Our program is so dynamic that different coordinates from the ones used in our simulation can be entered and any node can be chosen for the simulation. Figure 4 shows a snap shot of randomly
Energy-Aware Transmission Scheme for Wireless Sensor Networks
141
Start
Enter Sensor Id, and the distance to its nearest sensorǤ
Calculate the five transmission ranges of the sending sensor
Sensor Senses Data
No ͷǤ
Check sensed data against each rule in database. Rule(s) fulfilled?
Yes Message Priority Allocation assigns MP to fulfilled rules.
Power Priority Allocation assigns PP to Residual Energy Level.
No
PP≥MP? Yes
TR=PP
Transmit Packet with new TR
Decrease Residual Energy
Packet Received
No
Residual Energy≤0?
Yes
Stop
Fig. 3. Flowchart of IRT
142
A.I. Abdu and M. Salamah
Fig. 4. A snap shot of randomly distributed nodes used in the simulation
distributed nodes used in the simulation. Circles represent the maximum transmission range of sensors and lines represent possible communication link [7]. We chose node8 as it’s located in the middle. We assumed packets are transmitted every 5 minutes. 4.2 Simulation Results Our simulation was conducted to show how the network life time can be improved by using IRT technique proposed in this paper. The data generation in our simulation is not so important, what is important is whether rules are fulfilled or not, and the power priority of the residual energy. Node-08 was use in this simulation due to its location; any node could be taken as well. Node energy Depletion Times: Figure 5 shows the time it takes for a technique (traditional, IDEALS|RMR, IRT) to deplete its energy reserve. ‘100’ means nodes energy is full, ‘0’ means nodes energy is depleted. It can be seen that in the tradition case that node-8 depletes its energy reserve after around 10 hours, as it is sending packets every 5min. without taking into account the information contents, and energy levels. In the case of IDEALS|RMR, packets are not transmitted every 5min., as the packet importance are considered before transmission. Hence, the node lifetime significantly increases. Finally, in the IRT case, packets are also not transmitted every 5min., as the importance of packets are considered before transmission. Unlike IDEALS|RMR which uses one fixed transmission range, IRT can adjust its
Energy-Awaare Transmission Scheme for Wireless Sensor Networks
143
F 5. Node energy depletion times Fig.
transmission range based on n nodes energy level, thus allows the sensor more choice to reduce its energy consum mption. As a result, our proposed IRT scheme show ws a significantly high increase in i node life time compared to the other two cases. As it can be seen from the figure, th he attained improvement of IRT over IDEALS|RMR and traditional schemes reachess 80% and 530% respectively. It can be noticed from fiigure 5 that the energy level of the IDEALS/RMR and IIRT schemes drops suddenly an nd then becomes constant, this process continues until the battery is depleted. The sud dden dropping of energy level represents continues paccket transmission because PP≥M MP (battery level is high enough to allow a node to transsmit a packet of that importan nce) while the constant energy level represents PP< <MP (battery level is not high en nough to allow a node transmit a packet of that importannce, therefore no transmission occurs).In o a nutshell, if a packet is not transmitted duee to PP<MP, the same battery level will be maintained until a packet arrives in whhich fi 2. PP≥MP, as can be seen in figure
5 Conclusions In this paper, we proposed d an IRT scheme, which operates upon a combinationn of information management reeporting (determining the information contents of a paccket and how important such an information is, through a system of rules), eneergy he residual energy level with the packet importance) and management (balancing th
144
A.I. Abdu and M. Salamah
transmission range adjustment (determining a suitable transmission range for a node based on its residual energy level) which we believe has not be considered before. Simulation was done using c programming where a single node was simulated to show the operation of our algorithm and display the results. The result shows that by adjusting the transmission range of a sensor node based on its battery life, the nodes battery life is extended significantly compared with the other two methods (Traditional and IDEALS|RMR). We are currently working on a more intelligent technique to determine the transmission range, which will not be based on the residual energy alone, but on the message importance as well. Moreover, analysis of network connectivity (the measure of the ability of any node in the network to successfully transmit a packet) and packet success (the packet that were transmitted by the simulated node and successfully receive by the sink node) can be added.
References 1. Akyildiz, I.F., Su, W., Sankarasubramaniam, Y., Cayirci, E.: Wireless sensor networks: a survey, Comp. A Survey, Comp. Netw. 38, 393–422 (2002) 2. Anastasi, G., Conti, M., Francesco, M.D., Passarella, A.: Energy conservation in wireless sensor networks: A survey. Ad Hoc Networks 7, 537–568 (2009) 3. Elrahim, A.G.A., Elsayed, H.A., Ramly, S.E.L., Ibrahim, M.M.: An Energy Aware WSN Geographic Routing Protocol. Universal Journal of Computer Science and Engineering Technology 1(2), 105–111 (2010) 4. Lo, B.P.L., Yang, G.-Z.: Key technical challenges and current implementations of body sensor networks. In: Proceedings of the Second International Workshop Wearable and Implantable Body Sensor Networks (BSN 2005), London, UK (April 2005) 5. Werner-Allen, G., Lorincz, K., Ruiz, M., et al.: Deploying a wireless sensor network on an active volcano. IEEE Internet Computing 10, 18–25 (2006) 6. Simon, G., Maroti, Ledeczi, A., et al.: Sensor network-based countersniper system. In: Proc. Conf. Embedded Networked Sensor Systems, Baltimore, MD, pp. 1–12. Baltimore (2004) 7. Merrett, G.V., A‘l-Hashimi, N.M., White, N.R.: Energy managed reporting for wireless sensor networks. Sensors and Actuators A142, 379–389 (2008) 8. Mingming, L., Jie, W., Mihaela, C., Minglu, L.: Energy-Efficient Connected Coverage of Discrete Targets in Wireless Sensor Networks. International Journal of Ad Hoc and Ubiquitous Computing 4, 137–147 (2009) 9. The Distance Formula, http://www.purplemath.com/modules/distform.htm 10. Song, C., Liu, M., Cao, J., Zheng, Y., Gong, H., Chen, G.: Maximizing network lifetime based on transmission range adjustment in wireless sensor networks. Computer Communications 32, 1316–1325 (2009) 11. Busse, M., Haenselmann, T., Effelsberg, W.: Energy-efficient forwarding in wireless sensor Networks. Pervasive and Mobile Computing 4, 3–32 (2008)
PRWSN: A Hybrid Routing Algorithm with Special Parameters in Wireless Sensor Network Arash Ghorbannia Delavar1, Javad Artin1, and Mohammad Mahdi Tajari2 1 Payam Noor University, Tehran, Iran
[email protected],
[email protected] 2 Islamic Azad University, Mashhad, Iran
[email protected] Abstract. We will present a new Hybrid Routing Algorithm with Special Parameters in Wireless Sensor Network for network with many sensors. This algorithm will select the cluster heads (CH) based on the scale of average local energy and the density surrounding each node. In the presented algorithm a type of mechanism was used in which it performs cluster formation in special conditions with regard of the neighborhood principle and local information of a node and its neighbors. Also in PRWSN the data of each cluster is sent to BS via a hybrid method. In networks with high density which use multi-hop methods, the energy of nodes close to BS will be discharged with a higher speed while in single-hop methods the energy of nodes which are far from BS will be diminished earlier. To overcome these problems, we use a combination of the single and multi-hop methods for increasing the lifetime of the network. In the presented algorithm, parameters such as the distance to the Base Station and energy are used to choose the next step in the hybrid method. Finally, with the results of multiple simulations we were able to show that PRWSN, in comparison with the previous Clustering Algorithm has increased the lifetime of sensor network, and reduced the amount of node energy consumption by balancing the use of energy between nodes, therefore resulting to a more suitable distribution of clusters in the sensor network. Hence, this algorithm is more effective compared to the previous algorithms. Keywords: Wireless sensor network, energy balancing, energy efficiency, routing, clustering.
1 Introduction During the Twenties century, man has used sensors for the means of monitoring his surroundings. As a result of the advances in wireless communication and electronics technologies, wireless sensors are getting smaller, cheaper, and more powerful. The development of these miniaturized wireless sensors enables to use sensor networks for many applications such as military surveillance, environmental monitoring, infrastructure and facility diagnosis, and other commercial applications [1, 2, 3, 4]. These sensors monitored their environment and transformed the received information into an electrical signal. The signal processing method has a close relationship with A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 145–158, 2011. © Springer-Verlag Berlin Heidelberg 2011
146
A.G. Delavar, J. Artin, and M.M. Tajari
the event type and the subject under review. Sensors present in the environment send their data to a data center or base station (BS) for review and further decisions [5]. Despite proper performance and high efficiency levels have a number of limitations in high-scale development. Limitations of the sensors can be grouped in one of the following forms: low energy power levels, and the lack of battery replacement in most cases, limitation of their bandwidth and short radio range. Managing a large number of nodes with these limitations can provide many challenges [6]. Energy limitation of the nodes has caused that the reduction of energy consumption in all layers of sensor network design, be considered as one of the main cases. One of the most important layers in this case is the network layer in which the routing process is done. The routing notion in sensor networks is distinguished from other wireless communication networks because of some intrinsic characteristic and has placed more challenges in the path of the design. Some of its intrinsic characteristic includes: lack of a broad IP creation for every node and in result the inability of execution of the many common network routing algorithms, which result to an increase in data traffic in these networks and also limitations in the energy of message transmission, limitations in the power present in each node, and limitations in the calculation potency and the memory on nodes. Regarding the mentioned cases, a lot of routing methods were created for WSNs which can be divided into three group's base on the most common categorization: Data–centric Algorithms, Location Base Algorithms and Hierarchical Algorithms [6]. Data-centric protocols are query-based and depend on the naming of desired data, which helps in eliminating many redundant transmissions. Location-based Algorithms utilize the position information to relay the data to the desired regions rather than the whole network. Hierarchical Algorithms aim at clustering the nodes so that cluster heads can do some aggregation and reduction of data in order to save energy [6]. Cluster based methods benefit from a few characteristics: the first characteristic is that they divide the network into several parts and every part is directed by one cluster head. This characteristic causes the cluster based methods to be of a higher scalability level. The second characteristic is that a cluster head receives the data of its nodes and sends it BS after gathering data, which results in substantial reduction in data redundancy. We will provide a clustering algorithm, which uses a new distributed method, and a local threshold detector to perform clustering. Also in this algorithm, we use a combination of the single and multi-hop methods to send the data to BS in order to increase the lifetime of the network. By comparing PRWSN to previous algorithms we will evaluate it.
2 Related Work Grouping of SNs into clusters has been widely used as an energy-efficient organization of nodes in WSNs [7]. In each cluster, one of the member nodes, the cluster head, acts as a local coordinator of the basic operations within the cluster, namely communication as well as data aggregation operations. Clustering protocols are well suited for aggregation operations since nodes in the same cluster are in close proximity, and thus, data sensed by these nodes are correlated and can be effectively aggregated. [8]
PRWSN: A Hybrid Routing Algorithm
147
One of the most famous clustering algorithms is LEACH [9]. The operation of LEACH is divided into rounds and each round separated into two phases, the set-up phase and the steady-state phase. In the set-up phase, each node decides whether or not to become a cluster head for the current round. This decision is based on the threshold T (n) given by: p ⎧ ⎪ ⎪ T ( n) = ⎨1 − p ∗ (r mod ⎪ ⎪0 ⎩
1 ) p
if n ∈ G ,
(1) otherwise
Where p is the predetermined percentage of cluster heads (e.g., p = 0.05), r is the current round, and G is the set of nodes that have not been cluster heads in the last 1/p rounds. Cluster head broadcasts an advertisement message to the other nodes. Depending on the signal strength of the advertisement messages, each node selects the cluster head it will belong to. The cluster head creates a Time Division Multiple Access (TDMA) scheme and assigns each node a time slot. In the steady-state phase, the cluster heads collect data from sensor nodes, aggregate the data and send it to the base station. Since the decision to change the CH is probabilistic, there is a good chance that a node with much low energy gets selected as a CH. When this node dies, the whole cell becomes dysfunctional. Also, the CH is assumed to have a long communication range so that the data can reach the base-station from the CH directly. This is not always a realistic assumption since the CHs are regular sensors and the base-station is often not directly reachable to all nodes due to signal propagation problems. [7] HEED [10] provides balanced cluster heads and smaller sized clusters. They use two radio transmission power levels; one for intra-cluster communication and the other for inter-cluster communication. HEED does not select cluster head nodes randomly. Sensor nodes that have a high residual energy can become cluster head nodes. But the cluster heads election uses complex iterative algorithm, and considers only the remaining energy situation of some nodes. HEED will not play a significant role in evening energy consumption of cluster heads for the entire network. PEGASIS [11] adopts strategy to ease the issue of rapid consumption of sensor nodes energy caused by direct communicating with the base station. PEGASIS is limited to communicate with adjacent nodes. A cluster head is randomly selected to communicate with base station at each round, which decreases data traffic. But it increases the data delay. At the same time, in PEGASIS there exists serious "hot zone" problem, resulting in the imbalance of energy consumption. “Hot zone” problem is the phenomenon that the energy of nodes near base station is quickly consumed as transferring other clusters’ data in multi-hop routing [12]. Energy Residue Aware (ERA) [13] clustering algorithm is another energy-aware hierarchical approach. It is also improved from LEACH by including the communication cost into the clustering. The communication cost includes residual energy, communication energy from the CH to the sink and communication energy from the cluster members to the CH. There is a difference from HEED: ERA uses the same CH selection scheme as LEACH but provides an improved scheme to help
148
A.G. Delavar, J. Artin, and M.M. Tajari
non-CH nodes choose a ‘‘better” CH to join by calculating the clustering cost and finding CH according to maximum energy residue. PEBECS [14] focusing on the hot spot problem, PEBECS presents the solution by dividing a WSN into several partitions with equal area and then grouping the nodes into unequally sized clusters. The shorter the distance between the partition and the sink, the more clusters are created within the partition. Further, to select the CH, PEBECS uses the node’s residual energy, degree difference and relative location in network. PEBECS mitigates the hot spot problem by grouping nodes in smaller clusters to save more energy on their intra-cluster communication. As the result, PEBECS achieves longer network lifetime by better balancing node energy consumption. In previous work we proposed a Distributed Balanced Routing Algorithm with Optimized Cluster Distribution RCSDN [15]. In which the formation of clusters is locally done by a distributed method and through calculation of the average of energy in each node. After gathering the data in CH, they will be sent to BS using the singlehop method. Now by using the presented algorithm we can use a new selection conditions for creating clusters which reduces node energy consumption, and balances node energy consumption by the appropriate distribution of cluster head in the network; and create better conditions compared to previous methods.
3 System Model 3.1 Network Model The network model of the PRWSN which is under study contains the following characteristics: 1. 2. 3.
The base station and all sensor nodes are stationary after deployment. The basic energy of nodes is different. Sensor nodes do not require GPS-like hardware. So, they are not location aware.
3.2 Energy Model Generally, sensors consume energy when they sense, receive and transmit data [16]. Our energy model for the sensor network is based on the first order radio model as used [9]. In this model, the transmitter has power control abilities to dissipate minimal energy to send data to the receiver. In order to achieve an acceptable signal-to-noiseratio (SNR), the energy consumption of the transmitter is given by: ⎧ ETx ( n, d ) = n ( E elec + ε fs d 2 ) ⎪ ⎨ 4 ⎪ ETx ( n , d ) = n ( E elec + ε mp d ) ⎩
d < d0 d ≥ d0
(2)
Where, n is the number bit of the message and d is the distance. Eelec is the energy dissipated per bit to run the transmitter or the receiver circuit, and Ԑfs , Ԑmp is the energy dissipated per bit to run the transmit amplifier depending on the distance
PRWSN: A Hybrid Routing Algorithm
149
between the transmitter and receiver. If the distance is less than a threshold d0 , the free space (FS) model is used; otherwise, the multipath (MP) model is used. The energy consumption of the receiver is given by:
ERx (n) = n( Eelec )
(3)
4 Description of the PRWSN Algorithm PRWSN is created based on rounds in which each round contains two phases: (1) setup and (2) steady-state. In the Set-up Phase, the CHs are determined and the cluster structures are formed. In Steady-state Phase the nodes send the data they’ve received from the environment to the corresponding CH and after gathering data in the cluster head, data will be sent to the BS. The Set-up Phase starts by sending a start message via BS with a specified range R to the environment. After a node receives the start message from the BS, it provides a relative estimate of its distance from the BS through the intensity of the received signal. Then it broadcasts a message for its neighboring nodes including ID, the distance to BS, count of neighbors, and the level of remaining energy. Nodes bound in the radio range of this message, receive it and set this node as a neighbor node, and register its ID and energy level in their memory; Again they proceed to estimate their distance with the neighboring node by calculating the intensity of the received signal and finally calculate their distance from the BS. This is the done by all nodes in the network. Set-up phase contains two steps: 1.Cluster formation: In this step the CHs are selected and the structure of clusters is formed. 2. Route discovery: In this step, each CH, selects the next CH to transfer its data to BS. 4.1 Cluster Formation In the beginning of each round all of nodes have a normal state. In PRWSN we use a local threshold detector (TD) so that only nodes having appropriate energy participate in the competition for CH selection. This threshold detector is locally calculated in each node to prevent the lack-of-candidates problem in some areas because of the central selection. The following qualification function is used to prevent this problem:
TD
∑ (S ) = α × i
ncount ( S i ) k =1
E re ( S k )
(4)
k
Where ncount (Si) is equal to the number of neighbors of the node “i”. Ere (Sk) is the remaining energy of the neighbor “k”. And α ϵ [0.5, 1.5] is a coefficient which determines the candidate threshold; the larger this number is, the lesser number of nodes become candidates. Each node decides to become a candidate or not, based on the following relation: if Ere (Si) > TD (Si) state (Si) = candidate
(5)
150
A.G. Delavar, J. Artin, and M.M. Tajari
After distributing the start message based on the following algorithm which is implemented in each node, the proper CHs are selected and clusters are formed. In which Sn is the collection of neighbor nodes in radio range of each node and is defined as:
S n ( S i ) = {S j | d ( S i , S j ) ≤ R ( S i )}
(6)
Also Scn , is the collection of the neighboring nodes of a node which its remaining energy is equal to or more than TD.
S cn ( S i ) = {S j | S j ∈ S n ( S i ) ∧ E re ( S j ) ≥ T D ( S i )}
(7)
Algorithm 1. Cluster formation 1. calculate TD ( S i ) 2. 3. 4.
if Ere (S i ) > TD (S i )
state ( S i ) = candidate-CH create S cn ( Si )
5. 6.
end if Initialize T
7.
While ( state ( S i ) = candidate-CH OR state ( S i ) = normal ) AND timer < T if candidate-CH AND ( state ( S i ) =
8.
( ∀ S j ∈ S cn ( S i ) : n count ( S i ) > ncount ( S j ) ) ( ∀ S j ∈ S cn ( S i ) : n count ( S i ) = n count ( S j ) ⇒ 9. 10. 11. 12. 13. 14. 15. 16.
OR
E re ( S i ) > E re ( S j ) ) )
state ( S i ) = CH broadcast a CH( Id ( S i ) ) message break end if if received a CH( Id ( S j ) ) message if state ( S i ) = normal state ( S i ) = cluster-member CH ( S i ) = S j
17. 18.
break else if state( Si ) = candidate-CH
19.
state ( S i ) = cluster-member
20.
CH ( S i ) = S j broadcast a Abort ( Id ( S i ) ) message
21. 22. 23. 24. 25.
break end if end if if received a Abort ( Id ( S j ) ) message
PRWSN: A Hybrid Routing Algorithm
26.
remove S j from S n ( Si )
27.
calculate TD ( S i )
28.
if state ( S i ) = normal AND E re (S i ) > TD (S i )
29.
state ( S i ) = candidate-CH create S cn ( S i ) else if state ( S i ) = candidate-CH recreate S cn ( Si ) end if end if end while if state ( S i ) = candidate-CH OR state ( S i ) = normal state ( S i ) = CH broadcast a CH( Id ( S i ) ) message end if
30.
31. 32. 33. 34. 35. 36. 37. 38. 39.
151
In this algorithm, each node performs a reversed timer regarding its remaining energy level. If this timer concludes and no message of cluster formation is received from other nodes or the node has not been set as a CH, it will introduce itself as a CH and broadcast a CH formation message including its ID. First by applying TD, the nodes with suitable energy will be chosen as candidates. Then if a candidate node has the most neighboring number among its local neighboring candidates (Scn(Si)) or has an equal number of neighbors with another node and its remaining energy is more than the node, it will introduce itself as a CH via message broadcast. Because usually nodes of a local region will detect similar data, and the more nodes a cluster has, the number of clusters will be reduced and less data will be sent to BS. As a result, the energy consumption will be reduced and lifetime of the network will be increased. When a node receives the CH message, if it has a normal state, it will immediately joined the chosen CH, and change its state to cluster-member. But if it has the candidate-CH state, it will perform the following actions: A. Change its state to cluster-member. B. Create an abort message containing its ID and broadcast it to its neighbors. When a node receives an abort message from its neighbors, if the state of the node is set to normal or candidate-CH, it will delete the node which sent the message from its neighbors list (Sn(Si)) and proceed to recalculating the threshold detector. If the state of a node is candidate, the general collection of candidate node (Scn(Si)) will be gained regarding the new amount of TD. But if the state of a node is normal, again it compares its remaining energy to the threshold, and it’s possible that a node that used to be in a normal state, to turn to a candidate-CH state; this process of CH selection guarantees full network coverage. Remember that if in PRWSN, a node receives multiple CH messages; it will joins a CH which is closest to it. After this stage, each node will have one of the two states: CH or cluster-member.
152
A.G. Delavar, J. Artin, and M.M. Tajari
4.2 Route Discovery After the clusters are formed, we should determine the suitable path for transferring data in each cluster to BS. In algorithms like LEACH [9] which use single-hop method for transferring data to BS, the energy of nodes farther from BS will be discharged. Also as we said in related work section, in algorithms like PEGASIS [11] which use multi-hop method, the energy of nodes close to BS will be discharged with a higher rate because of transferring the data of farther nodes. In PRWSN we use a combination of the mentioned methods for exploiting the advantages of both of them. When a node is selected as CH, at first it chooses its next step as BS and during Route discovery step, specifies the optimized route to BS. Then broadcasts a message as "Route discovery" along with its ID, the amount of its remaining energy and distance to BS with a double R Radius (inter-clustered Radius) to make aware its neighbor CH of its existence. Each node performs the following algorithms after receiving a "Route discovery" message and chooses the next-CH for transferring its data to BS: Algorithm 2. Route discovery 1.
2.
if
received a Route-discovery ( Id (S j ) ) message if
d ( S i , S j ) + d ( S j , BS ) < d ( S i , Next − CH ( S i )) + d ( Next − CH ( S i ), BS ) AND E re ( S j ) > (( E re ( S j ) + E re ( S i )) / 2 ) × β )
3. 4.
5.
Next − CH ( Si ) = S j end if end if
In this algorithm, each CH compares the overall collection of its CH distance to itself and its CH to BS with the overall distance of its previous chosen Next-CH after receiving a "Route discovery" message from one CH, A CH will be chosen for transferring its data to BS if the expected amount is less and the CH energy is more than its determined threshold. In PRWSN at first data will be sent in multi-hop method to BS. But after a while the node energies close to BS will become less compared to further nodes and it's better that farther nodes send their data in single-hop to BS .We may do this by applying a threshold (in line 2) in which β ϵ [0.4, 1.4] is a coefficient which specifies the difference of acceptable energy between CH and its next CH for transferring the data to BS. This process of creating route, balances the energy consumption between nodes. Then each node sends its data to its corresponding CH, which after receiving and gathering the data of its cluster members, will send them to a BS.
5 Algorithm Implementation and Performance Evaluation We will analyze the presented algorithm in MATLAB. The parameters used in stimulation, are as following; in which the basic node energy is a random amount between 0.5 and 1; and the nodes are distributed randomly in a quadrangle square perimeter.
PRWSN: A Hybrid Routing Algorithm
153
Table 1. Simulation Parameters
Parameter Network size Number of Nodes Base station location Initial energy for node E elec
ε
Value 500 × 500 m 500 250,250 m rand [0.5,1] J 50nJ/bit 10pj/bit/ m2
fs
ε mp
0.0013pj/bit/m4
Data aggregation energy
5nj/bit/signal
d0
87m
5.1 Comparison of PRWSN, RCSDN and LEACH We have compared the presented Algorithm with the LEACH [9] and RCSDN [15] method. We will use the value of α = 1.05 and β = 0.9 to compare our algorithm to RCSDN and LEACH. The number of alive nodes Figure 1 shows the total number of nodes alive through simulation time. The figure suggests that in PRWSN, the nodes have longer lifetime than that in LEACH and RCSDN. PRWSN reduces energy consumption and then prolong network lifetime. The cause of this is the appropriate distribution of clusters in the network and the consideration of the local state of the node and its neighbors in cluster formation. Network lifetime with different number of nodes We have compared PRWSN in a network with a fixed size of (500×500) with a different number of nodes with the previous algorithms. As figure 2 shows the network lifetime (both the time until the first node dies and the time until 30% nodes die) in PRWSN is considerably more compared to the LEACH and RCSDN. This accounts for using a combined method of single and multi-hop for transferring CH data to BS and using a local threshold detector in the candidate selection. The average of energy consumption in each node per round with different number of nodes We have gained the average of energy consumption in different rounds to get the first dead account in the network. As figure 3 shows, the energy consumption in PRWSN is less than other algorithms. This accounts for using the number of neighbor criterion in selecting CH. Since the nodes which are more in the number of neighbors will be chosen as CH and as a result, the number of clusters will be reduced and less similar data will be sent to each area of BS.
154
A.G. Delavar, J. Artin, and M.M. Tajari
Fig. 1. Total number of alive nodes
Fig. 2. Network lifetime with different number of nodes
The number of clusters in different rounds We have compared the number of clusters in different rounds of the network lifetime with LEACH in a network with a fixed size of (200×200) with 200 nodes. As the figure 4 shows the number of clusters in LEACH method does not have a special balance, and in some rounds this number is very low or very high; but in our presented algorithm the number of clusters in its distribution throughout the network has a good balance, which is because of using local threshold detector, which results in the balance of node energy consumption and increased lifetime of the network.
PRWSN: A Hybrid Routing Algorithm
155
Fig. 3. The average of energy consumption in each node per round
Fig. 4. Total number of clusters in different round
5.2 The Effect of α and β Coefficient in Algorithm Efficiency We have investigated the effect of α and β coefficient in algorithm efficiency. As figure 5 shows, the more the α is in number, the more number of nodes will be chosen candidates, and we may perform a better job for selecting the best clusters based of the criterion of the count of the neighbors. But with the increasing number of candidate nodes, the number of controlling messages exchanged will be increased and as a result the energy consumption of the energy will be increased. We can see that for α=1.05, an optimal value of the network lifetime can be obtained.
156
A.G. Delavar, J. Artin, and M.M. Tajari
Fig. 5. The effect of α coefficient in number of alive nodes
The more we reduce the coefficient β, the accepted threshold in energy difference of CH and the next CH will be reduced and also the performance of algorithms will be like multi-hop methods. As a result the energy of nodes close to BS will be discharged with a higher rate. On the other hand, the more we increase the coefficient β, the harder it will be increased, and the performance of algorithm will be close to single-hop method. As a result, the nodes staying away from BS will be dying with a higher rate. In both states, the lifetime of the network will be reduced. We can see that for β=0.9, an optimal value of the network lifetime can be obtained.
Fig. 6. The effect of β coefficient in number of alive nodes
PRWSN: A Hybrid Routing Algorithm
157
6 Conclusions In this paper we have presented a novel algorithm to form clusters in wireless sensor networks. We have shown that it is possible to create an efficient method for creating clusters in sensor networks by using the average of local energy in each node and a count of their neighbors. Also we can increase the lifetime of the network to a substantial amount by combining single and multi-hop methods. This causes the reduction of loading on nodes which are away from BS or those which are close to BS. We have demonstrated the presented algorithm in an efficient fashion and have compared it with the LEACH and RCSDN methods in cluster formation; in which the results show a higher efficiency level of the PRWSN in node energy reduction and cluster distribution.
References 1. Jeong, W., Nof, S.Y.: Performance evaluation of wireless sensor network protocols for industrial applications. Journal of Intelligent Manufacturing 19(3), 335–345 (2008) 2. Sohrabi, K., et al.: Protocols for self-organization of a wireless sensor network. IEEE Personal Communications 7(5), 16–27 (2000) 3. Min, R., et al.: Low power wireless sensor networks. In: Proceedings of International Conference on VLSI Design, Bangalore, India (January 2001) 4. Rabaey, J.M., et al.: PicoRadio supports ad hoc ultra low power wireless networking. IEEE Computer 33(7), 42–48 (2000) 5. Akyildiz, I.F., et al.: Wireless sensor networks: a survey. Computer Networks 38(4), 393–422 (2002) 6. Akkays, K., Younis, M.: A Survey on Routing Protocols for Wireless Sensor Networks. Elsevier Ad Hoc Network Journal 3(3), 325–349 (2005) 7. Abbasi, A.A., Younis, M.: A survey on clustering algorithms for wireless sensor networks. Computer Communications 30, 2826–2841 (2007) 8. Konstantopoulos, C., Mpitziopoulos, A., Gavalas, D., Pantziou, G.: Effective Determination of Mobile Agent Itineraries for Data Aggregation on Sensor Networks. IEEE Transaction On Knowledge and Data Engineering 22(12) (December 2010) 9. Heinzelman, W.R., Chandrakasan, A., Balakrishnan, H.: Energy efficient communication protocol for wireless sensor networks. In: Proceedings of the 33rd Hawaii International Conference on System Science, vol. 2 (2000) 10. Younis, O., Fahmy, S.: Heed: a hybrid, energy-efficient, distributed clustering approach for ad hoc sensor networks. IEEE Trans. Mobile Comput. 23(4), 366–379 (2004) 11. Lindsey, S., Raghavendra, C.S.: PEGASIS: Power efficient gathering in sensor information systems. In: Proc of IEEE Aerospace Conference, IEEE Aerospace and Electronic Systems Society, Montana, pp. 1125–1130 (2002) 12. Ai, J., Turgut, D., Boloni, L.: A Cluster-Based Energy Balancing Scheme in Heterogeneous Wireless Sensor Networks. In: Proceedings of the 4th International Conference on Networking, Reunion, France, pp. 467–474 (2005) 13. Chen, H., Wu, C.S., Chu, Y.S., Cheng, C.C., Tsai, L.K.: Energy residue aware (ERA) clustering algorithm for leach-based wireless sensor networks. In: 2nd International Conference ICSNC, Cap Esterel, French Riviera, France, p. 40 (August 2007)
158
A.G. Delavar, J. Artin, and M.M. Tajari
14. Wang, Y., Yang, T.L.X., Zhang, D.: An energy efficient and balance hierarchical unequal clustering algorithm for large scale sensor network. Inform. Technol. J. 8(1), 28–38 (2009) 15. Ghorbannia Delavar, A., Artin, J., Tajari, M.M.: RCSDN: a Distributed Balanced Routing Algorithm with Optimized Cluster Distribution. In: 2011 3rd International Conference on Signal Acquisition and Processing, ICSAP (2011) 16. Wang, Q., Yang, W.: Energy consumption model for power management in wireless sensor networks. In: 4th Annual IEEE communications society conference on sensor, mesh and ad hoc communications and network, SECON 2007 (2007)
Cone Tessellation Model for Three-Dimensional Networks Gözde Sarışın and Muhammed Salamah Computer Engineering Department Eastern Mediterranean University KKTC, Mersin 10, Turkey
[email protected],
[email protected] Abstract. Wireless terrestrial networks are usually designed in 2D plane, but in real life they form 3D space. In these networks, node placement strategy is one of the most important design problems. The idea is to deploy a number of nodes in an effective way to achieve communication between them. The volumetric quotient, which is the ratio of the transmission range to the sensing range of each node, is used as the main measure of the placement strategy. Researchers use polyhedrons to model 3D networks. As the volumetric quotient increases, we need less number of nodes for full coverage. In this paper, we proposed a cone model which gives a higher volumetric quotient than polyhedrons. The inspiration comes from satellite foot-print. For example, the number of nodes for truncated octahedron placement strategy is found to be 46.35% higher than the cone placement strategy. We also achieved full coverage with cone tessellation. Keywords: Modeling, 3D networks, tessellation, Kelvin’s Conjecture, Kepler’s Conjecture, Sensor Networks.
1 Introduction In a terrestrial sensor network, the height of the network is usually negligible as compared to its length and width, and as a result a terrestrial network is generally modeled as a two-dimensional (2D) network where it is assumed that all nodes reside on a plane [1]. This assumption may no longer be valid if a network is deployed in space, atmosphere, or ocean, where nodes of a network are distributed over a 3D space. Although such a scenario may not be common at present, applications are being developed that will make three-dimensional networks increasingly common in the near future [2]. Nodes of an underwater sensor network can be deployed at different depths of the ocean. For example, ocean column monitoring requires the nodes to be placed at different depths of the water, thus creating a three dimensional network [3]. Additionally, underwater acoustic ad hoc and sensor networks have generated a lot of interest among the researchers [1], [4], [5], [6]. Weather forecasting and climate monitoring can also benefit if three-dimensional networks can be deployed in the atmosphere [2]. That means we need a good strategy for deploying A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 159–169, 2011. © Springer-Verlag Berlin Heidelberg 2011
160
G. Sarışın and M. Salamah
the nodes in 3D space where we need to sense the environment. In this paper, we focus on the coverage and connectivity issues of three-dimensional networks, where all the nodes have the same sensing range and the same transmission range. In 3D networks, the coverage area of a node forms a sphere. Kelvin’s Conjecture and Kepler’s Conjecture have been used for finding the most efficient shape to fill the sphere. Previous researches used Kelvin’s conjecture to place nodes at the center of polyhedron shapes which are created by Voronoi tessellation in 3D space. Space filling property is very important to cover a sphere and best space-filler regular polyhedrons are cube, dodecahedron, icosahedron, octahedron and tetrahedron [2]. Most of the previous works depends on convex polyhedrons like cube, hexagonal prism, rhombic dodecahedron and truncated octahedron to achieve full coverage and connectivity. Motivated from models like cube, hexagonal prism, rhombic dodecahedron and truncated octahedron, and by assuming the same sensing range (R) and same transmission range, we proposed a Cone model to achieve better coverage and connectivity for 3D networks. However, we placed nodes on vertices of cones based on satellite footprint idea. Our contributions, results, and conclusions of this paper can be summarized as follows: 1) We used volumetric quotient approach, which is the ratio of the volume of a shape to the volume of its circumsphere. We show that the volumetric quotient of cone is 1, much higher than other possible space-filling polyhedron (volumetric quotient should be less than 1 for any polyhedron). 2) We show how to place nodes using any of these placement strategies. For each placement strategy, we define a new u,v,w-coordinate system, where a node should be placed each integer coordinate of this new system. Relation of this new u,v,w-coordinate system with the original given x,y,z- coordinate system has been provided in equations (5), (6),(7) ,(8) and (10) in terms of the sensing range R and the location of an arbitrary node in the original x,y,zcoordinate system (cx,cy,cz). Strategies require only a constant number of arithmetic operations to compute the location of each node and hence is computationally very efficient [2]. We find that cone placement strategy requires that the ratio of transmission range to the sensing range must be at least 2.2360 in order to maintain connectivity among nodes. The rest of this paper is organized as follows: Section 2 presents background information and related work. Section 3 presents our proposed research. Section 4 gives the performance analysis and comparisons. Section 5 provides conclusion and future work.
2 Background Information A main objective in wireless sensor networks is to find the best deployment strategy with minimum number of nodes while 100% coverage is guaranteed [7]. In 2D cellular networks regular hexagon covers the circle, with most efficient way (with fewer gaps), in addition radius of each hexagon is equal to maximum range of a base station. For sensor networks, sensing coverage is very important. Any of the selected
Cone Tessellation Model for Three-Dimensional Networks
161
point in the range should be within the sensing range of at least one sensor. Many algorithms [8], [9], [10], [11] developed to get full sensing coverage in 2D network. Lifetime of a network is another important issue, to increase the lifetime energy harvesting can be used for sensor networks. Also many energy conservation protocols [12], [13], [14], [15] are developed. For a specific time period, only a subset of nodes are active to sense the environment. The effect of sensing coverage on performance was studied for 2D wireless sensor networks in [16], and only [17] for rhombic dodecahedron, and [18] for hexagonal prism studied the 3D cellular networks. Also in [2], they investigated the required number of nodes for truncated octahedron and maximum of the minimum transmission range. In this paper, cone is used to model the shape of cell, and for 3D space we achieved 46.35% fewer nodes than truncated octahedron model. Definition: For any polyhedron, if the maximum distance from its center to any vertex is R, and the volume of that polyhedron is V, then the volumetric quotient of the polyhedron is given as [2] V 4 3
which is the ratio of the volume of a polyhedron to the volume of its circumsphere. 2.1 Volumetric Quotients for Polyhedrons 2.1.1 Cube
length of each side of cube is . radius of its circumsphere is √3 /2. volumetric quotient = = 0.36755 √ /
2.1.2 Hexagonal Prism
length of each side of hexagon is . height of hexagonal prism is h. radius of circumsphere of hexagonal prism is √
volumetric quotient =
√
=
= 0.477
Fig. 1. Hexagonal Prism [2]
/4.
162
G. Sarışın and M. Salamah
2.1.3 Rhombic Dodecahedron
length of each edge of rhombic dodecahedron is √3 /2. total volume of rhombic dodecahedron is 2 . circumradius of a rhombic dodecahedron is . volumetric quotient = = = 0.477
Fig. 2. Rhombic dodecahedron [2]
2.1.4 Truncated Octahedron
length of each edge of truncated octahedron is . volume of truncated octahedron is 8√2 . radius of circumsphere of truncated octahedron is √10/2. volumetric quotient = √ = = 0.68329 √
√
Fig. 3. Truncated Octahedron [2]
More information can be found in [2] for cube, hexagonal prism, rhombic dodecahedron and truncated octahedron.
Cone Tessellation Model for Three-Dimensional Networks
163
2.2 Placement Strategies for Polyhedrons In this section, we explained the results for the placement of a node in the coordinate system briefly. Details can be found in reference [2]. Cube, hexagonal prism, rhombic dodecahedron and truncated octahedron placement can be achieved from the below formulas: For cube =
,
√
,
For hexagonal prism =
2
√
√
,
(1)
√
2
2
For rhombic dodecahedron = For truncated octahedron =
,
√
,
√
,
2 2
(2)
√
√
√
,
(3)
,
The real distances between any two nodes for polyhedrons with coordinates ( and ( , , ) is as follows: For cube =
(4)
√ ,
,
) (5)
√
For heaxagonal prism = = √2
(6)
For rhombic dodecahedron = √2
(7)
For truncated octahedron = 4 √5
3 4
(8)
Figure 4 shows tessellation for hexagonal prism, rhombic dodecahedron and truncated octahedron. More details can be found in reference [2]. They assumed nodes are placed at the center of each node.
3 The Proposed Cone Model Assume all nodes have the same sensing range R. Radius of the sphere is R, and sensor nodes are placed on vertex. Neighbor node should be placed on vertex also. Boundary effects can be negligible. Any point in the 3D space should be within the sensing range R from at least one node. If R is given, we can find the number of nodes. Placement strategy for cone is used to find the minimum number of the transmission range in terms of the sensing range R. (All nodes must be connected to their neighbors). We used volumetric quotient formula idea to find the 3D shape to fill the sphere with less gap.
164
G. Sarışın and M. Salamah
(a)
(b)
(c)
(d)
Fig. 4. Tessellation for rhombic dodecahedron (a), hexagonal prism (b), truncated octahedron (c), and for cube (d) [2]
Recall that Volumetric quotient is always less than 1 for any polyhedron. Our model can achieve the highest volumetric quotient which is 1, without using any polyhedron. Finding the optimal shape is very hard in 3D, it can take many years to prove like Kepler’s conjecture because it still has 99% certain of the correctness (we can accept it as a theorem). We compared our model with four polyhedron shapes and cone has much higher volumetric quotient than others. So, cone needs less nodes than other space filling polyhedrons for coverage in 3D network. Lastly, we developed a
Cone Tessellation Model for Three-Dimensional Networks
165
placement strategy for cone and then we worked on connectivity issue for finding the minimum transmission radius needed to achieve connectivity between neighboring nodes in the placement strategy. optimal height for cone is h = volume of cone is circumsphere of cone is equal to R=h
volumetric quotient approach =
=
=1
Fig. 5. Cone
Based on satellite footprint idea, we found optimal location for cone is its vertex and when we tile the space when vertices are intersected we eliminate one of the node and it helped us with volumetric quotient to use less number of nodes. And also cones does tile a plane in 2D as shown below:
Fig. 6. Elimination of the two duplicated nodes in the same coordinates
Nodes placed at vertex, then we calculated the location of the node according to x, y and z axis. Suppose that the coordinate system is defined by three axes : , and , which are parallel to the x, y and z axes, respectively. For a node, unit distance in u direction is which is equal to 5R. Node is placed at 2R, v axis is R and w axis is 2
166
G. Sarışın and M. Salamah
2 ,
,
5
(9)
Optimal height for cone to achieve highest volumetric quotient is h/2. R is the sensing range and it is equal to h for cone (R=h). The real distance between two points with and in the u, v, w- coordinate is coordinates , , , , =
4
5
(10)
There is a tessellation model for better understanding the node placement strategy and it is done by 3DMax. More nodes are needed to cover the area if network size increases. However for same network size, cone needs less number of nodes to achieve full coverage. Figure 7 shows the node placement for cone which is done by 3DMax.
Fig. 7. Tessellation for the cone model
4 Performance Analysis In 2D, there is no shape that can fill the plane without gaps, therefore volumetric quotient approach can never be exactly 1. But we have an advantage in 3D. The cells can fill gaps when we deploy the cells carefully (to have a full filled space). Figure 6 can give an idea about our node deployment strategy. If we have a constant sensing range, cone needs less cells to fill a specific 3D space. If we consider all models, cone gives the best volumetric quotient value according to our approach. As we mentioned before volumetric quotient is characteristic issue for determining number of nodes. Cube needs 1⁄0.36755 2.7207 times that of cone.
Cone Tessellation Model for Three-Dimensional Networks
167
For hexagonal prism and rhombic dodecahedron it is 1⁄0.477 2.0964 and for truncated octahedron it is 1⁄0.68329 1.4635. Table 1 shows the comparisons according to cone model. The achieved saving ratios are shown in the last column. Table 1. Volumetric Quotients of Models and Number of Nodes Compared to Cone Model
Volumetric Quotient
Number of Nodes Needed Compared to Cone
Cube Hexagonal Prism Rhombic Dodecahedron Truncated Octahedron Cone
0.36755 0.477 0.477
272.07 209.64 209.64
Saving Ratio Compared to Cone 172% 110% 110%
0.68329
146.35
46%
1
100
0%
Minimum transmission range is important to achieve connectivity between neighboring nodes and it depends on the choice of the model. The distance between two neighboring nodes for cube is 2R√3, then the transmission range must be at least 1.1547R. For hexagonal prism this value is √2R=1.4142R along the axes u and v, and 2R/√3=1.1547R along the w axis. For rhombic dodecahedron, the minimum transmission range is √2R=1.4142R for both axes. If the truncated octahedron is used, the transmission range must be at least 4R√5=1.7889R along u and v axes, and for w axis this value is 2√3/√5 =1.5492R. Finally, for cone the transmission range must be at least 2R for u axis, R for v axis and R√5= 2.2360 R for w axis. Table 2. shows the minimum transmission range for these different models. It is clear from the table that the proposed cone model manifests its superiority in terms of transmission range as well. Table 2. Comparison of Minimum Transmission Ranges for Different Models Model
Minimum Transmission Range uywaxis axis axis
Cube
Maximum of the Minimum Transmission Range
1.1547R 1.1547R
1.1547R
1.1547R
Hexagonal Prism
1.41 42R
1.41 42R
1.15 47R
1.4142R
Rhombic Dodecahedron
1.41 42R
1.41 42R
1.41 42R
1.4142R
Truncated Octahedron
1.78 79R
1.78 79R
1.54 92R
1.7889R
R
2.23 60R
2.2360R
Cone
2R
168
G. Sarışın and M. Salamah
5 Conclusion In this paper, we proposed a new model to place nodes in 3D space, unlike today’s networks (where they assume nodes are placed on 2D plane). Deployment of the nodes is not very easy in 3D. For 2D networks, hexagonal tiling is the best strategy for placing the base stations, so the covered area is maximized with fixed radius base stations. Here, the problem for 3D space is filling the empty spaces between nodes. Related works on this topic just have the polyhedron shapes to achieve space-filling property in 3D. We used the idea of volumetric quotient, which is the ratio of the volume of a polyhedron to the volume of its circumsphere, to compare different models. The proposed cone model results in the highest volumetric quotient which is 1. This shows with cones, one can cover the sphere better than polyhedrons. Consequently, the number of nodes required for coverage is changed. For example, if we apply truncated octahedron placement strategy, it needs 46% more nodes than the cone placement strategy to cover same network size. From other related models like cube, hexagonal prism, rhombic and dodecahedron, the achieved saving in terms of the number of nodes are 172%, 110% and 110% respectively. After finding the optimal placement strategy, we investigated the connectivity issues and we found that the best placement strategy is with our cone model which requires the transmission range to be at least 2.2360 times the sensing range in order to maintain full connectivity. For cube, hexagonal prism, rhombic dodecahedron and truncated octahedron transmission ranges are 1.1547, 1.4142, 1.4142 and 1.7889 respectively. We believe that our model can be used in many research areas for 3D networks.
References [1] Alam, S.M.N., Haas, Z.: Coverage and Connectivity in three-dimensional underwater sensor networks.Wireless communication and mobile computing (2008), http://www.interscience.wiley.com [2] Alam, S.M.N., Haas, Z.: Coverage and Connectivity in three-dimensional networks. In: Proceedings of ACM MobiCom (2006) [3] Akyildiz, I.F., Pompili, D., Melodia, T.: Underwater Acoustic Sensor Networks: Research Challenges. Ad Hoc Networks Journal (Elsevier), (March 2005) [4] Heidemann, J., Ye, W., Wills, J., Syed, A., Li, Y.: Research Challenges and Applications for Underwater Sensor Networking. In: IEEE Wireless Communications and Networking Conference. IEEE, Las Vegas (2006) (p. to appear) [5] Kong, J., Cui, J., Wu, D., Gerla, M.: Building Underwater Ad-hoc Networks and Sensor Networks for Large Scale Real-time Aquatic Applications. In: IEEE Military Communications Conference (MILCOM 2005), Atlantic City, New Jersey, USA, October 17-20 (2005) [6] Vasilescu, I., Kotay, K., Rus, D., Dunbabin, M., Corke, P.: Data Collection, Storage, and Retrieval with an Underwater Sensor Network. In: SenSys 2005, San Diego, California, USA (November 2–4, 2005) [7] Rappaport, T.S.: Wireless Communications: Principles and Practice. Prentice-Hall, Englewood Cliffs (2002)
Cone Tessellation Model for Three-Dimensional Networks
169
[8] Couqueur, T., Phipatanasuphorn, V., Ramanathan, P., Saluja, K.K.: Sensor deployment strategy for target detection. In: Proceeding of the First ACM International Workshop on Wireless Sensor Networks and Applications, pp. 169–177 (September 2002) [9] Chakrabarty, K., Iyengar, S.S., Qi, H., Cho, E.: Grid coverage for surveillance and target location in distributed sensor networks. IEEE Transactions on Computers 51(12), 1448–1453 (2002) [10] Meguerdichian, S., Koushanfar, F., Potkonjak, M., Srivastava, M.B.: Coverage problems in wireless ad-hoc sensor networks. In: INFOCOM 2001, pp. 1380–1387 (2001) [11] Zhang, H., Hou, J.C.: Maintaining sensing coverage and connectivity in large sensor networks. Wireless Ad Hoc and Sensor Networks: An International Journal 1(1-2), 89–123 (2005) [12] Tian, D., Georganas, N.D.: A coverage-preserved node scheduling scheme for large wireless sensor networks. In: Proceedings of First International Workshop on Wireless Sensor Networks and Applications (WSNAm 2002), Atlanta, USA, pp. 169–177 (September 2002) [13] Wang, X., Xing, G., Zhang, Y., Lu, C., Pless, R., Gill, C.D.: Integrated coverage and connectivity configuration in wireless sensor networks. In: Sensys (2003) [14] Yan, T., He, T., Stankovic, J.A.: Differentiated surveillance for sensor networks. In: SenSys 2003: Proceedings of the 1st international conference on Embedded networked sensor systems (2003) [15] Ye, F., Zhong, G., Lu, S., Zhang, L.: Peas: A robust energy conserving protocol for longlived sensor networks. In: 23rd International Conference on Distributed Computing Systems, ICDCS 2003, pp. 169–177 (May 2003) [16] Xing, G., Lu, C., Pless, R., Huang, Q.: On Greedy Geographic Routing Algorithms in Sensing-Covered Networks. In: Proc. of MobiHoc 2004, Tokyo, Japan (2004) [17] Carle, J., Myoupo, J.F., Semé, D.: A Basis for 3-D Cellular Networks. In: Proc. of the 15th International Conference on Information Networking (2001) [18] Decayeux, C., Semé, D.: A New Model for 3-D Cellular Mobile Networks. In: ISPDC/HeteroPar (2004)
Post Disaster Management Using Delay Tolerant Network Sujoy Saha1, Sushovan2, Anirudh Sheldekar2, Rijo Joseph C.1, Amartya Mukherjee2, and Subrata Nandi2 1 Department Of Computer Application , Department of Computer Science and Engg National Institute of Technology, Durgapur, Durgapur-713209, India {sujoy.ju,bubususpatra,anisheld, mail2rjc,mamartyacse1,subrata.nandi}@gmail.com 2
Abstract. Delay-tolerant Networking (DTN) is an attempt to extend the reach of traditional networking methods where nodes are intermittently connected and an end-to-end path from source to destination does not exist all the time. Real networks like military, various sensors, post disaster management, deep space communication, Vehicular ad-hoc (VANETs) networks, are some examples of DTN. Our work mainly concentrates on the applicability of different flooding based routing scheme of DTN in post disaster scenarios. Cluster mobility model which maps human mobility more realistically rather than any other mobility in the context of disaster scenario has been considered. Further we have customized cluster mobility model according to the disaster like scenario and performed the simulation for delivery probability with respect to various constraints like buffer-size, transmission range, speed and density of nodes in ONE SIMULATOR. We also analyze the effect heterogeneous nodes in delivery probability. Keywords: Disaster Management, Cluster Mobility Model, Heterogeneous Network, Delivery Probability, Overhead Ratio, Average Latency.
1 Introduction In disaster affected areas the existing communication infrastructures like WLL, GSM or PSTN may get disrupted or destroyed. Thus, there exists a strong need for rapid deployment of communication networks that would provide much needed connectivity and communication capabilities for rescue-workers and survivors of a disaster affected zone to restore normalcy through properly co-ordinate resource management. For managing a post disaster situation, the prime requirement is to establish communication among disaster management groups or agencies. There will be different teams working together for managing the distribution of necessary commodities for the affected population in disaster-affected regions [1][2]. Information must be relayed and reveled in the shortest amount of time possible in A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 170–184, 2011. © Springer-Verlag Berlin Heidelberg 2011
Post Disaster Management Using Delay Tolerant Network
171
order to co-ordinate and perform required activities. Disaster response network is one kind of delay tolerant network [3]. In a disaster scenario like a fire burst or a natural calamity, the communication between the nodes no longer remains organized in the original infra-structural setting. The original assembly of nodes changes with the nodes moving in groups of people helping for the cause. In these kinds of networks traditional Ad-Hoc routing protocols fail to transfer messages from source to destination. A delay tolerant network differs from Ad-Hoc network due to the simple fact that message would be transferred to the destination node even if the source has no end-to-end connectivity with the destination at the time when that message is sent. So delay tolerant routing strategies are employed for efficient packet delivery among the nodes of such networks. A disaster environment could be modeled as a specialized mobility model, since disaster management always takes place amongst groups of people. Consider some disaster scenarios like Cyclone in some specific area, earth-quake, burst of fire etc. Let the place of the accident be termed as Activity Point. Now there would be various groups of people moving around the activity point like Medical Staff, Police, people, etc. Thus, a group-based movement model would be a good choice for such a scenario where the mobile nodes exists in groups and communication takes place within the group as well as between the groups. Now, in such scenarios, the movement of node groups will be restrained to fixed local sites like Hospital, Police station, Activity point, etc. Thus, we can consider the scenario as one with different clusters of nodes that restrain to particular sites. Vehicles that move across these sites like police jeeps, ambulances and other relief vehicles can be carrier nodes between the clusters. The Cluster Mobility Model [4] can be used to model the movements of node for this type of scenario. In the next section of this paper, we summarize some other mobility models such as Random Waypoint, Random Walk, Shortest Path Map Based and Working Day movement models in order to justify our choice of Cluster Mobility Model as the movement model for the scenario. In Section 3 we summarize about the different routing strategies that exists for DTN. The most challenging issue in the post disaster environment is the rate of transmission of critical information. To enhance the packet delivery ratio we require intelligent DTN routing strategies. In Section 4 we have described and analyzed the simulation results of delivery probability that has been carried out for various routing algorithms on cluster mobility model for post disaster scenario with respect to buffer-size, transmission range, speed and density of nodes in the network. The effect of the heterogeneous nodes in delivery ratio in the context of DTN is also explored. The constraints are so chosen as to derive an optimal configuration for the nodes to be deployed for communication in postdisaster scenarios.
2 Mobility Model Overview Mobility model helps to emulate closely the real life scenario of mobile nodes. All mobility models are based on some basic parameters like starting location, ending point, velocity of mobile node, movement direction. Works have been carried out on mobility models seeking to increase their realism in simulations by gathering
172
S. Saha et al.
information on existing scenarios to provide insights of node mobility and how they affect the performance of routing algorithms. Significance: In scheduled or predictable contacts it is possible to predict the future in terms of the time for which contacts will be available and how long they will last. However, in disaster recovery networks, it is almost impossible to predict the future location of the nodes. Communication is to be enabled in such networks using those unpredictable contacts between the nodes which are also known as intermittent or opportunistic contacts. It is extremely important in DTN to understand the mobility of the relay nodes that carry messages from source to destination [4]. Mobility models establish relationship among individuals and help us to study their movements in real life. It is extremely important in DTN to understand the mobility of the relay nodes that carry messages from source to destination [4]. Even if few nodes in the network are mobile and others are static, then they might block the flow of data from source to destination. If majority of the nodes in the network are mobile, then the routing protocols will have more opportunities to deliver the message to the destination by exploring the mobility of the relay nodes. An example of this type of network is a vehicular network where the cars, trucks are all mobile nodes. Since real life experiments are not feasible, we resort to simulation experiments which give us reallike results. Mobility models establish relationship among individuals and help us to study their movements in real life. Mobility models can be broadly classified into Entity-Based mobility model and Group-based mobility models [10]. In the former model, the nodes move individually and their movement is not influenced by the other nodes whereas the in the latter the movement of nodes is influenced by that of the member nodes. Entity Based models generate results that are more non-human like. On the other hand, group mobility model provide results which are more real, as human mobility occurs mainly in groups. Random Waypoint [5][8 ] model is a very common Entity-Based mobility model in which each mobile node randomly selects one point as its destination and travels towards this destination with constant velocity chosen uniformly and randomly from [0, Vmax ]. Upon reaching the destination, the node stops for a duration defined by the ‘pause time’ parameter Tpause. After this duration, it again chooses another random destination and moves towards it. Random Walk [6] [8] is another Entity-Based movement model and can be considered as a type of Random Waypoint model with zero pause time. In Random Walk model, nodes change their speed and direction after a time interval. Each and every node randomly and uniformly chooses its new direction θ(t) from [0, 2π ] for every new interval t. Similarly, a new speed, v(t), is chosen from [0, Vmax] uniformly. Thus, during any time interval t, a node moves with the velocity vector (v(t).cosθ (t), v(t).sin θ (t)). If a node moves and touches the boundary of the simulation area, it gets bounced back to the simulation area with an angle of θ(t) or π − θ(t). Shortest Path Map Based [8] mobility model is a map based movement model that uses algorithms like Dijkstra's algorithm to find shortest path between two random map points. This model is also an Entity-Based movement model. Working day mobility [9] model is a Group-based movement model [10]. This model basically is the technical resultant of different sub-models of node mobility
Post Disaster Management Using Delay Tolerant Network
173
during the whole day. This model involves activities that are the most common and capture most of a working day for the majority of people. However, the activities of nodes differ from each other. These sub-models repeat every day, resulting in periodic repetitive movement. Cluster Mobility Model: As the name suggests this mobility model classifies the whole network in number of clusters. Depending upon the applicability and mobility, literature of cluster mobility model categorizes the nodes in two different ways. The nodes responsible for carrying data from one cluster to another or maintaining inter cluster relationship are known to be Carrier nodes. Other than Carrier nodes all the other nodes present inside the cluster are treated as internal nodes. Movement of the internal node is defined around a particular point within the cluster which is known as Cluster Center and move around this cluster center. Cluster mobility model falls under the umbrella of Group based mobility model which unlike Random mobility model try to establish a social relationship between nodes within the network based on their activities to define the cluster first.
Fig. 1. Snapshot of Cluster Mobility Model from ONE simulator
Due to social status, relationship, profession, and friendship human does have a tendency to move in group. Secondly this mobility model certainly makes sense in disaster and defense activities. From the theoretical point of view cluster mobility model certainly outperforms other mobility models in the context of mapping the human mobility in disaster scenario where human moves in a group. That actually motivates our work to simulate routing strategies cluster mobility model and explore the future directions. A post-disaster scenario can be easily modeled in cluster mobility model. Groups of people could be considered as clusters and the node movements could be modeled as movement of these people within and across the clusters. For example, consider a point in a city where a disaster strikes. The fire-station that involves in the postdisaster management can be mapped as a cluster and the firemen with communicating devices could be matched to the nodes of that cluster. A hospital could be considered as another cluster with doctors, nurses and other supporting staff matched as nodes of
174
S. Saha et al.
that cluster. A police station could be another cluster of nodes with policemen matched to nodes. The point in the city where the disaster has struck or hospital would soon become a cluster of nodes with rescuers and relief-teams including firemen, policemen, doctors, nurses and others who would rush towards the spot for post-disaster activities, thus making those as Activity Points.
Fig. 2. Activity points as clusters in a sample city-like scenario
The nodes involved in these rescue activities will start moving within the clusters as well as across them. It can be noted that at any point of time, majority of the nodes will be moving within some cluster with lower speeds and only a few nodes will be moving across the clusters and that too with higher speeds. Such a scenario basically resembles the Cluster Mobility Model rather than any other traditional mobility models.
3 Routing Protocol Overview In DTN literature, routing protocols are broadly categorized as Forwarding based or Flooding based depending upon whether or not the protocol creates message-replicas (copies of the same message) or not. Routing Protocols that use only a single copy of the message are called as Forwarding Based routing protocols. On the other hand routing protocols that do create more than one copy of the message are called as Flooding Based [10] protocols. Further, Flooding based routing algorithms [13] can be classified as Direct contact, Tree-based flooding, Exchange based flooding and Utility based flooding. Owing to the dynamicity of DTN one has to choose the suitable routing algorithm for message delivery. With the help of simulations, we attempt to study, analyze and discuss the performance of different routing schemes in cluster mobility model which maps human mobility in the best possible way in a post disaster perspective. Here, we will be considering only the flooding based routing protocols and we are only
Post Disaster Management Using Delay Tolerant Network
175
bothering about successful timely delivery of the message rather than concentrating on the overheads incurred. Flooding Families [16]: Routing protocols that belong to these families make use of replication technique. In our work we are taking the flooding algorithms like Epidemic Routing, PRoPHET, Spray & Wait, Spray & Focus and MaxProp. Epidemic routing [11], guarantees that through sufficient number of exchanges, all nodes will eventually receive the message. The nodes maintain a Summary Vector that will keep track of the messages they generate or receive during message delivery using unique message IDs. When two nodes meet they exchange their summary vectors and request the exchange of the messages they do not have. Extreme flooding in this routing technique leads to heavier resource consumption [2][11]. In PRoPHET [12] when two nodes meet, they exchange Summary Vectors which also contain the delivery predictability information stored at the nodes. Nodes make use of this information to update their internal delivery predictability vector. The information is also used to find which messages are to be requested from the other node. A node forwards a message to another node or multiple nodes, if the delivery predictability is higher than a fixed threshold value [4] [12]. MaxProp [13] routing algorithm is knowledge based flooding routing algorithm. It also works similar to Epidemic by trying to replicate and transfer message copies to whomever coming in contact. However, each node maintains a delivery likelihood vector, obtained by doing incremental averaging. When two nodes meet, these vectors are also exchanged. With the help of this vector each node can calculate the shortest path to the destination. Another specialty of MaxProp is its use of acknowledgments to remove the delivered messages from the buffers of all nodes thereby preserving resources for the use of undelivered messages. In MaxProp the nodes maintain a list of previous relays too in order to prevent data getting relayed for a second time to the same node. In Spray and Wait [14] the number of copies of a message in the network is limited in order to reduce the overhead of extensive flooding in message forwarding. It has two phases in routing: Spray Phase and Wait Phase. When a new message gets generated at the source and needs to be routed to a given destination, Spray and Wait algorithm first enters the “Spray phase” for this message. When a message is generated at the source it also creates L forwarding tokens for this message. Whenever two nodes encounter, they exchange those messages that the other node does not have based on number of forwarding tokens left for each message. Thus n copies of message m are spread to n distinct nodes in this phase. In Wait phase, each of n nodes carrying copy of message m waits for a chance to perform a direct delivery of message to the final destination. Spray and Focus [15] is an extension of Spray and Wait. Spray Phase in Spray and Focus algorithm is same as that in Spray and Wait Routing algorithm. When a relay has only one forwarding token for a given message, it switches to the “Focus phase”. Unlike Spray and Wait, where messages are routed using Direct Transmission [16][17] in the Wait phase, in the Focus phase of Spray and Focus a message can be forwarded to a different relay according to a given forwarding criterion.
176
S. Saha et al.
4 Simulation Result Simulation has been carried out in ONE simulator version 1.4.1. Five routing algorithms namely Epidemic, PRoPHET, Spray and Wait, MaxProp and Spray and Focus were simulated in the post-disaster scenario modeled on Cluster mobility. This section explains the environment modeling parameters and performance analysis metrics that were chosen and also analyses of the results of the simulations. 4.1 Environment Model Parameters of Simulation, Routing Algorithms and Mobility Model are specified in Table1, Table2 and Table3. Simulations were run for 24hrs with an update interval of 1s. Nodes have a 500MB buffer. Since scan interval is taken as 0s, nodes continuously scan for neighbors. Speed of cluster nodes is kept as 1.8kmph – 5.4kmph (pedestrian speed) and waittime as 0min – 2min in order to mimic the movement of rescuers in the scenario. Similarly, the carrier nodes have a speed of 18kmph – 54kmph and wait-time of Table 1. Simulation Parameters considered for ONE Simulator
Parameter Simulation Time Update Interval No. of nodes Buffer size of nodes Sp eed
Cluster Nodes Carrier Nodes
Scan interval of nodes Cluster WaitCarrier Time Nodes Message TTL MAC Protocol Bluetoo Range th Data rate Range Wi-Fi Data rate Message Creation Interval Message Size Simulation Area Size
Value 86400s = 24hrs 1s 120 ((25nodes × 4clusters) + 20carrier_nodes) 500MB 0.5mps – 1.5mps = 1.8kmph – 5.4kmph 5mps – 15mps = 18kmph – 54kmph 0s 0min – 2min 0min – 10min 240min = 4h 802.11, 802.15.1 10m 2Mbps 40m 18Mbps 25s – 120s 50KB – 1MB 15.3 sq.km (4.5km x 3.4km)
Post Disaster Management Using Delay Tolerant Network
177
Table 2. Parameters of Routing Algorithms
Routing Algorithm
Parameter
Epidemic PRoPHET MaxProp Spray And Wait Spray And Focus
Val ue N/ A
N/A Seconds In Time Unit ProbSet maximum size No. of Copies Binary Mode No. of Copies Binary Mode
30s 50 3 TR UE 3 TR UE
Table 3. Parameters of Mobility Model
Parameter No. of clusters Cluster Radius No. of nodes in a cluster No. of carrier nodes
Value 4 800m 25 20
0min – 10min. Wait-time is the time for which a node waits or pauses on reaching its destination. In all the simulations nodes uses Bluetooth interface with a range of 10m and data rate of 2Mbps, except in heterogeneous network scenario where some percent of nodes have Bluetooth interface and others have Wi-Fi interface with a range of 40m and data rate of 18Mbps. After every 25s – 120s any one node generates a message of size 50KB – 1MB, to be delivered to any other node in the network. In PRoPHET, if a pair of nodes does not encounter each other in a while, the delivery predictability values age. The aging equation is shown below:
where γ є [0, 1) is the aging constant, and k is the number of time units that have elapsed since the last time the metric was aged. In the simulations for PRoPHET 30s of simulation time makes one time unit, as given in Table2. In the simulations for MaxProp each node can estimate and maintain delivery likelihood values for a maximum of 50 neighbors, as given in Table2.
178
S. Saha et al.
Spray and Focus and Spray and Wait operates in binary mode and the number of copies of a message is limited to 3, a near to optimal value considering the number of nodes in each cluster. 4.2 Performance Metrics The metrics that are chosen to analyze the performance of the routing algorithms are Delivery probability, Overhead ratio and Average latency. Delivery probability is the ratio of number of delivered messages to that of created messages, making it a good metric to measure the efficiency of routing algorithms in delay tolerant scenarios.
Overhead ratio is calculated as the difference of relayed and delivered number of messages upon number of delivered messages. Overhead ratio thus gives a measure of the overhead incurred by the routing schemes in delivering messages.
Latency of a message delivery is the time elapsed from the creation of a message at source to its successful delivery at the destination. Thus Average latency is the average of latencies of all those successful message deliveries. 4.3 Results and Discussion Simulations were performed with varying constraints of buffer size, transmission range, Bluetooth interface density, Carrier node speed and Message size. Buffer size and transmission range were chosen in order to check the dependency of the routing algorithms on the factors that are device-dependent. Message size was chosen in order to study its effect on the bandwidth and buffer usage. Analysis on carrier node speed was done to find the effect of indirect delays in message delivery resulting from the speed variations of carrier nodes. Bluetooth interface density was chosen to study the effect of introducing heterogeneity in the scenario. 4.3.1 Delivery Probability and Overhead Ratio with Respect to Buffer Size From the simulation results plotted in Fig. 3, it can be seen that Spray and Wait does not produce higher delivery probability although it manages to set a lower benchmark in overhead ratio than the other flooding schemes. Low overhead and less delivery probability of Spray and Wait is a resultant effect of Wait Phase mainly. On the other hand Spray and Focus put up effective delivery probability with less overhead ratio in smaller buffer size. But as the buffer size increases the number of message relayed in Spray and Focus also increases which boosts up the overhead ratio. Epidemic and PRoPHET, two basic flooding schemes, start with higher overhead ratio. PRoPHET manages to outperform Epidemic in both parameter and set up higher benchmark in delivery probability than all other flooding schemes due to restricted flooding as well as probability based message delivery.
Post Disaster Management Using Delay Tolerant Network
179
Fig. 3. Performance of routing algorithms on varying Buffer size
Even though MaxProp shows best performance at lower buffer sizes, PRoPHET outperforms it at higher buffer sizes. The performance of MaxProp owes to the dynamic delivery probability calculation, application of Dijkstra’s algorithm and other complimentary mechanism. Its starts with very high overhead ratio due to transformation of the entire message destined for neighbors, relays of routing information vector to other nodes in the network as well as generating acknowledgement for all delivered message. However, it can be seen that above 60MB, the overhead incurred by MaxProp is slightly less than that of Epidemic itself. 4.3.2
Delivery Probability and Overhead Ratio with Respect to Transmission Range In a post-disaster scenario, the constraint of transmission range of nodes can be a real barrier to achieve good delivery ratio. Higher transmission ranges trades for higher power consumption which cannot be much tolerated by mobile nodes, especially in this scenario. In cluster mobility model we can relate both of these two terminologies called: Transmission range and Node Density. Both of these are products of increment of number of nodes within the network. So increment of transmission range for each of the node will cause identification of larger number of neighbors. On the other hand, node density severely affects the sparse nature of the network. All the flooding schemes in our simulation produce much better delivery probability with the increment of transmission range. But over head ratio differs a lot depending upon the number of copies made by particular routing strategies in order to ensure successful delivery of the message. From the simulation results plotted in Fig. 4, it can be seen that Epidemic, PRoPHET and MaxProp performs quite well as number of identified neighbors in single scan is large which is technically equivalent to increasing the number of copies. But these flooding schemes have shown tendency to produce huge overhead ratio with the gradual increment of transmission range. Spray and Wait scheme achieves lowest over head ratio because it does not deliver the single copy of the message at Wait Phase until there is a direct contact with the destination. But this wait for direct
180
S. Saha et al.
Fig. 4. Performance of routing algorithms for varying Transmit range
contact makes Spray and Wait vulnerable in the context of delivery probability. Spray and Focus is challenged by the initial time it takes to calculate the utility function and difficulties it might face to explore the network due to sudden identification of huge number of nodes. MaxProp achieves the highest deliver probability at high transmit ranges. It has shown optimum result when transmission range was kept 20-30 meters. It almost achieves .85 to .90 of delivery probability. But with the increment in transmission range it shows inclinations towards higher overhead ratio. As the number of internal nodes as well as carrier nodes does not increase generally, Spray and Focus also is a good enough routing algorithm to count on. 4.3.3 Delivery Probability and Overhead Ratio in Heterogeneous Network Structure Here one of the most realistic environments is chosen where we have varied the number of nodes with Bluetooth interface and gateway nodes which have both the interfaces of BT and Wi-Fi. Initially all the nodes are Wi-Fi interface enabled and we have increased this value until all the nodes are only having Bluetooth interface. Hence this scenario is much more practical than the previously discussed scenarios. As Wi-Fi interface does really mean increment of Transmission range and data rate, all the Flooding and Spraying Schemes achieves higher delivery probability when all the node are having Wi-Fi interface as can be seen in Fig.5. We have seen before that overhead ratio of Epidemic, PRoPHET and MaxProp are directly proportional with the transmission range. Here also, as Wi-Fi interface results in higher transmission range, overhead ratio increases for all of the above mentioned schemes. On the other hand overhead ratio is inversely proportional with the transmission range in case of both of the Spraying Schemes here it has shown exactly same result.
Post Disaster Management Using Delay Tolerant Network
181
Fig. 5. Performance of routing algorithms for varying Bluetooth interface percentage
4.3.4 Delivery Probability and Overhead Ratio in Carrier Nodes Speed Mobility of nodes is exploited in DTNs for relaying the message from source to destination. Speed of the nodes has got a lot to do with the timely delivery of message to the destination, which is of extreme importance in post-disaster scenario. Node Speed is very important issue in time of Post Disaster Management. Here we take realistic human walking of 1-5 Km/hr and varying the Carrier node speeds. The key thing to observe from the graphs in Fig.6 is that performance (delivery probability) differs substantially among the routing algorithms in cluster mobility model. Here we observed that the overhead ratio and the average latency decreased when we increased the carrier node speeds in comparison to other relative parameters like buffer size, transmit range etc, and it goes to constant except Spray and Focus routing Algorithm. Due to the high speed of carrier nodes, packets are brought in very short time to the
Fig. 6. Performance of routing algorithms for varying carrier node speed
182
S. Saha et al.
adjacent cluster. However, carrier nodes pause for a wait-time when these nodes are inside a cluster. Atan optimal speed of 5-10 m/s all the routing algorithms gives very good delivery probability with lower Average latency and Overhead Ratio. 4.3.5 Delivery Probability and Overhead Ratio in Message Size Message size is a challenging issue in the context of Social Network Structures. Increment of the Message Size is functionally dependent on sparse nature of the network as well as scalability of the network.
Fig. 7. Performance of routing algorithms for varying message size
As can be seen in the graph of Fig.7, the performance of all the routing strategies is severely challenged by the increment of message size above 500KB. The lower data rate (2Mbps) along with the reduced contact times of nodes can be a reason for this drop in performance. Since the messages has to be passed atomically in store-andforward message switching, successful node-to-node transfer of large sized messages is much difficult to achieve within the constraints of reduced contact times and low data rates. Since TTL value of the messages is taken as 4hrs, the buffer size limit of 500MB will not affect the performance much at lower message sizes. But when the message sizes are sufficiently big, the limited buffer size can also contribute to the drop in performance. In order to accommodate newer messages into their buffers nodes may drop older ones, magnifying the effect of increased message sizes on the performance of routing schemes. One interesting fact that can be noted from the results is that the overhead ratio is higher for Spray and Focus in most of the cases. This can be a side-effect of the forwarding technique used by the algorithm to focus the message to the destination in the focus phase. From the simulations it was noted that messages are getting carried away through longer relay-transfers in Spray and Focus than any other routing
Post Disaster Management Using Delay Tolerant Network
183
algorithm and many messages were even relayed through cyclic paths, thereby increasing the number of relayed messages.
5 Conclusion In this paper we have addressed Delay Tolerant Networking for monitoring disaster strike areas where infrastructure-based as well as Ad-Hoc networks fail to communicate owing to the unavailability of end-to-end connectivity and fully connected network. We simulated flooding and spraying based DTN Routing Algorithms where PRoPHET and MaxProp outperformed all other routing algorithms in cluster mobility model. Our work seems to be the first time to include Cluster mobility model for use in real life application like post disaster management. As the dimension of human communications and mobility are getting dynamic day by day, there are greater scopes to explore and modify the mobility model mentioned here. Disaster scenarios of Cyclone and Earthquake prone zones, coastal areas where transport communication system is quite different from city like environment, offer new challenges to merge the usability of infrastructure based network and DTN. This is a new dimension of research which we have kept for future works.
References [1] Mazumdar, C., Das, J., Saha, S., Upadhyay, M., Saha, S.: Rapidly Deployable Wireless data Communication Network (RDWCN) for Disaster Management- An Experiment. In: 20th Indian Engineering Congress, Kolkata, West Bengal, December 15-18 (2005) [2] Das, J., Saha, S., Kundu, A., Upadhyay, M., Chatterjee, K., Saha, S.: Rapidly Deployable Decentralized Disaster Management System and Information Network for Rural Areas. Presented at 37th IETE Mid – Term Symposium on Information Communication Technology – Initiative for Rural Development (ICTIRD 2006), Kolkata, West Bengal (April 2006) [3] Patra, S., Balaji, A., Saha, S., Mukherjee, A., Nandi, S.: A Qualitative Survey on Unicast Routing Algorithms in Delay Tolerant Networks. In: Proc. of AIM2011S, Nagpur (2011) [4] Uddin, Y.S., Nicol, D.M.: A Post-Disaster Mobility Model For Delay Tolerant Networking. In: Rossetti, M.D., Hill, R.R., Johansson, B., Dunkin, A., Ingalls, R.G. (eds.) Proceedings of the 2009 Winter Simulation Conference (2009) [5] Romoozi, M., Babaei, H., Fathy, M., Romoozi, M.: A Cluster-Based Mobility Model for Intelligent Nodes at Proceeding ICCSA 2009. In: Proceedings of the International Conference on Computational Science and Its Applications: Part I (2009) [6] Broch, J., Maltz, D.A., Johnson, D.B., Hu, Y.-C., Jetcheva, J.: A performance comparison of multi-hop wireless ad hoc network routing protocols. In: Proceedings of the Fourth Annual ACM/IEEE International Conference on Mobile Computing and Networking(Mobicom 1998), ACM, New York (1998) [7] Camp, T., Boleng, J., Davies, V.: A Survey of Mobility Models for Ad Hoc Network Research. In: Wireless Communication and Mobile Computing (WCMC): Special issue on Mobile Ad Hoc Networking: Research, Trends and Applications, vol. 2(5), pp. 483–502 (2002) [8] A Survey Of Mobility Models in Wireless Adhoc Networks Fan Bai and Ahmed Helmy University of Southern California, USA
184
S. Saha et al.
[9] Kumar, S., Sharma, S.C., Suman, B.: Mobility Metrics Based Classification & Analysis of Mobility Model for Tactical Network. International Journal of Next-Generation Networks (IJNGN) 2(3) (September 2010) [10] Ekman, F., Keränen, A., Karvo, J., Ott, J.: Working Day Movement Model. In: 1st SIGMOBILE Workshop on Mobility Models for Networking Research, Hong Kong (May 2008) [11] Camp, T., Boleng, J., Davies, V.: A Survey of Mobility Models for Ad Hoc Network Research. In: Wireless Communication & Mobile Computing(WCMC): Special issue on Mobile Ad Hoc Networking: research. Trends and Applications, vol. 2(5), pp. 483–502 (2002) [12] Evan, P.C., Jones Paul, A.S.: Ward, “Routing Strategies for Delay Tolerant Networks”, Submitted to Computer Communication Review (2008) [13] Lindgren, A., Doria, A., Schelen, O.: Probabilistic Routing in intermittently connected networks, vol. 3126, pp. 239–254 (2004) [14] Burgess, J., Gallagher, B., Jensen, D., Levine, B.N.: MaxProp: Routing for VehicleBased Disruption-Tolerant Networks [15] Spyropoulos, T., Psounis, K., Raghavendra, C.S.: Proceedings of the ACM SIGCOMM workshop on Delay-tolerant networking (2005) [16] Spyropoulos, T., Psounis, K., Raghavendra, C.S.: Spray and Focus: Efficient MobilityAssisted Routing for Heterogeneous and Correlated Mobility. In: Proceedings of the Fifth IEEE International Conference on Pervasive Computing and Communications Workshops (2007)
The Performance Comparison between Hybrid and Conventional Beamforming Receivers in a Multipath Channel Rim Haddad and Ridha Bouallegue Laboratory research in telecom systems 6’Tel@ SUP’COM High School of Communication of Tunis Route de Raoued, Km 3,5 2083 Ariana, Tunisia
[email protected],
[email protected] Abstract. The performance in term of Bit Error Rate (BER) of smart antenna receivers calls for some simplification of the interference reduction capability. Obviously, the receiver performances are strictly related on the efficiency of MAI reduction. In order to gain from the enhancements of both: multiuser detection and adaptive antenna , we propose in this paper a hybrid scheme of diversity and smart antennas called Hierarchical Beamforming (HBF), to jointly combat fading and MAI. Our analysis is based on modeling the HBF receiver and the description of the simulation strategy employed to simulate its performance. Moreover, we compare the performance of HBF receiver with Conventional Beamforming (CBF) one. The proposed model conforms the benefits of adaptive antennas in reducing the overall interference level (intercell/intracell) and to find an accurate approximation of the error probability. Keywords: Beamforming, Hierarchical Beamforming (HBF), Conventional Beamforming (CBF), Angle of Arrival (AoA), Rayleigh fading.
1 Introduction Smart antennas and associated technologies are expected to play a significant role in enabling broadband wireless communication systems. The demand for increased capacity in wireless communication networks has motivated recent research activities toward wireless systems that exploit the concept of smart antenna and space selectivity. The deployment of smart antennas at cellular base station installations has gained enormous interest because it has the potential to increase cellular system capacity, extend radio coverage, and improve quality of services [1,2]. Smart antennas may be used to provide significant advantages and improved performance in almost all wireless communication systems. In a typical mobile environment, signals from users arrive at different angles to the base station and hence antenna arrays can be used to an advantage. Each multipath of A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 185–196, 2011. © Springer-Verlag Berlin Heidelberg 2011
186
R. Haddad and R. Bouallegue
a user may arrive at a different angle, and this angle spread can be exploited using an antenna array [3,4]. The Bit Error Rate (BER) is considered to be one of the most important performance measures for communication systems and hence it has been extensively studies. The exact analytical evaluation of the probability of error in DS-CDMA, is still an open subject.Hence in this paper, we will describe the HBF receiver and the simulation strategy to simulate its performance. We propose a novel approach to evaluate the average probability of error by considering an approximation of the spatial filter. Hence, we will derive an analytical model for evaluating the mean BER of HBF and CBF receivers. The analysis is performed assuming Rayleigh fading multipath environments. We assume to make a comparison between two types of smart antenna receivers: the HBF receiver and CBF receiver. An analytical model provides rapid and accurate assessment of the smart antenna system performance under a variety of active users and channel scenarios. We organize the rest of the paper as follows: In section 2, we introduce the system model, followed by the smart antenna receiver model in section 3. The general simulation assumptions and simulation results are provided in section 4 and section 5 respectively. We conclude in section 6.
2 System Model 2.1 System Model of Conventional Beamforming We consider K the total number of active Mobile Stations (MS) in the system, which are randomly distributed in the azimuthal direction, along the arc boundary of the sector cell in the far field of the array. For simplicity, the conventional encoder and interleaver are ignored (this approach is widely used [5] for wireless communication systems employing multiple antennas). In fact, the signals, transmitted by the K users, pass through a multipath channel and are received by the BS array antenna. The location of each MS is identified by its Angle of Arrival (AoA) θ , which is conventionally measured from the array broadside. The BS receiver is equipped with a conventional Maximum Signal to Noise Ratio beamformer followed by an L finger non-coherent RAKE combiner [6]. The resultant signal goes into the in-phase (I) and quadrature (Q) channels simultaneously.The transmitted signal s of the k user can be written as [7]: s (t) = W
( )
()
(t)a (t) cos(ω t) ( )
W
( )
(t
( )
T )a
(t
T ) sin(ω t)
(1)
Where q = 1,2, … , Q, W (t) is a Hadamard-Walsh function of dimension Q which represents the q orthogonal signal of the k user’s long code sequence, a (t) is the () ( ) k user’s long code sequence, a (t) and a (t) are the in-phase and quadrature phase pseudo-noise (PN) sequences, T = T⁄2 is the delay for OQPSK signals. The power of each user is assumed unity (perfect power control). To simplify our study the PN codes are presented as follows:
The Performance Comparison between Hybrid and CBF Receivers
() a (t) =
a ()
( )
() a , (t) p(t
(t) =
a
( ) (t) p(t ,
187
T)
(2)
T)
(3)
( )
Where a , and a , are i.i.d variables taking the values 1 with equal probability and p(t) is the chip pulse shape which is assumed to be rectangular. The equation (1) can be written as follows: s (t) =
W
( )
(t)a( ) (t)
jW
jS
( )
( )
(t
T )a
(t
T) e
(4)
s (t)e
s (t) = () Where s (t) = S (t) transmitted signal.
( )
(t) is the complex low pass equivalent of the
The k user propagates through a multipath channel with (AoA) θ . The complex equivalent representation of the channel impulse response between the l multipath of the k user and the n element of array antenna is presented as follows: h
,,
(t) = β , e h
,,
,
(t) = β , e
(
) ,,
δ t
δ t τ
τ
,
(5)
,
where β , , Φ , and τ , are the path gain, phase and delay respectively, φ , , is the overall phase which includes the path phase and the difference in propagation delays between the antennas. In this case of transmitter we assume that path gains follow the Rayleigh and Ricean distributions respectively. To simplify our work, we assume that multipath channel parameters β , (t) and φ , , (t) remain constant in the duration of Walsh symbol [8], so β , (t) = β , and φ , , (t) = φ , , for t 0, T , where T is the Walsh symbol period. 2.2 System Model of Hierarchical Beamforming We consider a BS serving a single 120° angular sector. It is assumed that the BS is equipped with F co-linear sub-beamforming arrays. The number of array elements in each sub-array is B. That’s why the total number of array elements is = . The inter-element spacing in each sub-array is = /2, while the spacing between the adjacent sub-beamforming arrays ( ), is assumed large enough ( = 20 or more) to uncorrelated fading. The extreme case of = 1 and = corresponds to the conventional Beamforming. As the required spacing between sub-arrays for space diversity is much smaller than the sector radius, this AoA is assumed to be the same at each sub-array [9].
188
R. Haddad and R. Bouallegue
In this section, we consider that the BS is equipped with a hierarchical Beamforming receiver. Each sub array employs the functional block diagram of OQPSK receiver model. 2.2.1 Transmitted Signal We assume that the MS transmitter of each user employs offset Quadrature Phase Shift Keying (OQPSK) M-ary orthogonal modulation. user can be written as [10]: The transmitted signal of the ( )=
( )
( )
( )
( )
( )
)
( ) cos( ) ) sin(
(
)
(
)
( )
(
(6)
( )
where is a Hadamard-Walsh function of dimension which represents the orthogonal signal ( = 1,2, … , = 64) of the user, ( ) and ( ) are the in-phase user long code and quadrature phase pseudo-Noise (PN) sequences, ( ) is the =2 and is the carrier sequence, is the half chip delay for OQPSK signals, frequency. 2.2.2 Channel Model We assume in the following sections that the transmitter signal propagates over Rayleigh fading multipath channel. The complex equivalent representation of the channel impulse response between multipath of the the user and the the antenna in the sub-array is the given as: ( ) ,,
( )=
( )
( ) ,,
( ) ,
(7)
,
( )
Where , is the path amplitude, , , is the overall path phase and , is the path delay respectively. To simplify our work, we assume that multipath channel ( ) ( ) parameters , and , , remain constant in the duration of Walsh symbol.In vector notation, the spatial signature or channel response vector ( ) ,
=
( ) ,,
( ) ,,
…
( ) ,,
2.2.3 The Received Signal At the receiver, the total received signal for the notation as: ( )
( )=
,
( ) , (
( ) , (
) is given by: (8)
sub-array can be written in vector
)
( )
( )
(9)
The Performance Comparison between Hybrid and CBF Receivers
189
where , = + , , is the random delay of the user due to the effect of asynchronous transmission, ( ) is the noise which is assumed to be Additive White ( ) Gaussian Noise (AWGN) and , ( ) the channel response vector given in (2.2.2).
3 Smart Antenna Receiver Model 3.1 The CBF Receiver Model The receiver is divided in four main blocks which can be identified as follows: (1) the array antenna block, (2) the PN despreading, (3) the Beamforming and (4) Walsh correlation and demodulation. We will explain the function of each block: The first step of the receiver is to obtain the quadrature components at each antenna. We multiply the received waveforms by cos(ω t) and sin(ω t) respectively and then Low Pass Filtering (LPF) to remove the double frequency components that results from multiplication [11]. The output of the I-channel and Q-channel low pass filter is given by: ሺ୍ሻ ୩ǡ୪ǡ୬ ሺሻ ൌ ൣ୩ǡ୪ǡ୬ ሺሻ
ሺɘୡ ሻ൧
ɔ୩ǡ୪ǡ୬ ሺ୍ሻ ሺ୯ሻ ൌ ൜Ⱦ୩ǡ୪ ୩ ൫ െ ɒ୩ǡ୪ ൯୩ ൫ െ ɒ୩ǡ୪ ൯ ʹ ሺ୕ሻ ሺ୯ሻ Ⱦ୩ǡ୪ ୩ ൫ െ െ ɒ୩ǡ୪ ൯୩ ൫ െ ɔ୩ǡ୪ǡ୬ ൠ Ʉሺ୍ሻ ሺሻ െ ɒ୩ǡ୪ ሻ ʹ
(10)
ሺ୕ሻ ୩ǡ୪ǡ୬ ሺሻ ൌ ൣ୩ǡ୪ǡ୬ ሺሻ ሺɘୡ ሻ൧ ሺ୯ሻ
ɔ୩ǡ୪ǡ୬ ʹ ሺ୯ሻ ሺ୍ሻ െ Ⱦ୩ǡ୪ ୩ ൫ െ െ ɒ୩ǡ୪൯୩ ൫ ɔ୩ǡ୪ǡ୬ ൠ Ʉሺ୕ሻ ሺሻ െ ɒ୩ǡ୪ ሻ ʹ ሺ୕ሻ
ൌ ൜Ⱦ୩ǡ୪ ୩ ൫ െ ɒ୩ǡ୪ ൯୩ ൫ െ െ ɒ୩ǡ୪൯
ሺ୕ሻ
ሺ୍ሻ
୩ǡ୪ǡ୬ ሺሻ ൌ ୩ǡ୪ǡ୬ ሺሻ ୩ǡ୪ǡ୬ ሺሻ
(11)
(12)
The complex low pass of the received signal can be written as: r
,,
(t) = r ( ,), (t)
jr
( ) ,,
(t)
(13)
After filtering, each path is detected by one of the fingers immediately following the radio-frequency stages. The complex low pass equivalent of the post PN-despread signal is given as yk,l,n(t) : yk,l,nt=y
() ,,
(t)
jy
( ) ,,
(t)
(14)
190
R. Haddad and R. Bouallegue
The despreading sequences are denoted as [12]: a(t) = a( ) t τ We can also write as follows: y
() ,,
(t) =
a(t), r =
y
( ) ,,
(t) =
a(t), r =
,, () r ,,
ja
,
(t)a
()
t
τ
t
T
r
,
τ
( ) ,,
(t)a
r
,
( )
( ) ,,
t
T
(t)a
()
Where (a, b) = a · b the product between complex numbers.y vector notation as: = y
,
t
T
τ
,
.
(t)
, , (t) ( ) () r , , (t)a
Y
( )
,,
,y
,…,y
,,
τ
,
t τ
,
(15)
(16)
can be written in
,,
(17)
,,
In the next step, the signal after PN despreading is combined by the beamformer. In the Beamforming operation, the signals received by antenna elements are weighted by complex weights and then summed up. The smart antenna output is given by: Z
,
= W
,
()
Z , (t) = Z , (t)
Y
,
(18)
jZ
( ) , (t)
(19)
Where W , is the Beamforming weight vector given by: W
,
= W
,,
,W
,,
,…,W
,,
(20)
To simplify our work, we assume that the weights are set as W , = h , and these vector channel coefficients are assumed to be perfectly known. This provides the best case system performance. The last step is the correlation of the smart antenna output with stored replicas of the Walsh functions to form the decision variable for demodulation. The output of the q Walsh correlator (q = 1,2, … , Q) for single antenna is: ()
Z , (q) =
Z
( ) , (q)
=
1 T 1 T
,
()
Z , W(
()
)
t
τ
,
Z , W(
( ) ( ) , W
t
τ
,
Z
)
t
T
τ
dt
,
(21)
,
, ,
Z
( ) ( ) , W
t
T
τ
,
dt (22)
The Performance Comparison between Hybrid and CBF Receivers
The decision variable for the l previous values:
multipath of the k
() u , (q) = Z ,
Z
191
user is obtained from the
( ) ,
(23)
The overall decision variable is obtained by Equal Gain Combining (EGC) of all the decision variables from the L multipaths as [13]: u (q) =
u , (q) =
Z
() ,
Z
Finally, the receiver makes a hard decision on the q the Maximum Likelihood Criteria rule as: q = arg
,…,
( ) ,
symbol of the k
(24) user by using
max u (q)
(25)
3.2 The HBF Receiver Model The HBF receiver is divided in four main blocks which can be identified as follows: (1) the sub-array antenna blocks (2) the PN dispreading, (3) the Beamforming and (4) Walsh correlation and demodulation. The received signal at each sub-array antenna is first down converted. Each resolvable path is then detected by one of the RAKE fingers. To detect the l path, the signal at the different sensors is dispread using the sequence of the respective mobile and synchronized to the delay of the l path. The post PN-despread signal vector is:
Y
() ,
= y
() ,,
y
() ,,
…y
() ,,
(26)
In the next step, the signal after PN dispreading is combined by the Beamforming process. The Beamforming output is given by:
()
z , (t) = W Where W
() ,
() ,
Y
() ,
(27)
is the Maximum SNR Beamforming weight vector given by:
W
() ,
= W
() ,,
W
() ,,
…W
() ,,
(28)
192
R. Haddad and R. Bouallegue
To simplify our work, we assume that the weights are set equal to the channel response vector for the desired user. This provides a lower bound on the system performance. The last step is the correlation of the beamformers with stored replicas of the Walsh functions and then the overall decision variable is obtained by Equal Gain Combining (EGC) of all the decision variables from the multipath signals for the f sub-array. The overall decision is then made by selecting the decision outcomes from the respective sub-beamforming array with the best channel state [9].
4 General Simulation Assumptions The performance of HBF array antenna systems is evaluated by means of Montecarlo simulations runs over the variable of interest ( / or M). The figure of merit used in this work is the mean Bit Error Rate (BER). This is the mean BER taken over the set of channel Rayleigh fading parameters. The performance metric is collected and averaged over = 100drops. A drop is defined as a simulation run for a given number of MS. During a drop, the MS’s AoA increases or decreases linearly with angle change ∆ to crossover the entire sector azimuth range [-60°,60°]. During a drop, the channel undergoes fast fading according to the motion of the MS’s. To simulate the MS mobility, we assume that the snapshot rate is equal to the Walsh symbol rate and the angle change between snapshots is ∆ = 0,01° per snapshot (MS travelling at 300km/h at only 100m from the BS, this value is widely used in simulations). For clarity of investigations, the main parameters for HBF simulation assumptions are discussed below: a) Number of Antenna elements: To make the comparison between HBF and CBF, it is merely assumed that the number of antenna elements M is the same for both cases. b) Number of HBFbranches:We consider in simulations that the BS is equipped with F=2 co-linear sub-beamforming arrays. This choice of sub-arrays is motivated by practical array size considerations and is relevant to a BS serving three sectors, each covering 120° in azimuth. c) Channel:The channel considered is Rayleigh fading with L=1,2 paths/user respectively. d) Pdf in AoA:We assume a Gaussian pdf in AoA. The angular distribution of the waves arriving at the BS in azimuth is described by the pdf in AoA. e) Angle Spread:The values of angle spread used in simulations lie in the range 5°15° which corresponds to urban macrocellular areas.
5 Simulation Results The performance of HBF is determined by the interaction of a number of factors. These include: Beamforming gain via closely spaced antenna elements within each sub-array beamforming, space diversity gain via widely separated sub-arraysbeamforming,
The Performance Comparison between Hybrid and CBF Receivers
193
additional space diversity gain via angle spread and temporal diversity gain via the multipaths. We present in the following sections the impact of each parameter in the performance of HBF and we will make a fair comparison between HBF and CBF.
5.1 Effect of Varying Noise Level First of all, we study the performance of HBF and CBF for the case of a single user (K=1).Obviously, there is no MAI for the case of one user. We can notice from Figure 1 that both CBF and HBF for different number of antennas show a considerable improvement in mean BER compared to the conventional receiver (super imposed as reference). Besides, the improvement in mean BER increases with / .It is very clear from the figure that the performance of HBF is superior to CBF, e.g for a BER / of about 5dB is required for CBF, but threshold of 10-2, M=4 antennas, and only 2.5dB is required for HBF.The performance of HBF is superior to CBF due to space diversity gain offered by the widely separated sub-arrays, which is dominant factor (in the absence of MAI) for the case of a single user.
Fig. 1. Mean BER versus Eb/N0 for K=1 user, L=2 paths Rayleigh fading channel, σAoA=0°
5.2 Effect of Varying Angle Spread We can notice from Figure 2 that, both CBF and HBF improve the performance as the increases from 5° to 10°.It is obvious from the figure, that for low angle spread / , CBF is slight better than HBF. But, as / gets higher, diversity gain becomes dominant and HBF becomes better than CBF.
194
R. Haddad and R. Bouallegue
Fig. 2. Mean BER versus Eb/N0 for K=1 user, L=2 paths, M=6 antennas
5.3 Effect of Varying Number of Antennas It is noticed from the Figure 3, that for = 0°, HBF is better than CBF due to diversity gain provided by array architecture. Moreover, there is no much improvement in performance for both CBF and HBF, by doubling the number of antennas from 4 to 8.If we want to compare angle spread scenarios, for = 5°, HBF is better than CBF, but for larger angle spreads for = 10° and 15°, both array architectures show a similar performance for the number of users considered in simulations.
Fig. 3. Mean BER versus number of antennas M, K=15 users, L=1path/user
5.4 Effect of Varying Number of Multipaths = 0° and 5°, HBF is better than CBF. It can be observed from the figure that for But for = 10° and 15°, we notice that CBF outperforms HBF because the additional diversity gain from spatial fading becomes dominant with the increase of
The Performance Comparison between Hybrid and CBF Receivers
195
Fig. 4. Mean BER versus number of antennas M, K=15 users, L=2paths/user
number of antennas M and in the presence of path diversity, the MAI becomes the dominant factor.
5.5 Effect of Varying the Number of Users Finally, we examine in Figure 5 the performance of both HBF and CBF by varying the number of users K. Also in Figure 8, we re-confirm the trends identified in = 0° and 5°, Figures 1,2 and 3. It can be observed from the Figure 8 that, for HBF yields better mean BER results than CBF. However, for larger angle spread = 10° only for small number of users, HBF outperforms CBF. The behaviour of both schemes becomes different for larger number of users when MAI becomes the dominant factor, and to combat interference it is better to use CBF scheme.
Fig. 5. Mean BER versus number of users K for Eb/N0=10dB, M=6 antennas, L=2paths/user and σAoA=0°,5°,10° respectively
196
R. Haddad and R. Bouallegue
6 Conclusion In this paper, we have reported on the performance of hybrid scheme of diversity and Beamforming. Furthermore, its performance is compared with conventional Beamforming with moderate values of the system parameters such as angle spread number of antennas, number of multipath and number of users. It has be shown that while assuming zero angle spread, the performance of HBF is superior to CBF due to space diversity gain afforded by the well separated sub-arrays.The inclusion of angle spread produces spatial fading across the array, which results in additional diversity gain and improves the performance of both CBF and HBF schemes. For the case of moderate or large angle spread, when path diversity is present and the system is heavily loaded, CBF yields better mean BER results than HBF. All these results are based on the assumption of perfect channel estimation, that’s why the choice of optimum receiver architecture is dependent on the channel conditions.
References 1. Bellofiore, S., et al.: Smart antenna system analysis, integration and performance for Mobile Ad-Hoc Networks (MANET’s). IEEE Trans. AntennasPropagat. 50, 571–581 (2002) 2. Ho, M., Stuber, G., Austin, M.: Performance of switched-beam smart antennas for cellular radio systems. IEEE Trans. Vehic. Technol. 47, 10–19 (1998) 3. Haddad, R., Bouallègue, R.: BER Performance in Space-Time Processing receiver using Adaptive Antennas over Rayleigh Fading Channels. In: Proc. IEEE International Conference on signal Processing and Communication, November 2007, pp. 1483–1486 (2007) 4. Haddad, R., Bouallègue, R.: BER Performance of Smart Antenna Systems Operating over Rayleigh fading Channels. In: Proc. IEEE Wireless Days 2008, November 2008, pp. 1–5 (2008) 5. Bjerke, B.A., Zvonar, Z., Proakis, J.G.: Antenna diversity combining aspects for WCDMA systems in fading multipath channels. IEEE Transactions on Wireless Communications 3(1), 97–106 (2004) 6. Roberts, M.A., Thomas, M.: Introduction to Adaptive Arrays. Sc Tech Publishing (2004) 7. Lee, Y.H., et al.: Performance Analysis of conventional coded DS/CDMA System in Nakagami Fading Channels. In: Telecommunication Systems Journal 8. Jalloul, L.M., Holtzman, J.M.: Performance analysis of DS/CDMA with non-coherent Mary orthogonal modulation in multipath fading channels. IEEE Journal on Selected Areas in Communications 12(5), 862–870 (1994) 9. Bjerke, B.A., Zvonar, Z., Proakis, J.G.: Antenna diversity combining aspects for WCDMA systems in fading multipath channels. IEEE Transactions on Wireless Communications 3(1), 97–106 (2004) 10. Roberts, M.A., Thomas, M.: Introduction to Adaptive Arrays. Sc Tech Publishing (2004) 11. Rappaport, T.S.: Wireless Communications: Principles and Practice, 2nd edn. PrenticeHall, Englewood Cliffs (2002) 12. Stuber, G.L.: Principles of Mobile Communication, 2nd edn. Kluwer Academic Publishers, Dordrecht (2001) 13. Iskander, C.D., Mathiopoulos, P.T.: Performance of multicode DS/CDMA with M-ary orthogonal modulation in multipath fading channels. IEEE Transactions on Wireless Communications 3(1), 209–223 (2004)
A Qualitative Survey on Multicast Routing in Delay Tolerant Networks Sushovan Patra1, Sujoy Saha2, Vijay Shah1, Satadal Sengupta1, Konsam Gojendra Singh1, and Subrata Nandi1 1 Department of Computer Science and Engg, Department of Computer Application National Institute of Technology, Durgapur, 713209, India {bubususpatra,sujoy.ju,vjsah27,satadal.sengupta.nit, konsamsingh,subrata.nandi}@gmail.com 2
Abstract. Delay Tolerant Networks (DTNs) are a class of networks that make communication in stressed and challenging environments possible. DTN is characterized with a number of unique features by virtue of which a working environment is achieved in situations where traditional networking paradigms fail to deliver satisfactorily or entirely. The utility of multicasting in DTNs extends to numerous potential DTN applications i.e., crisis environments, battlefield situations, deep space communications, dynamic data size management, etc. In this paper, we propose taxonomy for the different multicast routing strategies and thereafter, we present a comprehensive up to date survey of these strategies. Further, we perform a qualitative comparison between the different multicast strategies with respect to important performance issues in DTN. We also highlight some unexplored areas in DTN multicasting that could inspire research in the near future.
1 Introduction Personal communication devices like as cellular phones have made voice and data communications possible by achieving global connectivity through infrastructure networks such as cellular and WLAN [1]. Additionally, local connectivity can be achieved through ad-hoc networks since mobile devices are nearly always turned on and possess the necessary attributes to act as routers. The classic TCP/IP-based communications necessarily require end-to-end connectivity. However, sparse ad-hoc networks do not support this due to frequent disruptions and partitions caused due to node mobility. Delay tolerant networks (DTNs) are a class of emerging networks that experience frequent and long-duration partitions. There is no end-to-end path between some or all nodes in a DTN [2]. These networks have a variety of applications in situations that include crisis environments like emergency response and military battle-fields, deep-space communication, vehicular communication, and noninteractive internet access in rural areas. Multicast involves the distribution of specific data to a group of users. While multicasting in the Internet and mobile ad hoc networks has been studied extensively, multicasting in DTN is a considerably different and challenging problem. It not only A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 197–206, 2011. © Springer-Verlag Berlin Heidelberg 2011
198
S. Patra et al.
requires new definitions of multicast routing algorithms but also brings new issues to the design of routing protocols. According to the best of our knowledge, our qualitative survey on multicasting in DTN is the first of its kind, and it includes even the most recently proposed multicast strategies. The remainder of our paper has been structured as follows. Section 2 discusses about the importance of multicasting in DTN and associated challenges. In section 3, we discuss the basis of our classification. Section 4 presents the proposed taxonomy tree by classifying various routing strategies. Section 5 concludes our paper and focuses on future work.
Multicast Routing Strategies in DTN
Multicast Flooding
Unicast Based Routing (UBR)
Group Static Based Tree Routing Based (GBR) Multicast
Tree Based Multicast
Dynamic Encounter Tree Based Based Multicast Multicast Routing (EBMR)
Dynamic Tree Based Routing (DTBR)
Probability Based Multicast
Context Forwarding Aware Group Multicast Based Routing Routing (CAMR) (FGBR)
Intelligent Multicast
SMART A-SMART RelayCast
On-demand Situation aware Multicast (OS Multicast)
Fig. 1. Classification of Multicast routing strategies based on a new taxonomy
2 Challenges and Applications of Multicasting in DTN Unicasting in DTNs has been researched upon to a large extent as opposed to multicasting. As mentioned before, multicast routing in DTN is a relatively fresh topic; however, the massive range of its applications makes its study an important one. Besides, because of its fundamentally different working principle with respect to unicast, multiple challenges are encountered while making headway with this topic. Some of the applications and challenges have been discussed below: 2.1 Identical Challenges between Unicast and Multicast in DTN While implementing multicasting in DTNs, due to large transfer delays, group membership of a particular multicast group may change during a message transfer, introducing ambiguity in multicast semantics. Under these situations, it is necessary to make a distinction between group members and the intended receivers of a message, i.e., endpoints to which the message should be delivered. Group members may change
A Qualitative Survey on Multicast Routing in Delay Tolerant Networks
199
with time as endpoints join and leave the group. The intended receivers, on the other hand, should be fixed for a message, even though they are defined based on group membership. In order to overcome the aforesaid challenges, various multicast routing strategies have been introduced by researchers, which we have tried to classify taking into consideration their working principles. 2.2 Applications of DTN Multicasting Multicast service supports the distribution of data to a group of users. Many potential DTN applications operate in a group-based manner and require efficient network support for group communication. For example, in a disaster recovery scene, it is vital to disseminate information about victims and potential hazards among rescue workers. In a battlefield, soldiers in a squad need to inform each other about their surrounding environment. Although group communication can be implemented by sending a separate unicast packet to each user, this approach suffers from poor performance. The situation is especially acute in DTNs where resources such as connectivity among nodes, available bandwidth and storage are generally severely limited. Thus efficient multicast services are necessary for supporting these applications.
3 Classification of Multicast Routing Strategies We have attempted to classify the proposed multicast routing strategies in DTNs on the basis of their basic working mechanisms. Multicasting can be implemented in DTNs in a variety of ways (as in Fig. 1). We classify them as follows: 1) Messages are flooded throughout the network, 2) Messages are forwarded along a multicast tree that stores node information leading to the destinations, 3) A probabilistic approach is used which employs history of encounters to select the best route, 4) An intelligent combination of flooding and forwarding techniques is used to make better use of available resources. Each of these techniques can be further classified into more specific ones (Fig. 1): multicast flooding can be achieved by using unicast transfer [3] [4], or by the broadcast strategy; tree-based multicast can be accomplished by using a static tree or dynamic tree to decide the shortest path to a destination; probabilitybased multicast can be implemented using the encounter-based technique that records history of node encounters to decide best route or by using the context-aware multicast-routing (CAMR) [11] which allows for excess power usage in extremely sparse networks; intelligent multicast can be achieved by segregating the entire message delivery process into two segments, each implementing either the flooding or the forwarding technique to achieve better performance, as in case of the forwarding group-based, SMART, A-SMART [12], and RelayCast [13] routing strategies. We discuss the above mentioned strategies in detail in the section that follows.
4 Proposed Taxonomy Based on the above mentioned bases of classification, we propose taxonomy for the various multicast routing strategies (as shown in Fig. 1). Each routing strategy has
200
S. Patra et al.
been discussed in brief and probable conclusions have been drawn on their performances. 4.1 Multicast Flooding Multiple copies of the messages are flooded into the network so that the message gets transferred to the intended receivers of the multicast group. The techniques that fall under this category are as follows. 4.1.1 Unicast-Based Routing (UBR) This can be considered to be the simplest way of implementing multicast in DTN. Here, the source sends multicast bundles to the destination through multiple unicast operations [6], [7]. Any existing DTN unicast scheme can implement this strategy by modifying its bundle header to include group information. Unicast routing schemes like Epidemic Routing [18] and Spray-and-Wait algorithm [15] already implement this strategy to achieve multicasting. Apparently, this strategy accomplishes least implementation overheads [7]; however, as number of receiver nodes in a multicast group increases, there is a chance that an intermediate node will forward the same bundle several times, thus decreasing delivery efficiency dramatically. 4.1.2 Broadcast-Based Routing (BBR) BBR [14] or Epidemic Routing [18] uses the technique of flooding in disruptiontolerant networks. In this routing scheme, flooding of messages throughout the network is carried out with the intention of reaching intended receivers [14]. BBR performs better when it has access to long-term information about the network topology, i.e., average interval between node contacts, etc. BBR generates redundant messages, a property which renders it inefficient in mobile networks where power supply for individual nodes is limited. It is probably safe to say that flooding based routing should work better in Random Walk/Waypoint models since node movement predictability is negligible. Delivery ratio must be very high with significantly low latency, although buffer overhead will be quite large. 4.2 Tree Based Multicast In tree-based multicast routing, a DTN graph is considered which consists of all the nodes present in the network [14]. The messages are forwarded along a tree in this DTN graph that has the source as its root and is connected to all the receivers in the network. The message passing technique is essentially forwarding, as messages are duplicated at a node of the tree if and only if it has more than one outgoing paths [4], [6], [7]. Tree-based multicast can be categorized into the following two strategies: 4.2.1 Static Tree Based Multicast As discussed earlier, a multicast tree is created at the start of a multicast session, with its root at the source [7]. The source first gathers information about discovered routes to all the intended receivers and then constructs a smallest cost tree using Djikstra’s algorithm based on this information [14]. As we can understand from the name, the
A Qualitative Survey on Multicast Routing in Delay Tolerant Networks
201
topology of the intermediate nodes of this tree does not change until the multicast session is complete. Bundles are replicated according to the number of downstream neighbours, i.e., number of messages coming out of a node equals the number of its downstream neighbours. Its demerit comprises of the fact that it loses flexibility of adjusting multicast routing decision according to variations in the topology during the course of a particular multicast session. This strategy is most appropriate where disruptions happen periodically in a scheduled pattern, e.g., data communication via LEO satellite. We can intuitively conclude that this strategy is supposed to work best in the Working Day mobility model where the node mobility follows a periodic pattern. 4.2.2 Dynamic Tree Based Multicast Contrary to the static tree, dynamic tree based multicast allows for dynamic adjustment of the multicast tree to incorporate changes in the network topology during the course of a particular multicast session. In this strategy, each bundle has an associated tree [7] that may change hop-by-hop depending upon up/down variations of DTN links. Each node having a bundle performs the three common steps: collection of information regarding availability of DTN links, computation of smallest cost tree and forwarding bundles using discovered multicast tree [14]. In addition, this strategy can take advantage of newly available routes to receiver nodes and can avoid forwarding messages through links that are now disconnected due to outward movement of nodes. Though this strategy is characterized with high overheads, it is better adaptive to topology variations in DTNs. We discuss two variations of this strategy in the text that follows. 4.2.2.1 Dynamic Tree Based Routing (DTBR). Each DTN node has knowledge oracle containing schedule or statistical summary of link up/down information in DTN overlay and thus the source computes a multicast tree for each bundle and forwards the current message along the tree [7]. Based on this, source computes a multicast tree for each bundle and forwards current message along tree. Thus, once a message leaves the source for a destination node, the strategy remains static virtually since it does not incorporate the changes in the topology thereafter. This will fail to work efficiently in networks where disruptions are random and frequent. 4.2.2.2 On-Demand Situation-Aware Multicast (OS-Multicast). It also builds up a dynamic multicast tree hop-by-hop for each copy of bundle [9]. However, contrary to DTBR, it doesn't rely on any global knowledge of network such as node position, or link up/down schedule. It assumes that underlying networks is able to record discovered routing information and report current availability of outgoing links to DTN multicast agent. It contains full list of intended receivers and thus each intermediate node that has a bundle is responsible for delivering multicast message to all receivers. This improves on DTBR since the intermittent topology changes are evaluated dynamically, thus optimizing performance. However, delivery latency is quite high.
202
S. Patra et al.
4.3 Probability Based Multicast Here nodes deliver messages to the other nodes only when its delivery predictability is higher than the certain threshold value. 4.3.1 Encounter Based Multicast Routing (EBMR) It is a scheme that is purely based on node encounters. EBMR scheme is built on top of PRoPHET Scheme [8]. Each node doesn’t pass bundle to a next hop node unless the next hop node has delivery predictability higher than a certain delivery threshold (Pthresh) value [10]. For multicast delivery each node will pick as many nodes as needed with highest delivery predictability to each of the multicast receivers. 4.3.2 Context Aware Multicast Routing (CAMR) Nodes are allowed to use high power transmission when locally observed node density drops below a certain threshold. Each node maintains 2-hop neighbourhood information and hence can deliver traffic without invoking a route discovery process if all receivers are within its 2-hop neighbourhood [10] [11]. Its advantage constitutes of the fact that it can achieve higher multicast delivery ratio than DTBR and OSmulticast. However it still relies on route discovery process and ability to control node movement. CAMR can be considered a special case of multicast routing where power resources can be exploited to achieve high delivery ratio in very sparse networks. 4.4 Intelligent Multicast Here dynamic intelligence is used by the algorithm to decide between flooding and forwarding techniques of delivering messages to the receivers. This strategy is based on a two-phase algorithm with each phase implementing flooding or forwarding to achieve optimal performance. Flooding technique is implemented to achieve high delivery ratio and low latency since all the intermediate nodes receive single or multiple copies of the message thus increasing the chances of message delivery to an intended receiver. Forwarding, on the other hand, achieves better efficiency and works with a significantly reduced buffer space since message replication is not allowed beyond the number of intended receivers. Intelligent multicast is able to take advantage of the merits of both these techniques. 4.4.1 Forwarding Group Based Routing (FGBR) FGBR implements the concept of a forwarding group [4] within which the message is flooded. The forwarding group is created by computing a shortest path tree (as in case of tree based multicast) to the intended receivers. The group consists of those nodes which are present in the shortest path tree, including the receiver nodes. Within this forwarding group, the message is flooded, thus decreasing latency and increasing delivery ratio. Performance of this strategy is better than in cases where only flooding is implemented. 4.4.2 SMART SMART uses travel companions of the destinations to increase the delivery opportunities. Here, routing is divided into two phases: 1) a fixed number of copies of
A Qualitative Survey on Multicast Routing in Delay Tolerant Networks
203
the message are injected into the network to propagate the message to the companions of the destination by Binary Spray algorithm [15], and 2) a companion of the destination only transmits the message to other companions of the destination until the message is delivered to the destination. 4.4.3 A-SMART In A-SMART, companion nodes are organized to form an anycast group [12] and periodically broadcast its group identifiers and hops to build the routing table. Routing is divided in two phases: 1) an anycast scheme is used to forward the message the companion node of the destination; 2) the companion node only transmits the message to other companions of the destination until the message is delivered to it. Anycast is virtually a unicast, for the reason that source node just sends the message to any one member of a destination group which is the best receiver according to the current environment. In addition, the message will be routed to an alternative member of the destination group when the previous path to a member disappeared, so anycast is a more reliable routing mechanism. 4.4.4 RelayCast RelayCast [13] is a routing scheme which extends 2-hop relay algorithm used in unicast to multicast in DTNs. In this strategy, a source forwards a single message to all the relay nodes, each of which in turn transmits the message to all intended multicast receivers. Mathematical analysis shows that the throughput achieved is better than in case of conventional multi-hop relay. Thus, RelayCast is able to achieve maximum throughput bound of DTN multicast routing. FGBR and A-SMART seem to perform well in most mobility models due to an efficient balance between flooding and forwarding techniques. Due to the partial flooding, delivery ratio and latency are taken care of, whereas buffer usage is somewhat controlled by the partial forwarding character. Recent reports show that use of multicast tree results in poor scaling behaviour which is efficiently dealt with using RelayCast algorithm. Table 1. Performance Comparison among Multicast Routing Strategies based on Performance Metrics
Buffer Usage
UBR
Routing Algorithms
Delivery Latency
Low
Highest
Lower than BBR
BBR
Multicast Flooding Based
Routing Strategies
Performance Metrics
Delivery Ratio
High
Low
Highest
Remarks
Higher delivery ratio is achieved at the cost of high buffer overhead and low efficiency. Should work well in Random Walk/Waypoint mobility models.
204
S. Patra et al. Table 1. (continued)
Routing Strategies
Routing Algorithms Static Tree Based
Higher than UBR
DTBR
Higher than UBR
Less usage
Higher than DTBR
Less usage
OSMulticast
Less than GBR
Higher than DTBR when network is sparse
EBMR
Less usage
Very high when node mobility is predictable
CAMR
Less than GBR
Highest, 8 times more than DTBR or OSMulticast
Low, almost identical to DTBR and OSMulticast
Medium
FGBR
Buffer Usage
High
More than Tree Based
Medium
SMART
Delivery Latency
Higher than ASMART
Low
Slightly lower compared to multicast flooding techniques
A-SMART
Delivery Ratio
High
Higher than SMART
Lower than SMART
RelayCast
Intelligent Multicast
Probability Based
Multicast Forwarding Based
Performance Metrics
Higher than EBMR
High
High
Medium
Remarks
Buffer usage reduced significantly; however, delivery ratio and latency are compromised with. Compatible with most mobility models.
Ideal for networks where node mobility is periodic and/or predictable. CAMR compromises heavily with power usage. Should work best with Working Day mobility model.
Highly efficient; uses intelligent combination of flooding & forwarding techniques to achieve optimal performance; Designed to work well with most mobility models.
Comparable to A-SMART
5 Conclusion and Future Work Multicasting in DTNs is a fresh area of research and there is a limited amount of research information on it. The information, however, is growing in volume as researchers realize the importance of multicast routing in challenging environments.
A Qualitative Survey on Multicast Routing in Delay Tolerant Networks
205
In this paper, we have tried to identify the reasons of considering multicasting an essential tool for routing in disruption-tolerant networks. We have presented a classification comprising the multicast routing schemes that have been proposed and have performed a comparative survey on their performances. The advantages and otherwise of each of the strategies have been studied with an eye for novelty. Though research in the field of DTN multicasting has made some headway in the recent past, there are many important areas that remain unexplored. We highlight some of those areas that can prove to be fodder for future research work. Firstly, security in DTNs is an area of huge concern, especially in those cases where the networking deals with personal information (such as in social networking) or classified information (such as in the battle-field scenario). Major practical contributions regarding security are yet to come up. Secondly, efficient usage of power is another aspect that needs to be considered. More power usage will lead to higher cost, which is both impractical and unsustainable. Another area of significant importance could be dynamic buffer management in DTNs. Data packets can range in size from a few KBs (such as text files) to some GBs (such as multimedia files). There is a need to provide for dynamic addition and reduction of buffer space in nodes depending upon the size of the data packet at being transmitted at a particular instant of time. This provision could contribute significantly in the reduction of buffer usage and thus make routing in DTN more sustainable. Last but certainly not the least; we should focus on the issue of scalability in DTN environment, i.e., sustainability of a particular routing strategy with increasing node density. The practicality of a strategy will depend hugely on its scalability.
References 1. Keränen, A., Ott, J., Kärkkäinen, T.: The ONE Simulator for DTN Protocol Evaluation. In: Simutools 2009 Proceedings of the 2nd International Conference on Simulation Tools and Techniques, Belgium (2009) 2. Fall, K.: A Delay Tolerant Network Architecture for Challenged Internets. In: Proc. ACM SIGCOMM, pp. 27–34 (2003) 3. Jones, E.P.C., Ward, P.A.S.: Routing Strategies for Delay Tolerant Networks. In: Proc. WDTN 2005 Proceedings of the 2005 ACM SIGCOMM workshop on Delay-tolerant networking, ACM, New York, USA (2005) 4. Santiago, J., Casaca, A., Pereira, P.R.: Multicast in Delay Tolerant Networks using Probabilities and Mobility Information. Ad-hoc and Sensor Wireless Networks, An International Journal 7(1-2), 51–68 (2009) 5. Narmawala, Z., Srivastava, S.: MIDTONE: Multicast in Delay Tolerant Networks. In: Proceedings of Fourth International Conference on Communications and Networking in China (CHINACOM 2009), Xi’an, China, pp. 1–8 (August 26-28, 2009) 6. Santiago, J., Casaca, A., Pereira, P.R.: Non-Custodial Multicast over the DTN-Prophet Protocol. In: International Federation for Information Processing Digital Library, Wireless Sensor and Actor Network 7. Ye, Q., Cheng, L., Chuah, M., Davison, B.D.: Performance comparison of different multicast routing strategies in disruption tolerant networks. Journal Computer Communication 32(16), 1731–1741 (2009)
206
S. Patra et al.
8. Xi, Y., Chuah, M.: Performance Evaluation of An Encounter-Based Multicast Scheme for DTN. In: 5th IEEE International Conference on Mobile Ad-hoc and Sensor System, pp. 353–358 (2008) 9. Bae, S.H., Lee, S.J., Su, W., Gerla, M.: The design, implementation, and performance evaluation of the on-demand multicast routing protocol in multi-hop wireless networks. IEEE Network 14, 70–77 (2000) 10. Chuah, M., Xi, Y.: An Encounter-Based Multicast Scheme for Disruption Tolerant Networks. Journal Computer Communications 32(16) (April 1955); ButterwothHelnemann, Newton, MA, USA (October 2009) 11. Yang, P., Chuah, M.: Context-Aware Multicast Routing Scheme for DTNs. In: Proc. Of ACM Workshop on PE-WASUN (August 2006) 12. Wu, J., Wang, N.: A-SMART: An Advanced Controlled-Flooding Routing with Group Structures for Delay Tolerant Networks. In: Second International Conference on Networks Security, Wireless Communications and Trusted Computing (2010) 13. Lee, U., Oh, S.Y., Lee, K.W., Gerla, M.: RelayCast: Scalable Multicast Routing in Delay Tolerant Networks. In: IEEE International Conference on Network Protocols (ICNP 2008), Orlando, FL (October 2008) 14. Zhao, W., Ammar, M., Zegura, E.: Multicasting in Delay Tolerant Networks: Semantic Models and Routing Algorithms. In: WDTN 2005 proceedings of 2005 ACM SIGCOMM workshop on Delay-tolerant Networking, USA. ACM Press, New York (2005) 15. Spyropoulos, T., Psounis, K., Raghavendra, C.S.: Spray and Wait: An Efficient Routing Scheme for Intermittently Connected Mobile Networks. In: WDTN 2005 Proceedings of the 2005 ACM SIGCOMM workshop on Delay-tolerant networking, USA, ACM Press, New York (2005) 16. Cerf, V., et al.: Delay Tolerant Network Architecture, IETF, RFC 4838 (April 2007) 17. Scott, K., Burleigh, S.: Bundle Protocol Specification, IETF, RFC 5050 (November 2007) 18. Vahdat, A., Becker, D.: Epidemic Routing for partially-connected ad hoc networks. Duke Technical Report CS-2000-06 (July 2000)
Integrating RFID Technology in Internet Applications Simon Fong Faculty of Science and Technology, University of Macau, Macau
[email protected] Abstract. Radio Frequency Identification (RFID) which is a mature identification and tracking technology recently is picking up its momentum with the emergency of a highly hyped "Internet of Things" (IOT). The ultimate goal of IOT is to let human and objects seamlessly talk to each other via the Internet. It finds great impacts on our lifestyles in various applications such as supply chain management, access control security, mobile health-care, etc. In response to this technology trend and market demands, a high level infrastructure called EPC (Electronic Product Code) network with Object Naming Service has been proposed that rides on the future IPv6 Internet. In order to enable the operation of EPC network, auxiliary mechanisms such as privacy and security protection, context-aware capabilities, middleware and interoperable data semantics are needed. In this paper we propose a prototype of wireless communication system incorporated with Internet capability, namely Internet RFID. In particular, this paper shows the preliminary design, implementation and testing of the Internet RFID prototype in a client-server environment. As an experiment, we implemented a typical RFID system with the additional feature of distance finding that can track and monitor access control, and at the same time a database located remotely in a Web Server is updated. Users can query the Web database server in real-time about the location of the client. The hardware module consists mainly of the construction of two micro-controller systems: one for the user card and the other one for the station transceiver controller. Recommendations are provided at the end of the paper for future development. Keywords: Wireless communication, Internet applications, RFID.
1 Introduction Internet has evolved in the past decades from “Internet of Information” that hyperlinked information over disparate websites in the 80’s, “Internet of Services” with the bloom of eCommerce in the 90’s, “Internet of People” as forms of social network and collaborative forums over the Millennium, to “Internet of Things” (IOT) [1] in year 2010 and beyond. IOT is tightly coupled with ubiquitous computing in which Radio Frequency Identification (RFID) technology plays a central part in auto-identifying and tracking of not just humans, but things like artifacts that can be tagged. The power of IOT as we can see is founded from the tracking and communication among humans-to-humans, humansto-objects and objects-to-objects [2], etc. A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 207–216, 2011. © Springer-Verlag Berlin Heidelberg 2011
208
S. Fong
The underlying support communication network is essentially important, for those “things” talk to one another. A lot of research works have emphasized on the RFID networks such as Mobile RFID network [3] and Wireless RFID network [4]. Mobile RFID network refers to using a mobile phone as a card reader; the information collected from a stationary tag in a fixed location (of course it can be moved) to a mobile phone that is equipped with a built-in RFID reader is sent to the network for tracking. The supporting network would be a mix of telecommunication GSM network and others. Examples are users who use a mobile phone to obtain detailed information from a product (a suit) in a departmental store. The price, origin of manufacturing, materials info, availability etc. would appear on the mobile phone screen. The other type of support network would be a set of local client-server radio network systems interconnected via Internet, generally known as just Wireless RFID network or Internet RFID in short. In this case the tags are referred to the identification objects that the users are wearing or holding. The tag holders (clients) move and their location information are being tracked by a wireless RF controller that usually connects to a computer (server). One example is micro-chips injected in the bodies of animals, so their particulars can be identified and their whereabouts can be traced in real-time. For human users this usually would assume a secure environment installed with sensors where the identification of the personnel and their access patterns matter. Induction from their movement data reveal whether they belong to normal or suspicious patterns. There are ample business opportunities for this kind of Internet RFID despite of applications of security and access controls. For instances, cargos and postal articles are being traced for logistics optimization; locations of supermarket products and trolleys are monitored for shoplifting prevention and for revealing shoppers behaviors; visitors who are wearing RFID badges are located in real-time for proactively providing them information of the exhibits near them or recommending to them the next show-time at their ear-phones. On a larger scale, Internet RFID applications can transfer the location information of users across the Internet, even cross-countries, without boundary. This is the focus of this paper, about a wireless client-server local platform for detecting the presences of tags/users, and the information can connect to another wireless client-server system via Internet. The primary objective of this project is to integrate hardware and software (to build a wireless system that can communicate through the Web) together, so that the system can operate through the Internet. The hardware consists of interfacing the RF modules with the microcontroller, allowing the microcontroller to control the operation of the RF modules. As could be observed from Figure 1, the whole architecture concept can be segregated clearly into global and local. The global architecture concerns about the messaging across Internet, such as checking EPC codes with the registry databases and the database maintenance, interoperable message semantics possibly using XML, messages privacy and security protection, etc., which are beyond the scope of this paper. Our focus is to propose a wireless client-server platform that acts as a part of RFID application across Internet (as indicated in the dotted rectangle in Figure 1.). We assume the messages that sent/received across Internet would be encrypted and the RFID EPC/TRE meta-data supports and middleware functions are already in place. The local Internet RFID system mainly then contains three main portions, namely:
Integrating RFID Technology in Internet Applications
209
Fig. 1. Architectural view of the Internet RFID (Source: Courtesy of csie.ndhu.edu.tw) • • •
The Microcontroller System Design: This comprises the microcontroller (Intel 8051) and the RF modules. The Client Design: This is made up of the user interface screen and it is where the user can communicate with the web server and the microcontroller. The Web Server Design: This is where the customer database is stored and the various CGI files are stored.
The three components work in synchronization so that the whole Internet RFID system operates smoothly. The main challenge is the integration of each component. The operation of the Internet RFID system is described as follows. Let us consider a scenario where the client computer and the web server are situated far away from each other and they communicate through Internet protocol. For the client, one side is connected to the Web server via Internet connection (ADSL/Broadband, e.g.). On the other side of the client, the client is connected to a microcontroller system based setup through an interface I/O port (e.g. USB or RS-232 link). The microcontroller system is then used to control the function of the transceiver. The preprogrammed user ID from the RFID card is received remotely at the receiver when the user was detected within the range of the transceiver station. Upon receiving the ID the microcontroller will send the information to the client computer through the I/O port. Then the client computer will send a query to the web server, providing it with the obtained user ID. This will invoke the CGI program running the web server to perform a search on the database. If the visitor’s particular is found in the database, the CGI program will return a positive result to the client computer for display on the screen. Otherwise, an invalid entry is flagged on the screen.
2 Hardware Design Each of the client station will be equipped with a receiver-transmitter pair (or station transceiver controller) to receive the user ID of the bypassing user. The user card constantly emits the user ID wirelessly and the microcontroller in the station
210
S. Fong
transceiver picks up this information, and transfers it to the client computer via the I/O interface. The working range of the transceiver pair is in the popular 902-928MHz band. With the inclusion of an antenna, the transceiver pair is capable of transmitting and receiving digital information over distances in excess of quarter mile. Intel 8051 series of controller is used for the controlling portion, because of its simple design and interfacing. The design at the user card consists of the 8051 controller and interfaces with the RF module. The schematic diagram of the user card is shown Figure 2.
Fig. 2. The schematic design of the user card
Fig. 3. The schematic design of the Station Transceiver Controller
The design of the station transceiver controller is similar to the user card but have some extra circuitry. Since 8051 uses multiplexed address and data bus concept, a 74HC373(U2) latch is used to facilitate the proper functioning of the whole circuit. At the first part of the execution cycle, the address is released by the microcontroller, and this is indicated by the activation of the Address Latch Enable (ALE) signal. The ALE signal will cause the 74HC373 to latch in the valid address and hold it stable for the rest of the execution cycles. Meanwhile the microcontroller will output the data to be sent on the second part of the current execution cycle. The outstanding part of this logic is to enable the interface of the microcontroller with the outside world and at the same time maintain a minimum pins package so as to reduce production and design cost. The purpose of the 74HC138 (U3: 3-to-8 decoder) is to enable the controller to generate a suitable chip select
Integrating RFID Technology in Internet Applications
211
signal so that at only time, only one device can output information onto the data bus. The schematic design of the station transceiver controller is shown in Figure 3.
3 Software Design The controlling portion is driven by the assembly language program that resided in the EPROM of the 8751 microcontroller. It is divided into two parts, namely the user card and the station transceiver controller. Hence, the user card program will act as a standalone program, controlling the behavior of the user card. For the station transceiver module, the program will continue to receive and transmit the user ID through the receiver and transmitter repeatedly. Apart from that, it will also be in charge of the sending of received user ID to the client PC, and at the same time receiving the new programmed ID from the PC. Thus, it has to handle a two-sided communication with the card and the client PC simultaneously. The logics of the programs for the user card and the station transceiver controller are shown below. Start
Start
Program Initialization
Program Initializtion
On Transmitter Off Receiver
On Receiver Off Transmitter
Transmit "***" and User ID
"**"?
Connect
No
Programming Mode?
Yes
Yes
Receive User ID
Send User ID to Screen
No
Connect
Off Transmitter On Receiver
No
Program Mode?
Yes
Off Reveiver On Transmitter Receive and program new user ID No
Transmit "***" and new User ID
Last byte of data?
Yes
No No
Normal Mode? Normal Mode?
Yes
Yes
On Transmitter Off Reciever
On Receiver Off Transmitter
End
End
Fig. 4. General flow of the user card algorithm (Left). General flow of the station transceiver controller algorithm. (Right)
212
S. Fong
The primary functions of the client software are listed as follows: • • • • • • •
To provide an easy-to-use user interface To enable the communication of the client computer with the microcontroller of the receiver module via I/O port. To enable user to query about user's particular. To allow for the programming of new card. To allow location setting. To allow searching of the database. To establish connection with the Web server.
The client side software program is written in MS Visual Basic that allows communication through the COM port and at the same time provides Internet control protocol for the communication with the Web server. The following shows the highly
Fig. 5. State transition diagram of the client software
Integrating RFID Technology in Internet Applications
213
simplified state transition diagram of the main program that runs on the PC client. The logic of the programs for the user card and the station transceiver controller are shown. On the Web server, there are seven CGI scripts that do the server processing. The CGI programs are written in C for efficient performance, and low-level control over the system. The database is located in the Web Server computer as well. When a request or query is sent to the Web Server from the client, the program to be invoked will be specified in the request. The following CGI programs stored in the directory cgi-bin handle the request accordingly. Table 1. List of CGI scripts for Web server processing enquire.cgi inupdate.cgi outupdate.cgi program.cgi programinfo.cgi search.cgi searchagain.cgi
for handling enquiry request from the clients. for updating the particular record upon user entry. for updating the particular record upon user exit. for updating the new card ID into the particular user record. return user information to the requesting client. for searching the database for the requested string in a first level search. for searching the database for the requested string in a 2nd level search.
4 Accuracy of Transmission From the RF module, there are 8 binary selectable reception frequencies available for operation. This means that at any one time, up to 8 users using different frequencies can access the system simultaneously. For example, we let the sensing distance be 200 meters, and a cardholder be travelling at a speed of 0.5 meter/sec. 200 meters will take 400 sec to cover, assuming that the round trip response time through the Internet is negligible. Since the server takes about 400msec to process a user ID, within a period of 400sec there can be 1000 transmissions for each channel. If all the 8 channels are available, the whole system is able to handle 8000 transmissions for a period of 400 seconds theoretically. An experiment is conducted to test the accuracy of transmission, and the result is plotted in Figure 7. The x-axis represents the number of time in transmitting an arbitrary string by the client repeatedly, and the y-axis is the number of erroneous reception in percentage. By sending some data repeatedly, we simulate the workload generated for the Web server. The errors reported are the reception errors that may be due to the external interference from the environment and
Fig. 6. RF Modules reception accuracy
214
S. Fong
Signals
RF Receiver with RSSI
Analog RSSI Signal
Analog -To- Digital Converter
8751 μC Data Bus and Control Signals
RS-232C Interface To Client
Data Bus and Control Signals
MA X 233
UART NS16550
MAX 233
Fig. 7. Block diagram of the new circuit for the station transceiver controller
the deterioration of the battery power during the testing period. The distance of the testing device remains fixed at a distance of 50 meters and the voltage level of the system are constantly check to ensure sufficient power. The results show that the error is kept less than 10% in most cases.
5 Distance Finding and Wireless Audio Card Paging One feature of the Internet RFID system is the ability to track object without line-ofsight restriction. With the help of Internet, the exact locations of the users can be tracked. It serves as a cheaper alternative to globe positioning system using satellites. In order to implement this distance finding facility, a number of amendments to the existing system are needed: 1) Additional of ADC circuit to the station transceiver controller, 2) Calibration of the RF receiver modules for measurement of signal strength, and 3) Software amendment to the station transceiver controller and the client. The additional circuitry is shown in the dotted area in the above figure. The ADC (e.g. National Semiconductor ADC0801) circuit is included so that the analog RSSI (Receive Signal Strength Indicator) signal can be converted to its digital equivalent and be read by the microcontroller. The RSSI signal will fluctuate with the strength of the receive signal and this will in turn determine how far the transmitter is away from the receiver. Upon the reception of the digital information, the microcontroller will compare it with pre-stored value in a lookup table and compute the actual distance. In order to accurately measure the incoming signal strength, the RSSI circuit in the RF module must be calibrated. Using a signal generator, inject a –40dBm unmodulated carrier into the receiver and measure the RSSI voltage. Repeat the same for an input signal level of –100dBm. This is known as a two-point “slope” calibration. To determine the slope of the RSSI response, use the following equation: M=60 / (V2-V1)
where
M = slope in dB/volts V1 = the voltage measured at –100dBm V2 = the voltage measured at – 40dBm
Integrating RFID Technology in Internet Applications
215
Using this slope, the signal strength of any RSSI voltage can be determined: SS = M * (V1- Vm) + 100dBm where
SS = signal strength in dBm M = previously determined slope Vm = measured RSSI voltage
The obtained SS value is then proportional to a certain distance. Hence, a lookup table is formulated to store a set of SS values versus distance, so that every time a SS value is needed, the lookup table can be checked to obtain the equivalent distance. The major amendment is on the assembly program running in the microcontrollers. The station transceiver controller needs to control the ADC to convert the analog signal into digital format, and at the same time process the information before passing it to the client software. From the Figure 8, when a client terminal is trying to locate a particular user holder, it can do so by locating him/her within its radius of operation. The efficiency of the distance measurement operation will largely depend on the range of the RF modules used. As it can be observed, the limitation of a single client environment is that when locating a user, the system can only tell how far the user is away from the client. The direction of the user cannot be determined, as the area of sensitivity (or coverage) of the RF modules is circular. A clear advantage of such system is the simple implementation and cost saving. The main disadvantage is the inability to exactly pinpoint the location of a user. This problem however can be solved by making use of at least two clients (multi client environment) to determine the location of a user. See Figure 9. The clients must reside in the same environment, and the range of coverage must overlap each other. When a request is made to locate a particular user, the clients in the environment may
Circle of a Particular Signal Strength
Fig. 8. General flow of distance finding Fig. 9. Distance finding using multiple clients algorithm for the microcontroller
216
S. Fong
begin to sense for the RSSI of the user. The interception of radio waves not only tells the distance but also the direction of the user. The principle underlying this technique is more difficult and it requires all the clients in the same environment to work in synchronization in order to locate a user. Another mod will be Wireless Audio Card Paging: With the advancement in microelectronics integration, manufacturers are able to produce components that are smaller in size but better function, hence equipment are also made smaller. The HP series modules are capable of transmitting or receiving a wide range of analog signals with minimal distortion. The transmitter is equally adept at transmitting complex waveform such as voice. Analog signals ranging from 50Hz to 25KHz may be applied to the data input pin. One can use this facility to send voice signal to a particular user wirelessly. This feature allows the paging of a cardholder within the reach of an Internet without the use of public announcement system like confidential paging. The figure below demonstrates a setup to transmit voice. Channel Selection Signal
μC 8751
Transmitter
Digital Data RS-232 Interface
Data Output
Multiplexer
Analog Audio Signal
Fig. 10. Setup of transmitter to transmit analog and digital data
6 Conclusion With the arrival of The Internet of Things (IOT), billions of wireless identifiable “objects” could communicate and interact with one another as an ecosystem. RFID is an enabling technology for IOT that allows easy gathering identification information physical objects from a distance. The communication platform will be an EPC Network whose design has already been proposed. In order to support this network, RFID system needs to be integrated into Internet. This paper proposed a relatively simple solution in hardware and software for integrating RFID into Internet.
References 1. Harrison, M.: The ’Internet of Things’ and Commerce. XRDS: Crossroads, The ACM Magazine for Students 17(3), 19–22 (2011) 2. Fong, S., Zhuang, Y.: A Security Model for Detecting Suspicious Patterns in Physical Environment. In: IEEE The Third International Symposium on Information Assurance and Security (IAS 2007), UK, pp. 221–226. IEEE Press, Manchester (2007) 3. Michael, M.P.: Architectural Solutions for Mobile RFID Services on Internet of Things, Master Thesis, University of Helsinki, pp. 1–95 (2007) 4. Liu, H., Bolic, M., Nayak, A., Stojmenovic, I.: Taxonomy and Challenges of the Integration of RFID and Wireless Sensor Networks. In: IEEE Network, pp. 26–32. IEEE Press, Los Alamitos (2008)
BPSO Algorithms for Knapsack Problem Amira Gherboudj and Salim Chikhi Computer Science Department, MISC Laboratory, Mentouri University, Constantine Algeria {gherboudj,chikhi}@ umc edu.dz
Abstract. Particle Swarm Optimization (PSO) is an evolutionary metaheuristic. It was created in 1995 by Kennedy and Eberhart for solving optimization problems. However, several alternatives to the original PSO algorithm have been proposed in the literature to improve its performance for solving continuous or discrete problems. We propose in this paper 4 classes of binary PSO algorithms (BPSO) for solving the NP-hard knapsack problem. In the proposed algorithms, the velocities and positions of particles are updated according to different equations. To verify the performance of the proposed algorithms, we made a comparison between algorithms of the 4 proposed classes and a comparison between the proposed algorithms with the Standard PSO2006 and the Standard BPSO. The comparison results showed that the proposed algorithms outperform the Standard PSO2006 and the Standard BPSO in terms of quality of solution found. Keywords: PSO, BPSO, knapsack Problem (KP).
1 Introduction The Particle Swarm Optimization (PSO) is one of population-based solution metaheuristics inspired by an analogy with the ethology. It was created in 1995 by Kennedy and Eberhart [3]. PSO mimics the collective behavior of animals living in groups such as bird flocking and fish schooling. Simplicity and performance of this method have attracted interest of several communities of researchers who have conducted studies on optimization and application of this metaheuristic for solving several optimization problems. In this paper, we propose 4 classes of Binary PSO algorithms (BPSO) for solving the knapsack problem. The knapsack problem (KP) is a NP-hard problem [1,2]. It can be defined as follows: Assuming that we have a knapsack with maximum capacity C and a set of N objects. Each object i has a profit pi and a weight wi. The problem is to select a subset of items from the set of N objects to maximize the value of all selected objects without exceeding the maximum capacity of the knapsack. KP can be formulated as: Maximize Subject to
∑ ∑
i xi ixi
(1) C
A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 217–227, 2011. © Springer-Verlag Berlin Heidelberg 2011
(2)
218
A. Gherboudj and S. Chikhi
1 If the object i is selected xi= 0 Otherwise
i=1, ……,N
(3)
The remainder of this paper is organized as follows: the principle of the PSO is described in section 2. The third section concerns PSO variants. In the fourth section we describe the algorithms of each class. Comparison and experimental results are provided in section 5 and a conclusion is provided in the sixth section of this paper.
2 PSO Principle The PSO method involves a set of agents for solving a given problem. This set is called swarm, each swarm is composed of a set of members, they are called particles. Each particle is characterized by position xid= (xi1, xi2,…, xid,…, xiD) and velocity vid= (vi1, vi2,…, vid,…, viD) in a search space of D-dimension. During the search procedure, the particle tends to move towards the best position (solution) found. At each iteration of the search procedure, the particle moves and updates its velocity and its position in the swarm based on experience and the results found by the particle itself, its neighbors and the swarm. It therefore combines three components: its own current velocity, its best position pbestid= (pbesti1, pbesti2,…, pbestid,…, pbestiD) and the best position obtained by its informants. Thus the equations for updating the velocity and position of particles are presented below: vid(t)= vid (t-1) + c1 r1 (pbestid (t-1) - xid (t-1)) + c2 r2 (gbestd (t-1) - xid (t-1))
(4)
xid (t)= xid (t-1) + vid (t)
(5)
(xid (t), xid (t-1)), (vid (t), vid (t-1)): Position and Velocity of particle i in dimension d at times t and t-1, respectively. pbestid (t-1), gbestd(t-1) : the best position obtained by the particle i and the best position obtained by the swarm in dimension d at time t-1, respectively. c1, c2: two constants representing the acceleration coefficients. r1, r2: random numbers drawn from the interval [0,1[. vid (t-1), c1 r1 (pbestid (t-1) - xid (t-1)), c2 r2 (gbestd(t-1) - xid (t-1)): the three components mentioned above, respectively. The position of particle i represents a solution of the addressed problem. The value of the objective function (or fitness) of the particle i is denoted by f (xid). To estimate the quality of particle i, it is necessary to calculate its fitness. This one is calculated using a special function for the addressed problem. In the knapsack problem, the fitness is calculated according to equation (1). The PSO algorithm begins by initializing the size of the swarm and the various parameters. Assign randomly to each particle an initial position and velocity. Initialize pbestid, then calculate the fitness of particles in order to calculate the best position found by the swarm (gbestd). At each iteration, particles are moved using equations (4) and (5). Their objective functions are calculated and pbestid, gbestd are updated. The process is repeated until the satisfaction of stopping criterion. A pseudo PSO algorithm is presented below:
BPSO Algorithms for Knapsack Problem
1.
2.
3.
219
Particle Swarm Optimization Algorithm Initialization : • Parameters and size of the swarm (S); • Randomly initialize particles positions and velocities; • For each particle, pbestid = xid; • Calculate f (xid) of each particle; • Calculate gbestd; // the best pbestid While (termination criterion is not met) { For (i = 1 to S) { • Calculate the new velocity using equation (4); • Calculate the new position using equation (5); • Calculate f (xid) of each particle; • If (f (xid) >f (pbestid)) pbestid = xid; // Maximization case • If (f (pbestid) >f (gbestd)) gbestd = pbestid; } } Show the best solution found gbestd;
3 PSO Variants The idea of the pioneers of PSO algorithm: Kennedy and Eberhart [3] has sought the attention of several researchers who have conducted studies in the aim of improving the performance of the proposed method (PSO) which is not a global convergenceguaranteed optimization algorithm [5]. In 1996, Eberhart and al [15] proposed to limit the velocity of the particles in [-Vmax, Vmax] to avoid the problem of deviation of the search space during the movement of particles. The role of the new parameter Vmax is to control the movement of particles. In 1998, Shi and Eberhart [4] proposed to apply the inertia coefficient ω, to control the particles velocities as follows: vid(t)= ω vid (t-1) + c1 r1 (pbestid (t-1) - xid (t-1)) + c2 r2 (gbestd (t-1) - xid (t-1))
(6)
ω is an inertia coefficient. It is used to control the influence of particle velocity on his next move to keep a balance between exploitation and exploration of the search space. On the other hand, Clerc and Kennedy [9] proposed an alternative of equation (4). Their solution is to add a constriction coefficient K in the aim of controlling the speed of the particles to escape the divergence problem of the swarm that causes premature convergence of the algorithm. The proposed equation is: vid (t)= K [vid (t-1) + c1 r1 (pbestid (t-1) - xid (t-1)) + c2 r2 (gbestd (t-1) - xid (t-1))] Where K=
;With
= c1+ c2 and
(7)
; c1 = c2 = 2.05, K=0.729844.
To ensure the diversity of the swarm, Hi et al [7] proposed to update the particle velocity according to equation (8):
220
A. Gherboudj and S. Chikhi
vid (t)= ω vid (t-1) + c1 r1 (pbestid (t-1) - xid (t-1)) + c2 r2 (gbestd(t-1) - xid (t-1) ) + c3 r3 (Prid (t-1) - xid (t-1)) (8) Prid is the position of a particle i of swarm in the dimension d of the search space, this particle is selected randomly at time (t-1). The role of the component (Prid (t-1) - xid (t-1)) is to ensure the diversity of the swarm based on the value of the coefficient c3.
4 BPSO Algorithm The first version of BPSO algorithm (The Standard BPSO algorithm) was proposed in 1997 by Kennedy and Eberhart [11]. In the BPSO algorithm, the position of particle i is represented by a set of bit. The velocity vid of the particle i is calculated from equation (4). vid is a set of real numbers that must be transformed into a set of probabilities, using the sigmoid function as follows: (9) Where S (vid) represents the probability of bit xid takes the value 1. To avoid the problem of the divergence of the swarm, the velocity vid is generally limited by a maximum value Vmax and a minimum value -Vmax, i.e. vid The position xid of the particle i is updated as follows: 1 if r < S (vid) xid = 0 Otherwise
r [0, 1[
[-Vmax, Vmax].
(10)
In addition to the version of the Standard BPSO algorithm they exist other versions of BPSO algorithm, such as those proposed in [8, 12, 13, 14]. 4.1 Representation To represent the positions and velocities of the particles we used binary vectors of size D. The representation of position of particle i is as follows: xid = [xi1, xi2,…, xid,..., xiD] 1 If the object is selected xid=
0 Otherwise
4.2 Velocity and Position Update To represent the PSO principle, we need a number of operations and operators which are defined in [6].
BPSO Algorithms for Knapsack Problem
221
4.3 Proposed Classes In the aim of solving the KP, we have proposed four classes of BPSO algorithm. In each class, we have proposed four algorithms with different equations and parameters. 4.3.1 The First Class. In the first class we adapt and use the PSO version with inertia coefficient ω, proposed in 1998 by Shi and Eberhart [4]. In the algorithms of this class, the position of particles is updated according to equation (5). 1) BPSO6: It is an adaptation of the Standard PSO2006. In BPSO6, the velocity of particles is updated using the following equation: vid(t)= ω ×vid (t-1) + r1c1× (pbestid (t-1) - xid (t-1)) + r2 c2× (lbestd (t-1) - xid (t-1))
(11)
lbestd (t-1) is the best position found by the particles in dimension d of a given neighborhood. c1 and c2 are chosen randomly at each iteration. But in contrast to the standard PSO2006, The size of the swarm is equal to the dimension of the problem. 2) BP3: In BP3, the velocity is updated using Equation (11). c1 and c2 are constants. 3) BP2: In BP2, the velocity is updated according to equation (12) defined below: vid (t)= ω ×vid (t-1) + r1c1× (pbestid (t-1) - xid (t-1)) + r2 c2× (lbestd (t-1) - xid (t-1)) + r 3 (12) c3× (gbestd(t-1) - xid (t-1)) c1, c2 and c3 are constants. 4) BP1: To provide greater diversification within the swarm, we were inspired by the PSOPC algorithm [7] and we proposed to update the velocity of particles in BP1 algorithm using the following equation: vid(t)= ω ×vid (t-1) + r1c1× (pbestid (t-1) - xid (t-1)) + r2 c2× (lbestd (t-1) - xid (t-1)) + r 3 c3× (13) (gbestd(t-1) - xid (t-1)) + r4 c4× (Prid (t-1) - xid (t-1)) Where c1, c2, c3 and c4 are constants. Prid is the position of a particle i of swarm in the dimension d of the search space, this particle is selected randomly at time t-1. 4.3.2 The Second Class. In the second class we drew mutation factor used in the C3DPSO algorithm proposed by Zhong and Zhang [8] and we proposed a new acceleration coefficient F that we used to update particle position. 1) BFP6: In the BFP6 algorithm, the position of particles is updated according to equation (14) defined below: xid (t)= rF×xid (t-1) + vid (t)
(14)
The velocity of particles is updated according to equation (11). 2) BFP3: In BFP3, position and velocity of particles are updated according to equation (14) and (11), respectively. But c1 and c2 are constants. 3) BFP2: In BFP2, the position of particles is updated according to equation (14) and the velocity is updated according to the equation (12). 4) BFP1: In BFP1, the position of particles is updated according to equation (14). The velocity is updated according to equation (13).
222
A. Gherboudj and S. Chikhi
4.3.3 The Third Class. In the third class, we adapted and used the PSO version proposed in [9], because we noticed that the PSO algorithm with constriction coefficient K is not widely used in the literature. To our knowledge, there is no paper that addresses the KP using PSO algorithm with constriction coefficient. In the algorithms of this class, the position of particles is updated according to equation (5). 1) BCP6: In the BCP6 algorithm, the velocity of particles is updated using the following equation: vid (t)= K×[vid (t-1) + r1c1× (pbestid (t-1) - xid (t-1)) + r2 c2× (lbestd (t-1) - xid (t-1))]
(15)
c1 and c2 are chosen randomly at each iteration. 2) BCP3: In the BCP3 algorithm, the velocity of particles is updated using the equation (16), but c1 and c2 are constants. 3) BCP2: In BCP2 algorithm, the velocity of particles is updated using the following equation: vid (t)= K×[vid (t-1) + r1c1× (pbestid (t-1) - xid (t-1)) + r2 c2× (lbestd (t-1) - xid (t-1)) + r 3 (16) c3× (gbestd(t-1) - xid (t-1))] c1, c2 and c3 are constants. 4) BCP1: In BCP1 algorithm, we proposed to update the velocity of particles using the following equation: vid (t)= K× [vid (t-1) + r1c1× (pbestid (t-1) - xid (t-1)) + r2 c2× (lbestd (t-1) - xid (t-1)) + r3 (17) c3× (gbestd(t-1) - xid (t-1)) + r4 c4× (Prid (t-1) - xid (t-1))] Where c1, c2, c3 and c4 are constants; K=0.7. 4.3.4 The Fourth Class. This class includes algorithms defined in the third class with application of the new acceleration coefficient F. The position of particles is updated according to equation (14). 1) BFCP6: In the BFCP6 algorithm, the velocity of particles is updated according to equation (15). 2) BFCP3: In the BFCP3 algorithm, the velocity of particles is updated according to equation (15), but c1 and c2 are constants. 3) BFCP2: In the BFCP2 algorithm, velocity of particles is updated according to equation (16). 4) BFCP1: In the BFCP1 algorithm, velocity of particles is updated according to equation (17).
5 Comparison and Experimental Results To verify and compare the performance of the algorithms of the 4 proposed classes, 7 instances with different numbers of items were generated. In the first instance the number N of objects is equal to 120, in the second instance N = 200, in the third one
BPSO Algorithms for Knapsack Problem
223
N = 500 then N = 700, 900, 1000 and 2000 in the fourth, fifth, sixth and seventh instances respectively. Initially, we conducted a comparative study between the proposed algorithms of the four classes. Then we compared the proposed algorithms with the Standard PSO2006 [10] that we have adapted to the binary representation used. We also compared the proposed algorithms with the Standard BPSO [11]. The algorithms are coded in Java. Each algorithm is executed 125 times. The capacity of the knapsack is calculated using the following formula: i
The weights wi and profits pi of objects were selected randomly. For the algorithms of each class, the size of the swarm is equal to the number of items. In the 1st and 2nd classes, ω = 0.7. In the 3rd and 4th Classes, K was not calculated from the formula defined by Clerc and Kennedy i.e. K=
, but it was set at 0.7.
The values of the parameters c1, c2, c3, c4 and F are equal to 0.3, 0.4, 0.6, 0.1, and 0.9 respectively. Exceptionally in BPSO6, BFP6, BCP6 and BFCP6, the parameters c1 and c2 are drawn randomly from [0, 1[. The positions of particles were randomly initialized for each execution. The velocities were initialized with the value 0. The number of iterations in each run is chosen equal to 15 and is used as stopping criteria for each run. Concern the parameters of the standard PSO2006, we kept the same parameters defined in [10], but with binary representation of positions and velocities of particles. About the Standard BPSO, we followed the equations, representation and parameters defined in [11], except that the values of c1 and c2 are equal to those used for testing the proposed algorithms, i.e. c1, c2 = 0.3, 0.4, respectively.
Fig. 1. Comparison of average computation time of the proposed algorithms with the Standard PSO2006 and the Standard BPSO
224
A. Gherboudj and S. Chikhi
Fig. 1 shows a comparison of average computation time with 1000 objects, estimated by seconds for the proposed algorithms, the Standard PSO2006 (SPSO2006 in the figure) and the Standard BPSO (SBPSO in the figure). In terms of computing time, Fig.1 shows that: - The BCP3 algorithm is the best one and the BFCP1 algorithm is the worst one among the proposed algorithms. - The B*P2 and B*P3 algorithms (i.e. BP2, BFP2, BCP2, BFCP2, BP3, BFP3, BCP3 and BFCP3) converge faster than the Standard BPSO algorithm. - The Standard PSO2006 converges faster than the proposed algorithms. Tables 1 and 2 show the experimental results of algorithms of each class, classes 1 and 2 in Table 1 and Class 3 and 4 in Table 2. First column of each table represents the instance i.e. the number of items. The second and third column (Class 1 and Class 2 in the first table and Class 3 and Class 4 in the second table) represent the best solutions and averages found for each instance by the algorithms of the relevant class. Table 3 completes the tables 1 and 2. It represents the experimental results of the proposed algorithms, the Standard PSO2006 and the Standard BPSO for each instance during 125 executions. The first column represents the instance. The second column represents the best values of best and averages obtained by the proposed algorithms of the 4 classes. The third and fourth columns represent the bests and Averages obtained by the Standard BPSO and the Standard PSO2006 respectively. For each instance in tables 1, 2 and 3, the first row represents the best solution and the second row represents the average. Table 1. Comparison results of the proposed algorithms of Class 1 and Class 2 Instance 120 200 500 700 900 1000 2000
BPSO6 4439
Class 1 BP1 BP2 4469 4552
BP3 4457
BFP6 4463
Class 2 BFP1 BFP2 4489 4497
BFP3 4564 4236,4
4180,8
4130,4
4136
4236,4
4203,2
4140,4
4137,4
7559
7522
7642
7490
7491
7339
7648
7624
7104,4
6989,9
6938
7169,8
7108,6
6972,2
6979,8
7168,4
17647
17682
17810
17642
17647
17974
17776 16949 24678 23252,6
16848,2 16598,4 24407
24101
18058 17224 24431
17094,2 16864,8 16564,2 17175,8 24809
24407
24469
24563
23335,6 23019,4 23867,6 23767,6 23368,4 23027,8 30507,4
31428
30898
31215
29659,4
29833,8
29369
31686
31192
31018
30509,8 30391,2 29841,6
34654 34319 34213 35019 34847 32912,2 33065 32558,6 33810,6 33637 68857 67914 66605 66546 68548 63232,75 65030,6 63754,6 66478 66185,2
34273 33023 67110 65037
31525
31276
29330
30507,4
33999 34596 32543,2 33882,4 67507 67829 63725 66716,6
BPSO Algorithms for Knapsack Problem
225
Table 2. Comparison results of the proposed algorithms of Class 3 and Class 4 Class 3
Instance 120 200 500 700 900 1000 2000
BCP6 4543 4226,2 7506 7139,4 17922 17069,2 24651 23589,6 31488 30111 34680 33291,6 66703 65275,2
BCP1 4497 4216,2 7614 7126,2 17983 17093,8 24603 23637,4 31295 30063 34606 33384,8 68266 66656,8
Class 4
BCP2 BCP3 4545 4533 4204 4242,4 7474 7556 7110,4 7183 17810 17906 17159,6 17265,4 24558 24729 23787,2 23926,8 31105 31240 30439,2 30537,6 34863 35897 33819,6 33910,8 67856 67995 66615 66779,75
BFCP6 4552 4229,8 7531 7151 17642 17175,8 24540 23876 31324 30448 34934 33879,4 68784 66322,8
BFCP1 4538 4230,4 7646 7133,6 17838 17192,4 24634 23835 31448 30504,6 34745 33882,4 68378 66740
BFCP2 4512 4207 7681 7097 17731 17129,4 24326 23772,6 31082 30455,2 34510 33776 67645 66682,4
BFCP3 4487 4257,4 7555 7203,6 18332 17268 24951 23934,6 31478 30495,2 34717 33863 67689 66630,8
Table 3. Comparison of best values obtained by the proposed algorithms, the Standard PSO2006 and the Standard BPSO Instance 120 200 500 700 900 1000 2000
Best Known 4564 4257,4 7681 7203,6 18332 17268 24951 30537,6 31686 30537,6 35897 33910,8 68857 66779,75
Standard BPSO
Standard PSO2006
4296 3840,8 7456 5703 13116 12471,2 18276 17097,4
4331 4027 7391 6819,4 17618 16244,4 23893 22400,2 30770 28574,2 34025 31682,2 67006 63265,8
22857 21736,6 24933 24050 47674 46538,8
226
A. Gherboudj and S. Chikhi
Tables 1, 2 and 3 show that: - The use of PSO algorithm version with constriction coefficient gives good averages compared to the PSO version with inertia coefficient. - In most cases, the application of the acceleration coefficient F on the algorithms of the first class (which gave birth to algorithms of the second class) has improved their results in terms of averages. - The use of the acceleration coefficient F in the algorithms of the third class improves their results. - In most cases, the application of acceleration coefficient F on the version of PSO algorithm with constriction coefficient gives good averages compared with its application to the version of PSO algorithm with inertia coefficient ω. - Best averages are obtained by B*P3 i.e. BP3, BFP3, BCP3 and BFCP3. - In most cases, best values are obtained by B*P3 i.e. BP3, BFP3, BCP3 and BFCP3. - The performance of the 16 proposed algorithms exceed those of the Standard PSO2006 and the Standard BPSO in terms of best solution found and average.
6 Conclusion PSO is a recent metaheuristic. It has sought the attention of several research communities. PSO has proved its simplicity of implementation and effectiveness. Several variants to the original PSO algorithm have been proposed in the literature to improve its performance. In this contribution, we drew some works and applications of the PSO algorithm presented in the literature, and we proposed 4 classes of BPSO algorithms with different equations for updating velocities and positions of particles. We have grouped the proposed algorithms into four classes: in the first class, we adapted and used the PSO version with inertia coefficient [4]. The new acceleration coefficient F is used in the second class for updating the particles positions. F was applied on the algorithms proposed in the first class which has given birth to the second class of algorithms. In the third class we adapted and used the PSO version with constriction coefficient [9] because we noticed that few studies use this version. In the fourth class, we used the acceleration coefficient F for the update of particles positions and the constriction coefficient for the update of particles velocities. We applied the proposed algorithms for solving the NP-hard knapsack problem using multiple instances (120, 200, 500, 700, 900, 1000 and 2000 objects). To verify the performance of the proposed algorithms, we conducted a comparative study between the proposed algorithms of the four classes and a comparison of the proposed algorithms with the Standard PSO2006 [10] and the Standard BPSO [11]. Comparative studies of the proposed algorithms show performance improvements with the use of the new acceleration coefficient F for the updating of position and the application of the constriction coefficient K for the updating of velocity. In terms of average and best solutions, experimental results show that the proposed algorithms outperform the Standard PSO2006 and the Standard BPSO. In terms of average, best solutions and computation time, experimental results show that the B*P2 and B*P3 algorithms (i.e. BP2, BFP2, BCP2, BFCP2, BP3, BFP3, BCP3 and BFCP3) outperform the Standard BPSO.
BPSO Algorithms for Knapsack Problem
227
References 1. Xie, X., Liu, J.: A Mini-Swarm for the quadratic Knapsack Problem. In: IEEE Swarm Intelligence Symposium (SIS), Honolulu, HI, USA, pp. 190–197 (2007) 2. Pisinger, D.: Where are the hard knapsack problems? Computers and Operations Research 32(9), 2271–2284 (2005) 3. Kennedy, J., Eberhart, R.C.: Particle Swarm Optimization. In: Proc. IEEE Int. Conf. On Neural Networks, WA, Australia, pp. 1942–1948 (1995) 4. Shi, Y., Eberhart, R.: Parameter Selection in Particle Swarm Optimisation. In: Proceedings of the 7th Annual Conference on Evolutionary Programming. LNCS, vol. 1447, pp. 591–600. Springer, Heidelberg (1998) 5. Wang, J., Zhou, Y.: Quantm-behaved Particle Swarm Optimization with Generalized Local Search Operator for Global Optimization. In: Advanced Intelligent Computing Theories and Applications With Aspects of Artificial Intelligence, pp. 851–860. Springer, Heidelberg (2007) 6. Gherboudj, A., Chikhi, S.: Algorithme d’OEPB pour Résoudre le Problème du Sac à Dos. In: Laouar, M.R. (ed.) Proceedings of the 1st International Conference on Information Systems and Technologies, ICIST 2011, Tebessa, Algeria, pp. 460–466 (2011) ISBN: 9789931-9004-0-5 7. He, S., Wu, Q.H., Wen, J.Y., Saunders, J.R., Paton, R.: A Particle Swarm Optimizer with Passive Congregation. Biosystems, 135–147 (2004) 8. Zhong, W., Zhang, J., Chen, W.: A Novel Discrete Particle Swarm Optimization to Solve Traveling Salesman Problem. In: IEEE Congress on Evolutionary Computation, CEC 2007, pp. 3283–3287 (2007) 9. Clerc, M., Kennedy, J.: The Particle Swarm: Explosion, Stability, and Convergence in Multidimensional Complex Space. IEEE Transactions on Evolutionary Computation 6, 58–73 (2002) 10. Standard PSO2006, http://www.particleswarm.info/Programs.html 11. Kennedy, J., Eberhart, R.C.: A discrete binary version of the particle swarm algorithm. In: Proceedings of the World Multiconference on Systemics, Cybernetics and Informatics, Piscatawary, NJ, pp. 4104–4109 (1997) 12. Afshinmanesh, F., Marandi, A., Rahimi-Kian, A.: A novel binary particle swarm optimization method using artificial immune system. In: Proccedings of IEEE international conference on computer as a tool, pp. 217–220 (2005) 13. Liao, C., Tseng, C., Luarn, P.: A discrete version of particle swarm optimization for flowshop scheduling problems. Computers & Operations Research 34(10), 3099–3111 (2007) 14. Zhan, Z.-h., Zhang, J.: Discrete particle swarm optimization for multiple destination routing problems. In: Giacobini, M., Brabazon, A., Cagnoni, S., Di Caro, G.A., Ekárt, A., Esparcia-Alcázar, A.I., Farooq, M., Fink, A., Machado, P. (eds.) EvoWorkshops 2009. LNCS, vol. 5484, pp. 117–122. Springer, Heidelberg (2009) 15. Eberhart, R.C., Simpson, P., Dobbins, R.: Computational PC Tools, ch. 6, pp. 212-22, AP Professional (1996)
Systematic Selection of CRC Generator Polynomials to Detect Double Bit Errors in Ethernet Networks Behrouz Zolfaghari1, Hamed Sheidaeian1, and Saadat Pour Mozafari2 1
Engineering Department, Islamic Azad University, Garmsar Branch, Garmsar, Iran
[email protected],
[email protected] 2 Department of Computer Engineering, AmirKabir University of Technology, Hafez Street, Tehran, Iran
[email protected] Abstract. CRC (Cyclic Redundancy Check) is used as an error detection code in Ethernet frames. This method attaches the residue of a modulo-2 division to the message in the sender side and recalculates the residue in the receiver side. The agreed-upon divisor is called the generator. The range of detectable errors in this method is determined by the properties of the generator. In this paper a systematic approach is proposed to identify all proper generators which can be used in Ethernet networks in order to detect double bit errors which invert a pair of bits in the message. Keywords: double bit errors; generator polynomials; OZO polynomials.
1 Introduction and Basic Concepts In our previous work [1], we developed a systematic scheme to select proper generator polynomials which can help detect burst errors in Ethernet frames by CRC. In this paper, we will augment our previous work by developing another method to select generators able to detect double bit errors. Let us shortly examine the CRC method before discussing the proposed approach. CRC works as follows. Whenever the sender has a message M to send, it first concatenates n zero bits to the right of the massage, converting it to M .2 n ( n is the length of an agreed-upon string called the generator subtracted by one It is also the length of the CRC. Especially, the CRC used by Ethernet is 32 bits long [10]). The sender divides the produced string ( M .2 n ) by the generator ( G ) in the next step and calculates the residue ( R = ( M .2 n ) ModG ). Then the residue is replaced for the n zero bits. The string is now converted to M ' = MR = M .2 n + ( M .2 n ) ModG . The string M ' = MR is transmitted instead of M .Figure 1 shows these steps. The addition, multiplication and division operations are performed modulo-2 here. Since addition and subtraction are the same in modulo-2 computations [1], we can think of M ' as M ' = MR = M .2 n − ( M .2 n ) ModG which is obviously divisible by G . A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 228–235, 2011. © Springer-Verlag Berlin Heidelberg 2011
Systematic Selection of CRC Generator Polynomials
229
The receiver divides what it receives by G again and calculates the residue. If the receiver gets exactly the string transmitted by the sender, the residue will obviously be equal to zero. Figure 2 shows this process.
Fig. 1. The transmitter side CRC Process
Fig. 2. The Receiver side CRC Process
But if an error has occurred through the channel, we can model the error as a string
E added to M ' [1]. In such a case, the receiver receives M '+ E instead of M ' . Since M ' is divisible to G , the calculated residue in this case will be equal to ( M '+ E ) Mod G = E Mod G .
The message, the generator string, the residue and the error vector are commonly represented in the form of polynomials. In the polynomial form of a bit string, 1s are shown by the exponents of a variable such as x . Such polynomials appear in the form of
∑a x i
example
i
. Each a i (being 0 or 1) represents the corresponding bit in the string. For the
bit
string
11101
can
be
represented
by
the
polynomial x + x + x + 1 . If a string contains n bits it is of degree n − 1 .An obvious result is that strings with odd lengths are presented by polynomials of even degrees and vice versa. In this paper, we use the terms string and polynomial interchangeably. Readers are referred to [10] for more details regarding these concepts. If E (called the error vector) is divisible to G , the receiver will come to zero as the residue and interpret s this situation as error-free reception of the message. Thus, CRC cannot detect errors whose vectors are divisible to the generator. Therefore the generator must be selected in a way that it does not have any multiples equal to the error vectors which commonly occur in the system. Applications of CRC [2, 4, 5, and 9] as well as developing methods for improving its efficiency [3, 6, 7, and 11] have been research focuses in recent years. In this paper we will propose a systematic method based on modulo-2 mathematics to list all 33-bit strings (polynomials of degree 32) which can detect double bit errors if used as generator polynomials. The rest of this paper is organized as follows. Section 2 present Preliminary discussions, section 3 explains the proposed method and section 4 is dedicated to conclusions and suggesting further works. 4
3
2
230
B. Zolfaghari, H. Sheidaeian, and S.P. Mozafari
2 Preliminary Discussions A double bit error is an error that inverts two distant bits of the message. Such an error has a vector like 00…010…0100…0. The substrings bounded between the two 1s (10….01) are called OZO (One-Zero-One) strings in this paper. An OZO string consists of two 1s in the right and the left. The equivalent polynomial form of an OZO string is like x n + 1 . We also refer to every polynomial that has an OZO multiple as an ODP (OZO Dividing Polynomial). Especially, every ODP of degree 32 is called an ODP32. It is obvious that ODP32 generators cannot detect double bit errors. Strings representing double bit error vectors can be shown by polynomials containing two exponents of x like x n + x m where n is the degree of the polynomial and m is the number of 0s in the right side of the string. As we will see later, OZO polynomials have divisors having all bits equal to 1 (like 111…11). Such strings are called ALO (All One) strings. ALO polynomials represent burst errors which change a number of consecutive bits in the message [1]. Every ALO polynomial of an even degree is called an even ALO polynomial and every ALO polynomial of an odd degree is called an odd ALO polynomial. An ALO strings is represented by a polynomial of the form: n x i in which n is the degree of
∑ i =0
the polynomial. OZO polynomials of odd degrees are referred to as odd OZO polynomials and those which have even degrees are called even OZO polynomials in this paper. The case is similar for even and odd ALO polynomials.
3 The Method Before explaining the proposed method, we need some lemmas presented below. Some of these lemmas appear without proof. Readers are referred to [1] for the proofs of these lemmas. Lemma 1: Every generator polynomial having a constant term (representing strings having a 1 bit at the right) can detect every single bit error. In order to exploit the above lemma, all standard CRC systems use generator polynomials which have constant terms (1s at the right). Standard generator n
polynomials of degree n also include x (the corresponding bit strings have 1s in the left).Therefore, in the rest of this paper, we assume that all generator strings have 1s in their left and right sides. Lemma 2: Every generator polynomial of degree m can detect all burst errors whose non-zero parts have degrees less than m . The non-zero part of an error vector is defined as the substring located between the first and the last 1 bit. The above Lemma states that we can focus on error vectors whose non-zero substrings have degrees larger than that of the generator (32 in Ethernet). Lemma 3: If the vector of a double bit error is divisible by the generator G , then its OZO substring is divisible by G .
Systematic Selection of CRC Generator Polynomials
Proof: Let us represent the error vector as x n + x m = x m ( x n−m + 1) . Since
231
x m cannot
be divided by G , the other term ( x n −m + 1 ) should be divisible by G . The latter term is the OZO substring of the error vector. The above lemma reduces our problem (finding generators able to detect double bit errors) to the problem of finding generators having OZO multiples. In other words, in order to detect double bit errors by CRC in Ethernet frames, the selected generator should not be an ODP. According to lemma 3, the main idea behind our proposed method is finding all ODP32 generators an excluding them from the list of possible generators of degree 32. The remaining generators have the ability to detect double bit errors. Therefore, in the rest of this paper, we attempt to find a solution to the problem of generating ODP32s. Separating ODP32s from all possible generators of degree 32 will determine generators which can detect double bit errors. Lemma 4: Every generator including an even number of 1s can detect every odd error. An odd error is an error that changes an odd numbers of bits in the message. Lemma 4 states that in order to detect errors which change odd numbers of errors, we should simply select generators having even numbers of 1s. Thus, we should focus on even errors which change even numbers of bits. Thus the main challenge in designing CRC systems is the selection of generators which can cover even errors. Lemma 5: Every odd ALO polynomial of degree 2 k + 1 can be factored as the product of an OZO polynomial of degree k + 1 and an ALO polynomial of degree k . The following lemma clears how to factor OZO polynomials of even degrees. Lemma 6: Every OZO polynomial of degree 2 k is the square of an OZO polynomial of degree k . Lemma 7: Every OZO polynomial of degree k + 1 can be factored to x + 1 and an ALO polynomial of degree k . Now we can prove a useful lemma that helps factor every OZO polynomial. Lemma 8: Every OZO polynomial is reducible to the product of an exponent of x + 1 and an exponent of an even ALO polynomial.
m can be factored as
Proof: First notice that every OZO polynomial of degree follows.
x
+ 1 = ( x + 1 ).
m
m −1
∑
x
i
(1)
i= 0
m − 1 is odd it can be written as 2 k + 1 and we will have
Now if m −1
∑
i= 0
x
i
= (x
k +1
+ 1 ).
k
∑
i= 0
x
i
= ( x + 1 ).
k
∑
i= 0
k
x i .∑
i= 0
x
i
232
B. Zolfaghari, H. Sheidaeian, and S.P. Mozafari
= ( x + 1 )(
k
∑
x i)
2
(2)
i= 0
In this case, we will have: x
m
+ 1 = ( x + 1 ) 2 .(
k
∑
(3)
2
xi)
i= 0
Again, if
k is odd, ∑k
x
i
will be factorable and this factoring can be continued
i= 0
until there remain only exponents of x + end of the factoring process, we will have: Or
x
m
+ 1 = ( x + 1)
1+
t −1
∑
2
i
i= 0
1 and ALO even polynomials. At the
2 r
∑
.(
xi)
2
t
(4)
i= 0
Equation 4 can be rewritten as follows.
x
m
+ 1 = ( x + 1)
2
t
2 r
∑
.(
xi)
2
t
(5)
i= 0
Where 2 t.2 r + 2 Now we can write:
x
2 t .( 2 r + 1 )
t
= 2 t .( 2 r + 1 ) = m
+ 1 = ( x + 1)
2
t
2 r
.(
∑
xi)2
t
(6)
i= 0
Again, equation 6 can be written in the following form.
x
2 t .( 2 r + 1 )
+ 1 = (( x + 1 ).
2 r
∑
i
x )
2
t
i= 0
= (x
2 r +1
+ 1)
2
t
(7)
Equation 7 states that we can factor any OZO polynomial of degree m , by dividing its degree by 2 (for t times) until there remains an odd number 2 r + 1 and 2 r +1
then writing the polynomial as ( x + 1) . It is obvious that If m is odd, t will be equal to 0. Considering the fact that x + 1 is a prime polynomial, lemma 8 reduces the problem of selecting generators for the detection of double bit errors into prime factorization of even ALO polynomials. Lemma 8 exploits previous lemmas and gives the straightforward way to produce ODP32s using x + 1 and even ALO polynomials. Considering the fact that x + 1 is a prime polynomial, the above lemma reduces the problem of prime factorization of all ALO polynomials to that of even ALO polynomials. 2
t
Systematic Selection of CRC Generator Polynomials
An obvious result of lemma 8 is that For each m , n , r ∈ N
233
∪ {0 },
if
2 r
∑
2 r . m + n > 0 then ( x + 1 ) n .(
x i ) m will be an ODP32. The
i= 0
equation 2 r . m + n = 32 has 56 sets of answers If it is solved for 2 r , m and n . It is obvious that the greatest possible value for 2 r can be 32. Thus, we continue by prime factoring ALO polynomials of degrees 2 r ∈ [0 ,32 ] . Table 1 shows the prime factorizations of the mentioned polynomials. Table 1. Prime factorizations of even ALO polynomials
Poly
Prime Factorization Prime
0
∑
x
i
Polyl
∑
i= 0
x
i
i= 0
2
∑
Prime Factorization Prime
18
i
x
Prime
20
∑
i
x
2
.( x 3 + x 2 + 1).( x 3 + x + 1).∑ x i
i= 0
i= 0
(x6 + x5 + x4 + x2 + 1).(x6 + x4 + x2 + x + 1)
i=0
Prime
4
∑
i
x
22
∑
i= 0
i= 0
( x + x + 1).( x + x + 1) 3
6
∑
i
x
2
24
3
∑
8
i
x
i= 0
∑
( x 6 + x 3 + 1). ∑ x i
x
∑
4
x
2
( x18 + x 9 + 1).( x 6 + x 3 + 1).∑ x i i=0
x
i
x
i
Prime
i= 0
12
∑
5
i=0
28
i= 0
6
4
i
i= 0
Prime
i
( x + x + x + x + x + x 2 + 1) .( x11 + x 9 + x 7 + x 6 + x 5 + x + 1) 10
( x 20 + x 15 + x 10 + x 5 + 1). ∑ x i
i
26
2
i=0
10
∑
x
i= 0
i= 0
∑
i
x
11
x
i
30
Prime
∑
i= 0
i= 0
(x5 + x4 + x3 + x +1).(x5 + x4 + x3 + x2 + 1) .(x5 + x4 + x2 + x +1).(x5 + x4 + x2 + x + 1)
.( x 5 + x 3 + 1).( x 5 + x 2 + 1) 2
14
∑
x
i
16
i= 0
x
i
i =0
i= 0
∑
∑ x .( x
i
32
4
+ x 3 + 1).( x 4 + x + 1). ∑ ∑ i
i= 0
x
i
( x10 + x9 + x5 + x + 1).(x10 + x7 + x5 + x3 + 1) 2
10
i =0
i =0
.∑ x i .∑ x i
( x 8 + x 7 + x 6 + x 4 + x 2 + x + 1)
.( x 8 + x 5 + x 4 + x 3 + 1)
The factorizations listed in Table 1 have been obtained from a program written in C. This program first generates and stores prime polynomials of degree 2 ( x 2 + x + 1 and x 2 + 1 ). Then generates all third degree polynomials and divides each of them to the stored prime polynomials. If a polynomial is divisible to none of the stored polynomials, it is stored as a new prime polynomial. Higher degree prime polynomials are detected and stored in a similar way. This program produces ALO
234
B. Zolfaghari, H. Sheidaeian, and S.P. Mozafari
polynomials of different degrees and divides each of them by each of the stored prime polynomials in the next phase. This phase gives the prime factorization of each of the ALO polynomials. Now can exploits the factorizations showed in the above table in order to solve the factorization problem. For example, let us solve the equation for 2 r = 6 . In this case, 6 must be an ALO polynomial. We can 2 −1 2 i ( x + 1)
t
.(
∑
x )
t
i= 0
replace ( x 3 + x 2 + 1).( x 3 + x + 1) for
6
∑
i
x
using table 1. This replacement shows that
i= 0
( x + 1 ) 2 − 1 .(( x 3 + x 2 + 1 ).( x 3 + x + 1 )) 2 must be an ALO polynomial. An immediate result is that for every x , y , z ∈ N ∪ {0 } , a 3 2 b 3 c ( x + 1 ) .( x + x + 1 ) .( x + x + 1 ) is an ODP32 if a + b + c = 32 . Thus, every answer to the equation a + b + c = 32 (considering a , b , c ∈ N ∪ {0 }) gives an ODP32. As another example, Let us solve the factorization problem for 2 r = 18 . In this t
case, ( x + 1 ) 2
t
t
−1
18
.(
∑
t x i ) 2 must be an ALO polynomial. But since
Thus for every a , b ∈ N ∪ {0 },
x
i
is
i= 0
i= 0
not reducible, we conclude that
18
∑
( x + 1) 2
t
−1
18
.(
∑
xi)2
t
must be an ALO polynomial.
i= 0
( x + 1 ) a .(
18
∑
x i)b
is an ODP32 if x + y = 32 . This,
i= 0
every answer to the equation a + b = 32 (considering a , b , c ∈ N ∪ {0 }) gives an ODP32 in the form ( x + 1 ) a .( 18 x i ) b .Through a similar process, we can solve
∑
i= 0
the equation for 2 r = 0 , 2 , 4 ,..., 32 . This way, we will find 3809 ODP32s. Now let us calculate the number of polynomials that can detect double bit errors in Ethernet frames. We know that there 2 33 different polynomials of degree 32. Half of them ( 2 32 ) include even numbers of 1s and 1 of this polynomials ( x 30 ) include 4 32 x and 1 . Since there are 3809 polynomials unable to detect double bit errors, the result is that there are 2 30 − 3809 polynomials of degree 32 that can be used as CRC generators in order to detect double bit errors in Ethernet frames.
4 Conclusions and Further Works This paper proposed a systematic non-exhaustive method based on modulo-2 mathematics to list all polynomials of degree 32 which can be used in order to detect double bit errors in Ethernet frames. The method proposed in this paper eliminates the need for time-consuming exhaustive searches and finds the considered polynomials in a short time. It was demonstrated that there are 2 30 − 3809 such polynomials. This work can be continued by proposing methods for simplifying CRC computation with such generators.
Systematic Selection of CRC Generator Polynomials
235
References 1. Zolfaghari, B., Mozaffari, S.P., Karkhane, H.: A Systematic Approach to the Selection of CRC Generators to Detect Burst Errors in Ethernet Networks. In: Proceedings of the IEEE International Conference on Intelligent Network and Computing (ICINC 2010), Kuala Lumpur, Malaysia (November 2010) 2. Deng, I., Rong, M., Liu, T., Yuan, Y., Yu, D.: Segmented Cyclic Redundancy Check: A Data Protection Scheme for Fast Reading RFID Tag’s Memory. In: Proceedings of IEEE Wireless Communications & Networking Conference (WCNC 2008), March 31- April 3, pp. 1576–1581. IEEE Computer Society Press, Las Vegas (2008) 3. Mathys, W.: Pipelined Cyclic Redundancy Check (CRC) Calculation. In: Proceedings of 16th International Conference on Computer Communications and Networks, ICCCN 2007, pp. 365–370 (1963) 4. Ahmad, A., Hayat, L.: Algorithmic Polynomial Selection Procedure for Cyclic Redundancy Check for the use of High Speed Embedded Networking Devices. In: Proceedings of International Conference on Computer and Communication Engineering 2008 (ICCCE 2008), Kuala Lumpur, Malaysia, pp. 13–15 (2008) 5. Pana, Y., Ge, N., Dong, Z.: CRC Look-up Table Optimization for Single-Bit Error Correction. Tsinghua University Journal of Science & Technology 12(5), 620–623 (2007) 6. Assaf, R., Shmuel, T.: The EasyCRC Tool. In: Proceedings of 2007 International Conference on Software Engineering Advances (ICSEA 2007), pp. 25–31 (August 2007) 7. Young, M.: The Technical Writer’s Handbook. University Science, Mill Valley (1989) 8. Dongliang, X., Jianhua, L., Chunlei, L., Bingli, J.: A Perturbation Method for Decoding LDPC Concatenated with CRC. In: Proceedings of Wireless Communications and Networking Conference (WCNC 2007), pp. 667–671 (March 11-15, 2007) 9. Zhanli, L., Xiao, L., Chunming, Z., Jing, W.: CRC-Aided Turbo Equalization For MIMO Frequency Selective Fading Channels. Journal of Electronics(China) 24(1), 69–74 (2007) 10. Tanenbaum, A.S.: Computer Networks, 3rd edn. Prentice Hall, Englewood Cliffs (1996) 11. Sudha, R., Wilson, G.S., Yalamarthy: Near-ML Decoding of CRC Codes. In: Proceedinggs of 41st Annual Conference on Information Sciences and Systems, pp. 92–94 (March 14-16, 2007)
Security Analysis of Ultra-lightweight Protocol for Low-Cost RFID Tags: SSL-MAP Mehrdad Kianersi, Mahmoud Gardeshi, and Hamed Yousefi Dep. Communication and Information Technology, IHU Tehran, Iran {Mehrdad_3264,Mgardeshi2000,Hamed.yousefi}@yahoo.com
Abstract. In this paper, we analyze the security vulnerabilities of SSL-MAP, an ultra-lightweight RFID mutual authentication protocol recently proposed by Rama N, Suganya R. We present two effective attacks, a de-synchronization attack and a full-disclosure attack, against this protocol. The former permanently disables the authentication capability of a RFID tag by destroying synchronization between the tag and the RFID reader. The latter completely threats a tag by extracting all the secret information that are stored in the tag. The de-synchronization attack can be carried out in three round of interaction in SSL-MAP while the full-disclosure attack is accomplished across several runs of SSL-MAP. We also discuss ways to counter the attacks. Keywords: RFID, Mutual authentication, Low-cost RFID Tag, SSL-MAP.
1 Introduction Radio Frequency Identification (RFID) systems offer improved efficiency in inventory control, logistics, and supply chain management. As such, they are of great interest to enterprises intensively reliant on supply chains, particularly large retailers and consumer product manufacturers. The long-term goal of these organizations is to integrate RFID on the retail level. Without proper protection, widespread adoption of retail RFID could raise privacy concerns for everyday consumers. RFID systems consist of three main components: tags, readers and back-end databases. Tags are radio transponders attached to physical objects. Each tag contains a microchip with a certain amount of computational and storage capabilities and a coupling element. Such devices can be classified according to memory type and power source. Another relevant parameter is tag price, which creates a broad distinction between high-cost and low-cost RFID tags. Radio transceivers, or readers, query these tags for some (potentially unique) identifying information about the objects to which tags are attached. Although readers are often regarded as a simple conduit to a back-end database, for simplicity we treat a reader and a back-end database as a single entity.
2 Related Works In [13], Chien proposed a tag classification mainly based on which were the operations supported on-chip. High-cost tags are divided into two classes: A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 236–245, 2011. © Springer-Verlag Berlin Heidelberg 2011
Security Analysis of Ultra-lightweight Protocol for Low-Cost RFID Tags
237
“full-fledged” and “simple”. Full-fledged tags support on-board conventional cryptography like symmetric encryption, cryptographic one-way functions and even public key cryptography. Simple tags can support random number generators. Likewise, there are two classes for low-cost RFID tags. “Lightweight” tags are those whose chip supports a random number generation and simple functions like a Cyclic Redundancy Code (CRC) checksum, but not cryptographic hash function. “Ultralightweight” tags can only compute simple bitwise operations like XOR, AND, OR, etc. These ultra-lightweight tags represent the greatest challenge in terms of security, due to their expected wide deployment and very limited capabilities. In 2006, Peris et al. proposed a family of Ultra-lightweight Mutual Authentication Protocols (henceforth referred to as the UMAP family of protocols). Chronologically, M2AP [4] was the first proposal, followed by EMAP [5] and LMAP [6]. These protocols are based on the use of pseudonyms to guarantee tag anonymity. Specifically, an indexpseudonym is used by an authorized reader to retrieve the information associated with a tag (tag identification phase). Additionally, a key-divided in several sub-keys is shared between legitimate tags and readers (back-end database). Both readers and tags use these sub-keys to build the messages exchanged in the mutual authentication phase. In continue of their real processing capabilities, tags only support on-board simple Operations. Indeed, these protocols are based on bitwise XOR, OR, AND and addition mod 2m. By contrast, only readers need to generate pseudorandom numbers; tags only use them for creating new messages in the protocol. In the UMAP family of protocols, the proposed scheme consists of three stages. First, the tag is identified by means of the index-pseudonym. Second, the reader and the tag are mutually authenticated. This phase is also used to transmit the static tag identifier (ID) securely. Finally, the index-pseudonym and keys are updated (the reader is referred to the original papers for more details). Since publication of the UMAP family of protocols, their security has been analyzed in depth by the researchers. In [7, 8] a de-synchronization attack and a full-disclosure attack are presented. These require an active attacker and several incomplete run executions of the protocol to disclose the secret information on the tag. Later, Chien et al. proposed a attack based on the same attack model, more efficient from full-disclosure attack [9]. Additionally, B´ar´asz et al. showed how a passive attacker (an attack model that may be, in certain scenarios, much more realistic) can find out the static identifier and particularly secret information shared between reader and tag after eavesdropping on a few consecutive rounds of protocol [10, 11]. In 2007 Hung-Yu Chien proposed a very interesting ultra-lightweight authentication protocol providing Strong Authentication and Strong Integrity (SASI) for very low-cost RFID tags [13]. However, In 2009 Hernandez-Castro et al. have showed that the protocol was not carefully designed [14]. Indeed, a passive attacker can obtain the secret static identifier of the tag (ID) after observing several consecutive authentication sessions.
3 Review of SSL-MAP The protocol comprises three stages: Tag identification, mutual authentication, and updating as shown in Fig. 1.
238
M. Kianersi, M. Gardeshi, and H. Yousefi
, ,
, ,
, , ,
, ,
, , ,
,
,
Where c = 0x3243F6A8885A308D313198A2
Fig. 1. SSL-MAP Protocol
3.1 Tag Identification The reader sends a “hello” message to the tag. The tag responds with indexpseudonym (IDS). The reader uses this ID as a reference number to search for the shared keys of the tag in its database. If the database has an entry associated an IDS, next phase starts, otherwise the reader requests for older IDS to identify the tag. 3.2 Mutual Authentication With IDS, the reader acquires private information linked to the tag from the database. Then the reader generates pseudonyms n1 and n2, constructs three concatenated public messages ‖ ‖ and sends them to the tag. Where c is a 96 bit length constant. The tag in reply sends a public message D or an error message depending on successful reader authentication. So we have two authentications as follow: 1. Reader Authentication: From messages A and B, the tag extracts pseudonyms n1 and builds a local version of and n2 respectively. Then it computes n3, k1*, k2*, message C as C′. This is compared with the received value C. If both values are same, the reader is authenticated. 2. Tag Authentication: Finally, the tag sends message D to the reader. On receiving D, this value is compared with a computed local version. If they are same, the tag is authenticated; otherwise the protocol is abandoned. 3.3 Updating After successfully completing the mutual authentication phase between the reader and the tag, both locally update IDS and keys k1, k2 as follows:
Security Analysis of Ultra-lightweight Protocol for Low-Cost RFID Tags MixBits n , n
239
(1) n
,n
n
,n ,
,n
(2)
,
(3) ,
(4)
4 Security Analysis Security analysis of SSL-MAP protocol result that following attacks are possible to carry out. 4.1 De-synchronization Attack The tag updates its values irrespective of whether the reader has received message D and verified it or not whereas, the reader updates its values only after receiving and verifying message D. This causes a difference between the storage of the tag and the reader in case that message D does not received by the reader. To avoid this desynchronization, in Gossamer, the tag is considered to be keeping the older values of IDS and keys in memory. So in such case that a de-synchronization occurs the reader can ask for the older IDS (not updated) and both can be in synchronization again. However, a de-synchronization attack can still be launched successfully using following procedure [3]: 1. Suppose a tag keep the synchronized value as: 1.
=
2. 3. This tag now communicates with a reader. The attacker records the corresponding message as Ax, Bx, Cx (being public messages and under the assumption that communication between the reader and the tag is not secure). Now the attacker interrupts message Dx and does not allow it reaches to the reader. The tag does not know whether the reader has verified D or not and updates its value as: 1.
2.
3.
4.
5.
6.
2. Next, the attacker allows the tag and the reader to run the protocol without intervening. As IDSy is not recognized by the reader (did not update its value as D was not received), so it asks the tag for the older values. The tag sends IDSx which is recognized by the reader and they complete the protocol. After completion, the tag updates its values as: 1. 4.
2. 5.
3. 6.
240
M. Kianersi, M. Gardeshi, and H. Yousefi
3. Now, the attacker intervenes and sends a ”hello” message to the tag. The tag responds with IDSz. The attacker pretends that it cannot identify IDSz and asks for the older values. The tag responds with IDSx. Attacker has already copied Ax, Bx, Cx, which are legitimate sub-messages against IDSx and n1x, n2x generated during step1. Protocol is completed and tag has the following values in memory: 2. 5.
1. 4.
3. 6.
Whereas, the reader has the following values in its database: 1.
2.
3.
4. As a consequence, the synchronization between them is failed. Since, reader has IDSz, k1z, k2z in its database, and does not recognize both triple IDSx, k1x, k2x and IDSy, k1y, k2y. The tag is unable to establish an association with reader, the next time that they communicate. 4.2 Full-Disclosure Attack Here we establish another attack that leads to disclosure of all secret information on tag. In this attack we need observe several rounds of protocol. This attack works if the following condition is satisfied: n1, n2 mod 96 = 0. In this case because 0,0 0, n3, , all becomes Zero. ,
96 0
0 96 , 0
96
0
96
,
,
(5)
,
,
(6) (7)
Then
(8) (9)
Then
(10) (11) Then
(12) (13)
By using the equations (11) and (13) we have: (14) (15) (16)
Security Analysis of Ultra-lightweight Protocol for Low-Cost RFID Tags
241
Now, attacker observes the exchanged public messages, if two consecutive run of protocol satisfy equation (6), then attacker results that n1, n2 mod 96 = 0. Now, he/she finds k1, k2, by solving system of equations (5), (11). These values are k1n and k2n. 2
2
(17) ,
, n
(18) n
n
n+1
n+1
n+1
Now the attacker has the values of IDS , k1 , k2 , IDS , k1 , k2 . He/she continues the attack as follows. In next session, the tag sends IDS for reader and receives ‖ ‖ from it. Now, the attacker Using this messages and secret values that he/she gained, computes n1, n2 and lets them in D for calculating Y. It’s apparent that computing , n3, k1*, k2* values is easy. Now attacker constructs a system of 12 equation 12 unknown using values n1, Y and calculates ID. (19) (20) (21) (22) (23) (24) (25) (26) (27) (28) (29) (30)
242
M. Kianersi, M. Gardeshi, and H. Yousefi
5 Proposed Solutions In this section we propose efficient countermeasures for existence attacks: 5.1 A Countermeasure for De-synchronization Attack To address this vulnerability, we propose a simple solution. In our countermeasure, both old and new keys and IDS will be stored in the reader side as the tag side. In this case reader and tag have at least one common triple (IDS, k1, k2) to authenticate each other. We launch the same attack as discussed above on this extended protocol at follows. If we suppose that the initial values of tag and reader are: Tag : Reader :
, ,
, ,
Step 1- The attacker interrupts message D, so the tag updates its values but the reader doesn’t: ,k ,k ,k
:
Tag
Reader :
, k ,k ,k
Step 2- The tag and the reader run protocol completely: Tag
:
Reader :
, , , ,
k k k k
,k ,k ,k ,k
Step 3- The attacker and the tag negotiate together: Tag
:
Reader :
, k ,k , k ,k
, k ,k ,k ,k
In this case even though the attacker has successfully completed all the steps, the tags and database are still synchronized since valid communication can take place using the old values of the keys. 5.2 A Countermeasure for Full Disclosure Attack MixBits function in SSL-MAP guarantees that if both of its two inputs are zeros mod 96, its output will be zero mod 96. Hence modifying MixBits function [3] as shown in below guarantee that in case of its two inputs is zeros mod 96; its output will not be zero mod 96. Then this modification will enhance the security of the protocol. The extra countermeasures are modifying the structure of some messages or internal states as follows:
Security Analysis of Ultra-lightweight Protocol for Low-Cost RFID Tags
243
Z=MixBits(X,Y) Z=x for (i=1:i ::= author will deny [ if [ ] [ with absence ] ] ::= | | | | ::= | , ::= | | | | ::= author says ::= author grants to <subject> during ::= below( , ) | separate( , ) | during( , ) | starts( , ) | finishes( , ) | before( , ) | overlap( , ) | meets( , ) | equal( , ) <subject> ::= <subject-constant> | <subject-variable> ::= | ::= | ::= author says that <subject> can use the during ::= author asks does <subject> have <privilege> rights to <xpath-statement> during ::= author creates ::= role( , <sign>, <xpath-statement>, <privilege> ) <sign> ::= + | <xpath-statement> ::= in <document-name>, return <xpath-expressions> <document-name> ::= <document-name-constant> | <document-name-variable> <xpath-expressions> ::= <xpath-node> | <xpath-node>, <xpath-expressions> <xpath-node> ::= [ / ] <node-name> [<xpath-predicate>] / <node-name> ::= <node-name-constant> | <node-name-variable> | * | // <xpath-predicate> ::= <predicate-relationship> | <predicate-relationship> | <xpath-axis-expression> ::= | ::= | <predicate-relationship> ::= < | > | = <privilege> ::= read | write
An Implementation of Axml(T ) : An Answer Set Programming
381
to XML documents. Rules are what provide the non-deterministic nature of Axml(T ) . They are conditional statements that consist of a head statement and body statements where the head statement is validated true if the body statements are found to be true as well. Rules are reasoned upon primarily to determine when a subject is allowed to access a particular XML document. Facts are additional information such as , which define role or temporal interval relations, or , which specify that a subject be granted membership to a role. We define a policy base as follows: Definition 1. A policy base P B is a finite set of facts and rules defining the access control rights that subjects have over XML objects in a database. Subjects, roles, XML objects, temporal intervals and all the relationships that exist between them exist within a domain D and can be represented in P B using the formal language Axml(T ) . Most of the syntax and semantic elements of the language are quite straightforward, however, a comprehensive explanation is available in [11]. As mentioned, Axml(T ) is based on the Role-based access control model and utilises temporal logic. Our formal language also uses the XML query language XPath for object specification [13]. We will briefly discuss these three areas of Axml(T ) . The Role-based Access Control model [6] defines the structure and relationships of authorisations we use. This primarily means rather than applying authorisations directly to subjects, we create roles that can have one or more specified authorisations. Each role contains a privilege, either positively or negatively granted, and an object that is being authorised access to. We grant subjects membership to these roles to allow them the privileges specified. Temporal interval reasoning is incorporated in our language so that we can specify temporal constraints on authorisations. We utilise Allen’s Temporal Interval algebra in Axml(T ) [1]. By doing so, we can define an interval and any relationships it may have with any other intervals directly in the policy base. With the existence of an interval, we can then use it in a to specify when that authorisation is applied. As mentioned earlier, we incorporate the XML query language, XPath, for specifying objects in roles. An XPath is a string representation of traversing through an XML document to return one or more elements from within it. However, the XPath query language includes more expressive power than just this. It also allows for predicate queries to specify an even more expansive set of XML document elements [13]. The following is an example of a security policy base demonstrating expressions in Axml(T ) utilising some of the mentioned features. Example 1. Axml(T ) expressions author creates role(graphicDesigner, +, /graphics/photographs/, read). author creates role(graphicDesigner, +, /graphics/photographs/, write). author creates role(graphicDesigner, +, /graphics/illustrations/paint @year=‘2011’], write). author creates role(graphicDesigner, -, /graphics/clipart//cartoons, read). author creates role(photographer, +, /graphics/photographs/, read).
author author author author author author if
creates role(photographer, +, /graphics/photographs/, write). says separate(graphicDesigner, photographer). says before(photoShoots, designMeetings). says during(presentations, designMeetings). grants graphicDesigner to Jemma during designMeetings. grants photographer to Bob during photoShoots author grants graphicDesigner to Jemma during designMeetings.
After defining a security policy base with Axml(T ) , the next step is to reason upon and query what authorisations it allows. However, to do so we must first
382
S. Policarpio and Y. Bai
translate it into a logic program. The semantic translation of Axml(T ) is called ALP [11]. The ALP program is computed using an Answer Set solver which produces a stable model (an answer set) containing authorisation statements. These statements specify which subjects have access to what and when and are produced due to the fact that the subject has been granted membership to a role and his membership does not conflict with any other rules or principles of the Role-based Access Control model.
3
Implementation
This implementation will test the true expressiveness, capability, and limitations of Axml(T ) in a software application by 1) defining an Axml(T ) policy base, and 2) translating the policy base and reasoning upon it to discover its authorisations. Although we are utilising the formal language to protect some arbitrary XML documents, the implementation will however not actually be restricting user access to those documents. At this point, we are mainly interested in seeing if the Axml(T ) authorisation model is feasible. The system will allow us to define a policy base over some XML documents so that we can query it to see if the system properly reasons the authorisations it should allow. This will also help us find new aspects of the language we should add or fix. Also, we will be able to see if those principles we inherited from the Role-based Access Control model are working properly. For the structure of the system, we designed a management module called pb mgr (policy base manager) that contains a majority of the functionality required. For ease of use, we incorporate a web-based user interface to execute the module. Besides the functionality already mentioned, we must also note the inclusion of an ASP solver and XML Documents database. The Answer Set Program solver represents the software tools we will use to ground1 the variables in the translated policy base and also compute a stable model (answer) from it. With respect to an XML database, for the sake of simplicity, we stored the XML documents in a local directory rather than a sophisticated XML storage system. This allows us to create and utilise simple programming structures to retrieve XML documents. The policy base manager pb mgr is written in the Python scripting language while the web interface was written in PHP. For testing purposes, we setup a local Linux server with the Apache HTTP Server Project (httpd 2.2.15), PHP (5.3.2), and Python (2.6.5). We discuss the important functions and design aspects we created for the pb mgr object class and PHP interface. Axml(T ) Policy Base Management. We utilised XML for the storage and retrieval of Axml(T ) rules and facts. We created a document structure using the XML Schema specification from the W3C [14]. The schema allows us to define a static structure that the policy base must adhere to. This was important because it ensured that any rules or facts that were inserted into the policy base were valid 1
A procedure that ensures we have a variable-free logic program.
An Implementation of Axml(T ) : An Answer Set Programming
383
because we wrote the XSD with respect to Axml(T ) ’s syntax. For the addition of rules or facts to the policy base file, we first parse the document using the Python Lightweight Document Object Model (minidom) library. This turns the XML document into an object that we can easily manipulate in Python. Rules and facts are created and removed using the web interface. For creation, the user has the choice of selecting from a set of Axml(T ) statements. The interface temporarily saves all created rules and facts to an XML document, validates them (using the XSD), and then passes the document to pb mgr so that they may be copied into the parsed policy base object. Translating Axml(T ) to a Logic Program. With the existence of a security policy base, the manager can translate it into an ALP logic program. As we mentioned earlier, ALP is our Answer Set Programming semantic translation of Axml(T ) . We must translate a policy base so that we may reason upon it to find valid authorisations from it. ALP translation consists of first parsing our XML policy base document and then sequentially translating each rule and fact. Utilising various string manipulation techniques and regular expressions, we perform a translation on the policy base and export each rule and fact to a Python list data structure. Algorithm 1 exhibits this process2 . Algorithm 1. Policy Base Translation Data: policy base pb, ALP policy base alpPb 1 ppb = minidom.parse(pb); 2 alpRules = list(); 3 foreach rule ∈ ppb do // translate each rules head and body to ALP 4 alpHead = alpTranslate(rule.head); 5 alpBody = alpTranslate(rule.body); // perform XPath rewriting if necessary 6 if alpHead contains a dynamic XPath then 7 expandedAlpHead = expandXPaths(alpHead); 8 expansionH = true; 9 end 10 if alpBody contains a dynamic XPath then 11 expandedAlpBody = expandXPaths(alpBody); 12 expansionB = true; 13 end // append head(s) and body(ies) together 14 if expansionH AND expansionB == true then 15 alpRules.append(combine(expandedAlpHead, expandedAlpBody)); 16 else if expansionH == true then 17 alpRules.append(combine(expandedAlpHead, alpBody)); 18 else if expansionB == true then 19 alpRules.append(combine(alpHead, expandedAlpBody)); 20 else 21 alpRules.append(combine(alpHead, alpBody)); 22 end 23 end // extract list of constants for explicit definition 24 cons = list(); 25 foreach alpRule ∈ alpRules do 26 cons.append(extractConstants(alpRule)); 27 alpPb.append(alpRule); 28 end 29 alpPb.append(cons); 30 return alpPb;
Computing Authorisations and Querying Them. For the computation of the policy base logic program, we utilise the Potassco tools set developed by 2
The methods shown for XPath expansion are explained in [10].
384
S. Policarpio and Y. Bai
the University of Potsdam [12]. These represent the ASP solver we mentioned earlier. From within the policy base manager class, we create a sub-process to execute gringo, Potassco’s grounding program that will take care of the variables in the policy base. We output a grounded version of our logic program and save it temporarily to file. Next, we create another sub-process to execute clasp, Potassco’s answer set solver, with the grounded logic program. The output from the answer set solver contains the authorisations which we can query to discover what XML objects subjects have access to. A query is created on the interface and submitted to the manager for comparison. If a match in the answer set is found for the query, the manager will indicate that the subject does in fact have the privilege to do some action on the XML document during the specified interval. Outcomes. We tested the prototype with various policy base scenarios to ensure that the formal language is working properly. Expectedly, Axml(T ) does perform the way we hoped in terms of the basic features and principles we incorporated into its formalisation and semantic translation. Although we are generally happy with the direction of the prototype, it still requires further and extensive testing to ensure that it meets the requirements of an access control model for XML. Presently, the source code, example XML document, and other files are available at http://www.scm.uws.edu.au/∼spolicar/ for the purpose of demonstration or examination. Naturally, there was some challenges we encountered during development that needed to be alleviated immediately. Due to space constraints, we cannot discuss these challenges in depth. They can be found in full detail in [10].
4
Case Study
In a health and well-being store, nutritional information about particular foods are stored in an XML database so that employees of the store can access and provide the data to customers. Users also access the document to reference and update the information for the purpose of creating meals that utilise the recorded foods. The XML was originally authored and located at www.alistapart.com/articles/usingxml/. For this case study we have four different roles with varying privileges over the document. Their access rights to specific XPath’s are shown in the following table: Role name Privilege XML Object generaluser +read /nutrition cookingstaff +read /nutrition/food/name nutritionist +write /nutrition/dailyvalues nutritionist +write /nutrition/food -write /nutrition/dailyvalues assistant
For the scenario there are 5 subjects (Joel, Sarah, Sinead, Patrick, Sean) assigned to the various roles and temporal intervals. We will also specify the following rules:
An Implementation of Axml(T ) : An Answer Set Programming – – – –
385
All nutritionists must not be allowed to work two days in succession. If an assistant is allowed access on a day, they must have or be allowed access the following day. If Sarah has the nutritionist role on any day, then Joel must have it the following day. The cooking staff are not allowed to access the document on Monday and Friday.
The following is the Axml(T ) policy base containing these rules and roles as well as the grant statements for the subjects. This policy base was generated by using the prototypes PHP interface to enter each rule individually. Normally, the policy base would be saved in the XML schema we specified, but we present it here in normal Axml(T ) syntax for the purpose of readability.
author creates role(generaluser, +, in nutrition, return /nutrition, read). author creates role(cookingstaff, +, in nutrition, return /nutrition/food/name, read). author creates role(nutritionist, +, in nutrition, return /nutrition/dailyvalues, write). author creates role(nutritionist, +, in nutrition, return /nutrition/food, write). author creates role(assistant, -, in nutrition, return /nutrition/dailyvalues, write). author says below(nutritionist, generaluser). author says below(assistant, nutritionist). author will deny if author grants nutritionist to S during T1, author grants nutritionist to S during T2, author says before(T1, T2). author grants assistant to S during T2 if author grants assistant to S during T1, author says before(T1, T2). author grants nutritionist to joel during T2 if
4.1
author grants nutritionist to sarah during T1, author says before(T1, T2). author will deny if author grants cookingstaff to S during monday. author will deny if author grants cookingstaff to S during friday. author says meets(monday, tuesday). author says meets(tuesday, wednesday). author says meets(wednesday, thursday). author says meets(thursday, friday). author grants cookingstaff to sinead during friday. author grants nutritionist to joel during tuesday. author grants nutritionist to joel during wednesday. author grants cookingstaff to patrick during thursday. author grants nutritionist to sarah during thursday. author grants assistant to sean during wednesday. author grants assistant to sean during thursday. author grants assistant to patrick during thursday.
Translating and Computing the Authorisations
Executing the translation functions of the implementation produces a logic program. However, when reasoning upon that logic program, the prototype did not produce an answer set. This was expected as we had purposely included grant statements in our policy base that would fail with respect to the rules we specified. We highlight those failed grant statements and provide a solution to alleviate the conflict. 1. Sinead is granted membership to the cooking staff role on Friday which is an off limits day. We will change her grant to Wednesday. 2. Joel is attempting to use the nutritionist role for two days in a row. We will change his first grant to occur on Monday instead of Tuesday so that a gap is provided between his access to the role on Wednesday.
The rest of the grant statements validate with the rules. With the policy base and translated logic program reflecting the changes listed above, the prototype generated an answer set. We are primarily concerned with the following correct authorisations (auth statements) excerpted from that answer set:
386
S. Policarpio and Y. Bai
auth(sarah,nutrition, nutrition dailyvalues,write,thursday) auth(sarah,nutrition, nutrition food,write,thursday) auth(joel,nutrition, nutrition dailyvalues,write,monday) auth(joel,nutrition, nutrition food,write,monday) auth(sinead,nutrition, nutrition food name,read,wednesday) auth(sinead,nutrition, nutrition food name,read,wednesday) auth(joel,nutrition, nutrition dailyvalues,write,friday) auth(joel,nutrition, nutrition food,write,friday) auth(joel,nutrition, nutrition dailyvalues,write,wednesday) auth(joel,nutrition, nutrition food,write,wednesday) auth(patrick,nutrition, nutrition food name,read,thursday)
5
auth(patrick,nutrition, nutrition food name,read,thursday) auth(sarah,nutrition, nutrition,read,thursday) auth(sean,nutrition, nutrition,read,friday) auth(sean,nutrition, nutrition food,write,friday) auth(sean,nutrition, nutrition,read,wednesday) auth(sean,nutrition, nutrition food,write,wednesday) auth(sean,nutrition, nutrition,read,thursday) auth(sean,nutrition, nutrition food,write,thursday) auth(joel,nutrition, nutrition,read,monday) auth(joel,nutrition, nutrition,read,friday) auth(joel,nutrition, nutrition,read,wednesday)
Experimental Results
In [10], we examined various policy bases and their effects on computation time with the Axml(T ) system implementation. Unfortunately, this examination was quite in depth (31 experiments) and we can not display all of the details here. In summary, these experiments were concerned with the following: 1. 2. 3. 4.
Computation time as the size of the policy base increases uniformly (without complex rules) Evaluating computation time as the complexity of Axml(T ) /ALP rules increases Effects of computation time as the number of XML documents requiring authorisation increases The effects of the number of temporal intervals and their complex relationships on computation time 5. Examining larger scale policy bases with complex rules against computation time and space
From the results of these experiments we found that the computation times for those grounded and translated policy bases which were under approximately one million rules are relatively acceptable from an implementation point of view. We also discovered that careful consideration must be taken for the number of temporal intervals and relationships specified in the policy base due to the increase in reasoning with their presence. Oppositely, we learned that the number of XML documents specified in the policy base is not much of a concern. In terms of the implementation, we also found that with larger scale policy bases, the output generated from computation became increasingly larger with each experiment. We now know that large inputs can take a toll on our system. We saw a drastic hinder in performance with Experiment 31 (the largest and most complex policy base in our test set) when the implementation halted due to a lack of system memory. Increasing system memory could be a short term solution, but from a software engineering point of view, is not the definitive one. At this point, we believe best practice would be to consider limiting the size of the authored policy bases. From our concluding results, we saw that the produced answerset data remained under one megabyte of data; an amount that is still easily searchable. Therefore, although we see a difficulty with the computation of larger scale and complex policy bases, we still deem the system feasible and desirable due to the fact that performing searches is relatively easy.
6
Conclusion
We have presented a prototype for our formal language of authorisation for XML documents with temporal constraints. After a brief overview of the Axml(T ) language, we discussed the features, structure, technical specifications, and algorithmic details of the implementation. Using a simple case study, we further
An Implementation of Axml(T ) : An Answer Set Programming
387
explained how the Axml(T ) implementation works. Finally, we presented some summarised results from experiments we ran to test the limitations of our prototype. With this first phase of the prototype complete, it would be beneficial in the future to extend it with other features such as implementing user authentication or linkage with a current XML database management system.
References 1. Allen, J.F.: Towards a general theory of action and time. Artif. Intell. 23(2), 123–154 (1984) 2. Anutariya, C., Chatvichienchai, S., Iwaihara, M., Wuwongse, V., Kambayashi, Y.: A rule-based xml access control model. In: RuleML, pp. 35–48 (2003) 3. Baral, C.: Knowledge Representation, Reasoning and Declarative Problem Solving. Cambridge University Press, Cambridge (2003) 4. Bertino, E., Carminati, B., Ferrari, E.: Access control for xml documents and data. Information Security Technical Report 9(3), 19–34 (2004) 5. Damiani, E., Vimercati, S.D.C.d., Paraboschi, S., Samarati, P.: A fine-grained access control system for xml documents. ACM Trans. Inf. Syst. Secur. 5(2), 169–202 (2002) 6. Ferraiolo, D.F., Cugini, J.A., Richard Kuhn, D.: Role-based access control (rbac): Features and motivations. In: 11th Annual Computer Security Applications Proceedings (1995) 7. Gelfond, M., Lifschitz, V.: The stable model semantics for logic programming. In: Kowalski, R.A., Bowen, K. (eds.) Proceedings of the Fifth International Conference on Logic Programming, pp. 1070–1080. The MIT Press, Cambridge (1988) 8. He, H., Wong, R.K.: A role-based access control model for xml repositories. In: WISE 2000: Proceedings of the First International Conference on Web Information Systems Engineering, vol. 1, p. 138. IEEE Computer Society, Washington, DC, USA (2000) 9. Lifschitz, V.: What is answer set programming? In: AAAI 2008: Proceedings of the 23rd national conference on Artificial intelligence, pp. 1594–1597. AAAI Press, Menlo Park (2008) 10. Policarpio, S.: An Answer Set Programming Based Formal Language for Complex XML Authorisations with Temporal Constraints. PhD thesis in Computer Science, University of Western Sydney (2011) 11. Policarpio, S., Zhang, Y.: A formal language for specifying complex XML authorisations with temporal constraints. In: Bao, F., Yung, M., Lin, D., Jing, J. (eds.) Inscrypt 2009. LNCS, vol. 6151, pp. 443–457. Springer, Heidelberg (2010) 12. University of Potsdam. Potassco, the potsdam answer set solving collection (2010), http://potassco.sourceforge.net/ 13. WWW Consortium, Xml path language (xpath) version 1.0. (1999), http://www.w3.org/TR/xpath 14. WWW Consortium, W3c xml schema (2004), http://www.w3.org/XML/Schema.html 15. WWW Consortium, Extensible markup language (xml) 1.0., 5thedn. (November 2008), http://www.w3.org/TR/REC-xml/
On Cloud Computing Security Yun Bai and Sean Policarpio Intelligent Systems Laboratory School of Computing and Mathematics University of Western Sydney Locked Bag 1797, Penrith, NSW 2751 Australia {ybai,spolicar}@scm.uws.edu.au
Abstract. Could computing is the latest development of the modern computing technology. It is the next stage of the Internet evolution. Cloud computing provides the organizations with the infrastructure management, various software services and the datacenter maintenance. The organizations can reduce their operational cost and concentrate on their strategic planning by using services provided by cloud computing. Generally, the success of cloud computing depends on three key issues: data security, fast Internet access and standardization [17]. Among the three issues, the biggest concern is data security. In this paper, we investigate the security issue related to datacenter of cloud computing. By analyzing the properties of the data stored at the datacenter, we propose a logical approach to specify the data and employ intelligent agents to enforce appropriate security policies on it. We expect such approach will protect the datacenter by only allowing the legitimate users accessing the data and preventing any malicious attempt to it. Keywords: Access Control, Security Model, Cloud Computing, Formal Specification.
1 Introduction Nowadays, Internet is used everywhere from personal leisure browsing to business trading, from our everyday routines such as banking, shopping to high technology applications such as spacecraft launching, satellite controlling. Cloud computing is the current development of the Internet and the modern computing technology. It is the next stage of the Internet evolution. With more and more organizations dealing with increasing amount of data relying on the Internet, cloud computing plays increasingly important role. Cloud computing is a structure that allows an organization to access applications, services or data that resides in a remote site or datacenter. It provides these organizations with the infrastructure management, software development and the datacenter maintenance. Hence, the organizations don’t need to spend much effort on investment and maintenance of the infrastructures which keep their datacenter running. They can instead concentrate on their strategic projects planning and development, and can reduce their operational and capital expenditure. ¨ A. Ozcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 388–396, 2011. c Springer-Verlag Berlin Heidelberg 2011
On Cloud Computing Security
389
Apart from datacenter maintenance and management, cloud computing also provides a variety of services to help the organizations to reduce their operational cost and to increase their productivity. Some major organizations such as Amazon, Google, Microsoft now offer cloud services to the public. Amazon offers virtual machines and extra CPU cycles as well as storage service. Google offers database, online documents, and other online software. Microsoft provides the organizations with the Window applications, datacenters, database services [19]. When increasing organizations enjoy the convenience, efficiency and increased productivity by using cloud computing, at the meantime, they also need to face a major challenge comes with it: the security. The introduction of the cloud computing brings major security concern about the organization’s data hosted in a non-local datacenter. Overall, the success of cloud computing depends on three key issues: data security, fast Internet access and standardization. Among the three issues, the biggest concern is data security. In this paper, we address the security issue of accessing the datacenter. We propose an approach using authentication and access control to ensure secure access to the data. The aim of the approach is to only allow the legitimate users accessing the organization’s data, to prevent any unauthorized attempt to the datacenter. Authorization or access control is a mechanism to ensure that all accesses to the system resources occur exclusively according to the access polices and rules specified by the security agent of the system. Authorizations or access control has been extensively studied in [1], [4], [10], [22] etc. and a variety of authorization specification approaches such as access matrix [6], [8], role-based access control [5], access control in database systems [3], authorization delegation [14], procedural and logical specifications [2] have been investigated. Since logic based approaches provide a powerful expressiveness [9] as well as flexibility for capturing a variety of system security requirements, increasing work has been focusing on this aspect. However, how these approaches can apply to cloud computing environment has not been explored much. [11] presented a design for a file system to secure file system storage service for Web 2.0 application. [15] proposed approaches to provide catch based security and performance isolation for cloud computing environment. In this paper, we try to incorporate formal approaches into cloud computing security. The paper is organized as follows: In section 2, we analyze the general features of the cloud computing including its architecture, various services it provides, the structure of the datacenter and the properties of the datacenter. Section 3 proposes a structure for secure datacenter with authentication and access control functions provided. In this section, the authentication mechanism is outlined and an authorization approach is also described for ensuring the datacenter security. Section 4 presents a detailed authorization mechanism by using a formal logical specification. The access control evaluation is discussed. Section 5 concludes the paper with some future work outlined.
2 Properties of Cloud Computing In a cloud computing structure, the organizations using the cloud services are normally referred to as clients. These clients can be located geographically differently. They
390
Y. Bai and S. Policarpio
access services provided by cloud computing via the Internet by distributed servers. The actual servers or the datacenter can be located geographically differently as well. Cloud computing provides various services to the clients. Software as a service (SaaS) is a software distribution model in which software application is hosted by a software provider to the clients. The clients, instead of purchasing and installing the software to their computer systems, they may just access such software via the Internet as a service. This kind of service is also called ’pay as you go’, the clients only pay the software services they actually use. With the development of the service oriented architecture (SoA) which supports the SaaS model and the increasing availability of the Internet access, SaaS is becoming a prevalent software service distribution model and accepted by ever increasing clients around the world. With SaaS model, the software is hosted off site by the provider, the clients do not need to develop, maintain or upgrade it. They benefit from using low-cost, more reliable software. They may just concentrate on their internal strategic planning for increasing their productivity. Platform as a service (PaaS) is another application distribution model. In this model, the PaaS provider supplies all the infrastructure needed to build application and service over the Internet. PaaS can provide the clients with application design, development, testing, etc. by an efficient, cost-effective delivery model. The clients do not need to download and install the required software to build and maintain the infrastructure. They may direct this cost to other part of their operation to ensure their overall productivity. Datacenter is an important service provided by cloud computing. It is a collection of data and servers the clients subscribe. As an organization using cloud computing service, its data and application are located on servers geographically different from its local site. As discussed previously, when a client’s data is housed in a datacenter, it saves the cost of maintaining it. When client from different location needs to access its data, the data seems to be just located locally. However, since the data is located in a non local site, ensuring the security of the data is not as simple as if housed locally. It is not feasible to enforce the same local security measurement. To ensure safe access to the datacenter, it needs a coordinated security measurement between the datacenters and the clients accessing the datacenter.
3 Approaches towards Securing Datacenter Generally there are two steps to ensure the security of a system such as a datacenter hosted by cloud computing: authentication and authorization. Authentication controls which user can access the datacenter. With authentication mechanism, only the legitimate users are allowed to enter the datacenter. On the other hand, authorization controls that the legitimate user only performs legitimate operations on the data once it has been successfully authenticated. The two mechanisms work together to effectively provide the datacenter with secure accesses as Figure 1 shows. Where C1 , C2 , ...Cn represent the client servers, each of them acts on behalf of a group of users who need to access the datacenter. AeS is the authentication server and AoS acts as the authorization server. AeS manages a database in which all the
On Cloud Computing Security
391
data center
authorization
authentication
C1
C2
...
Cn
Fig. 1. Secure Datacenter Structure
users information such as password, identification number etc. are registered. It ensures that only the registered, legitimate users are allowed to access the datacenter. AoS also manages a similar database for every user about its access right to the datacenter. It controls that the authenticated users only perform legitimate operations on the data of the datacenter. When a user needs to access the datacenter, it requests to its client server first, then the client server passes the information to AeS. AeS then checks its database about this user to either grant or deny such a request. If the authentication is successful, it passes on to the AoS. AoS checks the user request with its database about this user’s access right. If the user request is within its specified right, the request is granted, otherwise, it is denied. For our authentication mechanism, we employ a Kerberos [18] like system to fulfill the function. To simplify the description, we only illustrate the authentication process for one client server CS. All the rest client servers follow the same procedure for their user authentication. Figure 2 shows the authentication mechanism for a client server CS. The authentication process works as follows: suppose AeS and AoS share a secret key for encrypting and decrypting the messages between them and AeS has already have a database in which the user’s password, identification and other information are stored. When a user U1 enters the system and requests access to the datacenter, it needs to be authenticated by AeS first. To begin with, U1 sends its request to CS in step 1; on behalf of U1 , the CS sends the request to AeS in step 2; AeS checks its database to confirm if U1 has supplied the correct password and other required information and if U1 is permitted access to the datacenter. If all are confirmed correctly, then U1 is an authentic user and be issued a ticket encrypted by the secret key shared by AeS and
392
Y. Bai and S. Policarpio
AeS 2 3 CS 1
4
5
U1
AoS
Fig. 2. Authentication Structure
AoS in steps 3 and 4. U1 cannot alter the ticket since it does not possess the secret key. It can only pass it to AoS for requesting access to the datacenter in step 5. Similar to AeS, AoS manages a database as well about the users entitled access rights to the datacenter. It compares the user’s request to its database record about what operations the user can perform. If the request is within the specification of U1 ’s entitled rights, then U1 can access the datacenter as it requested. Otherwise the request is denied. The detailed specification, function and evaluation of the AoS will be presented in the next section.
4 A Formal Method for AoS In the system described above, each client server provides service for a group of users who access the datacenter generally located at a remote site. These client servers may locate at different sites and the user groups they represented may have different access requests to the datacenter. We assume that the Internet access on which the system relies is safe and sound. For the authorization mechanism, each agent manages one client server; to have the whole system coordinated, all the agents are managed by a super agent. In this section, we concentrate on the investigation of a single agent by proposing a logic model for its specification and evaluation. All the other agents follow the same model. 4.1 The Language for AoS Specification We introduce a formal model for representing AoS security rules based on a first order language. We give both syntactic and semantic descriptions for our policy base model. Let L be a sorted first order language with equality, with three disjoint sorts for legitimate object, legitimate user and group legitimate user respectively. Assume L has the following vocabulary:
On Cloud Computing Security
393
1. Sort legitimate object: with object constants O, O1 , O2 , · · ·, and object variables o, o1 , o2 , · · ·. 2. Sort legitimate user: with user constants U, U1 , U2 , · · ·, and user variables u, u1 , u2 , · · ·. 3. Sort group legitimate user: with group user constants GU, GU1 , GU2 , · · ·, and user variables gu, gu1 , gu2 , · · ·. 4. A binary predicate symbol request which takes arguments as legitimate user or group legitimate user and legitimate object respectively. 5. A binary predicate symbol can which takes arguments as legitimate user or group legitimate user and legitimate object respectively. 6. A binary predicate symbol ∈ which takes arguments as legitimate user and group legitimate user respectively. 7. A binary predicate symbol ⊆ whose both arguments are group legitimate user. 8. Logical connectives and punctuations: as usual, including equality. In this specification, a legitimate user U can access legitimate object O of the datacenter is represented by a ground formula can(U, O). A ground formula is a formula without any variables. A user U requests access to object O of the datacenter is represented by a ground formula request(U, O). The group membership is represented as follows: for example, let GU be a group constant representing a specific group users called expert. “U is an expert” means U is a member of the group GU , this can be represented using the formula U ∈ GU . We can also represent inclusion relationships between different user groups such as GU1 ⊆ GU2 . Furthermore, we can represent constraints among users’ authorizations. For example, the rule stating that “a technician can access an object file O”, “Sue is a technician” can be represented as follows. We use T to represent the group technician. ∀u.u ∈ T ⊃ can(u, O),
(1)
Sue ∈ T,
(2)
Usually we define that if a group entitles access to certain object, then all the members of the group can access the same object unless otherwise specified. This is called the inheritance property of authorizations. This can be represented as: ∀u.u ∈ GU ∧ can(GU, O) ⊃ can(u, O).
(3)
Where u represents any member of the group GU and O is a data object of the datacenter that GU can access. 4.2 The Security Rule Base Specification and Evaluation We can now give a formal definition of the security rule base of the AoS by using the language L. Definition 1. A security rule base SRB is a quaternary of (LU, LO, F, C) where LU is a finite set of legitimate users; LO is a finite set of legitimate data objects of the datecenter; F is a finite set of ground literals and C is a finite set of closed first order formulas.
394
Y. Bai and S. Policarpio
A formula without any free variables is defined to be a closed formula. In our formalism, both facts and rule constraints are represented by closed formulas of L. For example, can(Sue, T ) is a fact, and so is (2). can(Sue, O) ⊃ can(Amy, O) is a rule constraint, (1) and (3) are also rule constraints. These rule constraints are viewed as access constraints which should be always satisfied. We refer to a fact as a ground formula, a ground literal, or an atom. A model of a security rule base is the assignment of a truth value to every formula of the security rule base in such a way that all formulas of the security rule base are satisfied [7]. Formally, we give the following definition. Definition 2. A model of a security rule base SRB = (LU, LO, F, C) is defined to be a Herbrand model [7] of LU ∪ LO ∪ F ∪ C. SRB is said to be consistent if there exists some model of SRB. The set of all models of SRB is denoted as M odels(SRB). A formula ψ is a consequence of SRB, denoted as SRB |= ψ, if LU ∪ LO ∪ F ∪ C |= ψ. In this case, we also say ψ is satisfied in SRB. We use an example to show how the security rule base works. Example 1. Consider a situation where the facts are: U1 and U2 are technicians(T ), they both are legitimate users; and technician can access a data record(D-records) of the datacenter where D-records is a legitimate data record. The constraint states that if someone belongs to a group then he(she) inherits the group’s access rights. In our security rule base, this situation can be specified as SRB = (LU, LO, F, C), where LU = {U1 ∈ LU, U2 ∈ LU }, LO = {D-records ∈ LO}, F = {U1 ∈ T, U2 ∈ T, can(T, D-records)}, and C = {∀ug.u ∈ g ∧ can(g, o) ⊃ can(u, o)}. It is not difficult to see that facts can(U1 , D-records) and can(U2 , D-records) are consequences of SRB, and SRB has a unique model m where: m = {U1 ∈ LU, U2 ∈ LU, P -records ∈ LO, U1 ∈ T, U2 ∈ T, can(T, P -records), can(U1 , D-records), can(U2 , D-records)}. Example 2. Consider another situation where the facts are: U1 and U2 both are legitimate users, D-records is a legitimate data record. U1 can access D-records, U2 also can access D-records. The constraint is that D-records can only be accessed by one user at one time. Formally, this can be represented as SRB = (LU, LO, F, C), where LU = {U1 ∈ LU, U2 ∈ LU }, LO = {D-records ∈ LO}, F = {can(U1 , D-records), can(U2 , D-records)}, and C = {can(U1 , D-records) ⊃ ¬can(U2 , D-records)}. This security rule base is not consistent as there does not exist a model for it.
On Cloud Computing Security
395
Now we discuss the SRB evaluation. When a user U requests access to D of the datacenter, the task of the AoS is to evaluate such a request and make the decision to either grant or deny the request. For a request request(U, D), Generally, the AoS will first check its corresponding SRB to find out if U and D are legitimate user and data object or not. If yes, it then checks the facts of the SRB, if can(U, D) is presented, request(U, D) then is explicitly granted. Otherwise, it does reasoning about the related facts and rules, calculates the model of the SRB. If can(U, D) is in the model, then can(U, D) can be deduced, hence the request is implicitly granted; otherwise, the request is denied. Definition 3. For an access request request(U, D), the AoS evaluates the SRB = (LU, LO, F, C) by calculating its model m. If can(U, D) ∈ m, or SRB |= can(U, D), request(U, D) is to be granted; otherwise, it is to be denied. Example 3. The SRB is as described as in Example 1. In addition, U3 is also a legitimate user. The access requests are: request(U1 , D-records) and request(U3 , D-records). In this case, the SRB = (LU, LO, F, C), where LU = {U1 ∈ LU, U2 ∈ LU, U3 ∈ LU }, LO = {D-records ∈ LO}, F = {U1 ∈ T, U2 ∈ T, can(T, D-records)}, and C = {∀ug.u ∈ g ∧ can(g, o) ⊃ can(u, o)}. Again, the a unique model m is: m = {U1 ∈ LU, U2 ∈ LU, U3 ∈ LU, P -records ∈ LO, U1 ∈ T, U2 ∈ T, can(T, P -records), can(U1 , D-records), can(U2 , D-records)}. Obviously, SRB |= can(U1 , D-records), request(U1 , D-records) is granted; SRB |= can(U3 , D-records) does not hold, so request(U3 , D-records) is denied.
5 Conclusions In cloud computing environment, since the datacenter normally located in a remote site, it poses a great security concern of data hosted in the datacenter. In this paper, we have examined the security issue of the datacenter in cloud computing environment. We believe both authentication and authorization machenisms are essential to protect the data in the datacenter. We proposed a structure to perform the user authentication and to control their accesses to the datacenter in order to protect the datacenter from malicious attempt. We have sketched a framework for the authentication process and introduced a detailed formal approach for the access control mechanism. We investigated a logic approach for representing authorization rules and evaluating user’s access request. The implementation issue will be considered in our future work. Also more detailed access rights, different operations on data object needs to be investigated. This shall be part of the future work as well.
396
Y. Bai and S. Policarpio
References 1. Atluri, V., Gal, A.: An authorization model for temporal and derived data: securing information protals. ACM Transactions on Information and System Security 5(1), 62–94 (2002) 2. Bertino, E., Catania, B., Ferrari, E., Perlasca, P.: A logical framework for reasoning about access control models. ACM Transactions on Information and System Security 6(1), 71–127 (2003) 3. Bertino, E., Jajodia, S., Samarati, P.: Supporting multiple access control policies in database systems. In: Proceedings of IEEE Symposium on Research in Security and Privacy, pp. 94–107 (1996) 4. Chomicki, J., Lobo, J., Naqvi, S.: A logical programming approach to conflict resolution in policy management. In: Proceedings of International Conference on Principles of Knowledge Representation and Reasoning, pp. 121–132 (2000) 5. Crampton, J., Khambhammettu, H.: Delegation in role-based access control. International Journal of Information Security 7, 123–136 (2008) 6. Dacier, M., Deswarte, Y.: Privilege graph: an extension to the typed access matrix model. In: Proceedings of European Symposium on Research in Computer Security, pp. 319–334 (1994) 7. Das, S.K.: Deductive Databases and Logic Programming. Addison-Wesley Publishing Company, UK (1992) 8. Denning, D.E.: A lattice model of secure information flow. Communication of ACM 19, 236–243 (1976) 9. Fagin, R., Halpern, J.Y., Moses, Y., Vardi, M.Y.: Reasoning about knowledge. MIT Press, Cambridge (1995) 10. Fernandez, E.B., France, R.B., Wei, D.: A formal specification of an authorization model for object-oriented databases. In: Database Security, IX: Status and Prospects, pp. 95–109 (1995) 11. Hsu, F., Chen, H.: H Chen, Secure File System Services for Web 2.0 Application. In: ACM Cloud Computing Security Workshop, pp. 11–17 (2009) 12. Hurwitz, J., Bloor, R., Kaufman, M., Halper, F.: Cloud Computing for Dummies. Wiley Publishing Inc., Chichester (2010) 13. Jajodia, S., Samarati, P., Sapino, M.L., Subrahmanian, V.S.: Flexible support for multiple access control policies. ACM Transactions on Database Systems 29(2), 214–260 (2001) 14. Murray, T., Grove, D.: Non-delegatable authorities in capability systems. Journal of Computer Security 16, 743–759 (2008) 15. Raj, H., Nathuji, R., Singh, A.P.: England Resource management for Isolation Enhanced Cloud Services. In: ACM Cloud Computing Security Workshop, pp. 77–84 (2009) 16. Reiter, R.: A logic for default reasoning. Artificial Intelligence 13, 81–132 (1980) 17. Rittinghouse, J.w., Ransome, J.F.: Cloud Computing, Implementation, management, and Security. CRC Press, Boca Raton (2010) 18. Stallings, W.: Cryptography and Network Security - principles and Practice, 5th edn. Pearson, London (2006) 19. Velte, A.T., Velte, T.J., Elsenpeter, R.: Cloud Computing - A Practical Approach. McGraw Hill, New York (2010) 20. Winslett, M.: Updating Logical Databases. Cambridge University Press, New York (1990) 21. Woo, T.Y.C., Lam, S.S.: Authorization in distributed systems: A formal approach. In: Proceedings of IEEE Symposium on Research in Security and Privacy, pp. 33–50 (1992) 22. Zhou, J., Alves-Foss, J.: Security policy refinement and enforcement for the design of multilevel secure systems. Journal of Computer Security 16, 107–131 (2008)
PAPR Reduction in OFDM by Using Modernize SLM Technique Ashutosh K. Dubey1, Yogeshver Khandagre2, Ganesh Raj Kushwaha1, Khushboo Hemnani1, Ruby Tiwari2, and Nishant Shrivastava1 1 Dept. of Computer Science & Engineering Trinity Institute of Technology and Research Bhopal, India
[email protected],
[email protected],
[email protected],
[email protected] 2 Dept. of Electronics and Communication Trinity Institute of Technology and Research Bhopal, India
[email protected],
[email protected] Abstract. One major Disadvantage of OFDM is the high peak-to average Power ratio (PAPR). One investigated technique Selected Mapping (SLM) is PAPR reduction techniques for Orthogonal Frequency Division Multiplexing (OFDM). In this paper we proposed a Modernize SLM (MSLM) scheme to reduce the PAPR by using the complex signal separate into real & imaginary parts and individually phase sequence multiple real as well as imaginary part of complex signal then select minimum PAPR signal of real & imaginary and these are combine. The simulation show achieves good PAPR., which is one of the strong candidate for Future wireless communication. Keywords: OFDM, PAPR, SLM and MSLM.
1 Introduction Recently orthogonal frequency division multiplexing (OFDM) [5] has been regarded and used as one of the technologies for the communication systems. Especially OFDM has been adopted for various wireless communication Systems [14] such as wireless local area networks (WLANs) [4], wireless metropolitan area networks (WMANs), digital audio broadcasting (DAB), and digital video broadcasting (DVB) [19]. OFDM is an attractive technique for achieving high data rate in the wireless communication systems and it is robust to the frequency selective fading channel [11]. However, an OFDM signal can have very high peak-to-average power ratio (PAPR) [9] at the transmitter. Which causes the signal distortion such as the in-band distortion and the out-of-band radiation [16] due to the nonlinearity of high power amplifier (HPA) [18], and induces the degradation of bit error rate (BER)? Thus, the A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 397–405, 2011. © Springer-Verlag Berlin Heidelberg 2011
398
A.K. Dubey et al.
Fig. 1. Linear and Nonlinear Region
PAPR reduction is one of the most important research interests for the OFDM systems. The Linear region and nonlinear region is shown in Fig 1. In Fig1.we want to avoid such undesirable nonlinear effects [18] a waveform with a high peak must be transmitted in the linear range of the HPA by decreasing the average power of the input signal. This is called input back-off (IBO) and results in a proportional output back-off (OBO) after the amplification. However high back-offs reduce the power efficiency of the amplifier and cause a reduction in the coverage range. The operating point of the nonlinearity is defined by the input back-off (IBO) that corresponds to the ratio between the saturated and the average input powers. To deal with the high PAPR a number of approaches have been proposed. Such as clipping [17], tone reservation [4], and selected mapping (SLM) [4]. All of the mentioned schemes need an accurate PAPR calculation based on over sampling. The remaining of this paper is organized as follows. We discuss System Description in Section 2. In Section 3 we discuss about PAPR Reductions. The Modernized SLM Scheme in section 4. The conclusions and future directions are given in Section 5. Finally references are given.
PAPR Reduction in OFDM by Using Modernize SLM Technique
399
2 System Description In an OFDM system, shown in fig.1 [5] data is modulated in the frequency domain to N adjacent subcarriers. These N subcarriers span a bandwidth of B Hz and are separated by a spacing of Δf =B/N .The continuous-time baseband representation of this is
Where T=1/ f is the symbol period. 2.1 Peak-to-Average Ratio The most popular quantification metric of envelope variation is the peak-to-average ratio (PAR). Rightfully so as PAR captures the most important aspect of a signal that has to pass through a peak-power limited device the peak power. The use of PAR in communications signals is a result of the use of PAR in radar applications. A radar system shares certain similarities with a communications system namely they both have to transmit an amplified radio signal of a certain spectrum. For radar the spectrum shape is often the only signal constraint which makes waveform shaping that minimizes peaks a relatively straightforward problem. However in an OFDM communication system there is the additional constraint that each subcarrier (Fourier coefficient of the spectrum) is modulated with an information bearing complex number. This additional degree of constraint significantly complicates the problem.
Fig. 2. OFDM Block Diagram
The PAR of an OFDM signal x according to Fig 2.
400
A.K. Dubey et al.
where x and be any signal representation (critically sampled baseband oversampled base-band continuous-time pass band etc.) defined over one symbol period. Because the denominator of an expected value and strictly speaking not an “average" it is true that the term PAR is a bit of a misnomer. Despite this slight technical inaccuracy PAR is the most widely used term and we will keep with convention here. Also note that the ensemble average power and the expectation in the denominator of only differ for non-constant modulus constellations. The PAR of x[n] is different for levels the PAR increases with N. First we assume that N the number of subcarriers is large enough so that the discrete-time domain signal has an approximate complex Gaussian distribution [5]. It then follows that the instantaneous power of the discrete time domain samples which is Chi-Squared distributed.
The IFFT each discrete time sample can be treated as independent of all other samples. With these two approximations the probability that the power of at least one x[n] out of N samples is above a given level is
Finally if E [|x[n] |2] is normalized to unity, then the CCDF of the PAR is
We know that OFDM is a promising high-speed communications technique however it suffers from high PARs.
3 PAPR Reductions Ideally a SLM scheme will create D independent mappings of a discrete-time domain signal xL. If we assume that each mapping is independent of all other mappings then the CCDF of the PAR in a SLM scheme is simply
Fig 4 is a plot of the theoretical PAR CCDF curves for a critically sampled OFDM symbol where N = 64, L=1and d=1,2,10,100. The CCDF of the PAR in a Nyquist sampled OFDM symbol is
PAPR Reduction in OFDM by Using Modernize SLM Technique
401
where N is the number of subcarriers and D is the number of independent phase mappings. SLM Complexity It is obvious that SLM has significant PAR reduction capabilities. However, this reduction is not free. The most significant costs are the D -1 additional IDFT operations and the D-1 N – length where N is the number of subcarriers and D is the number of independent phase mappings. These complexities can be mitigated slightly by using the inverse fast fourier transform (IFFT) in place of the IDFT and by using binary phase sequences so that all of the phase multiplications are just sign changes.
4 Modernized SLM Scheme Firstly real and imaginary part of the complex modulating signal Am is separated as
Fig. 3. Modernized SLM PAPR system
402
Where
A.K. Dubey et al.
is the real part and
is the imaginary part of the frequency
domain vectors and for both are generated. Since the signals are real valued. The phase vector DR and DI have to be real. These candidates are transformed into time domain using IFFTs. Then each combination of one real and one imaginary from these the best candidate with minimum PAR is selected. The CCDF of the PAR in Modernized SLM OFDM symbol is
Fig. 4. PAPR CCDF curves for N =128, L=1 and d=1, 5, 10, 50,100
PAPR Reduction in OFDM by Using Modernize SLM Technique
403
Fig. 5. Simulation results MSLM and SLM
In the modernized SLM system we separate the complex baseband signal into real and imaginary part. These real parts converted into serial to parallel and individual phase sequence is multiply into every parallel and selected minimum PAPR. Similarly imaginary parts converted into serial to parallel and individual phase sequence is multiply into every parallel and selected minimum PAPR. Both real and imaginary parts again combine and transmitted.
404
A.K. Dubey et al.
Fig. 6. Clipping in dB
5 Conclusions and Future Work In this paper we proposed the modernized SLM which has a lower PAPR than traditional SLM. We derived the PAR CCDF in a modernized SLM system. Its performance is analyzing in mat-lab 6.5 version. The simulation results show the MSLM scheme with 1024 carriers and different phase sequence reduce the PAPR about 0.25dB to 1dB. The computational complexity reduction ratio increase as the phase sequences increases. This makes the proposed scheme more suitable for high speed data rate OFDM system.
PAPR Reduction in OFDM by Using Modernize SLM Technique
405
References [1] Abouda: A PAPR reduction of OFDM signal using turbo coding and selective mapping. In: Proceedings of the 6th Nordic Signal Processing Symposium, June 2004, pp. 248–251 (2004) [2] Fischer, B.R., Huber, J.: Reducing " the peak-to-average power ratio of multicarrier modulation by selected mapping". IEE Electronics Letters 32, 2056–2057 (1996) [3] Baxley, R.J., Zhou, G.T.: Assessing "peak-to-average power ratios for communications applications". In: Proc. IEEE MILCOM Conference (October 2004) [4] Breiling, M., Muller-Weinfurtner, S.H., Huber, J.B.: SLM peak-power reduction without explicit side information. IEEE Communications Letters 5, 239–241 (2001) [5] Chang, R.W.: Orthogonal frequency division multiplexing. U.S. Patent 3 488 445 (January 1970) [6] Chen, N., Zhou, G.T.: Peak-to-average power ratio reduction in OFDM with blind selected pilot tone modulation. In: Proc. IEEE Intl. Conference on Acoustics Speech and Signal Processing (March 2005) [7] Cimini, L.J., Sollenberger, N.R.: Peak-to-average power ratio reduction of an OFDM signal using partial transmit sequences. IEEE Comm. Letters 4, 86–88 (1991) [8] Eevelt, P., Wade, M., Tomlinson, M.: Peak to average power reduction for OFDM schemes by selective scrambling. IEEE Electronics Letters 32, 1963–1964 (1996) [9] Lim, D.W., No, J.S., Lim, C.W., Chung, H.: A new SLM OFDM scheme with low complexity for PAPR reduction. IEEE Signal Processing Letters 12, 93–96 (2005) [10] Muller, S., Huber, J.: OFDM with reduced peak-to-average power ratio by optimum combination of partial transmit sequences. IEE Electronics Letters 33, 368–369 (1997) [11] Krongold, B.S., Jones, D.L.: PAR reduction in OFDM via active constellation extension. IEEE Trans. Broadcast 49(3), 258–268 (2002) [12] Zhou, G.T., Baxley, R.J., Chen, N.: Selected mapping with monomial phase rotations for peak-to-average power ratio reduction in OFDM. In: Proc. Intl. Conf.on Communications Circuits and Systems ( June 2004) [13] Wang, C.L., Ouyang, Y., Hsu, M.Y.: Low-complexity peak-toaverage power ratio reduction techniques for OFDM systems. Submitted to IEEE Transactions on Circuits and Systems - Part I (2004) [14] Tellado, J.: Multicarrier Modulation with Low PAR Applications to DSL and Wireless. Kluwer Academic Publishers, Dordrecht (2000) [15] Sathananthan, K., Tellambura, C.: Partial transmit sequence arid selected mapping schemes to reduce ICI in OFDM systems. IEEE Communications Letters 6, 313–315 (2002) [16] Zou, W.Y., Wu, Y.: COFDM: An overview. IEEE Trans. Broadcasting 41, 18 (1995) [17] Mesdagh, D., Spruyt, P.: A method to reduce the probability of clipping in DMT-based transceivers. IEEE Trans. Communications 44, 1234–1238 (1996) [18] Cripps, S.C.: RF Power Amplifiers for Wireless Communications. Artech House, Norwood (1999) [19] Jayalath, A.D.S., Tellambura, C.: Side information in PAR reduced PTS-OFDM signals. In: Personal, Indoor and Mobile Radio Communications conference, September 2003, vol. 1, pp. 226–230 (2003)
Application of Integrated Decision Support Model in Tendering Fadhilah Ahmad1 and M. Yazid M. Saman2 1
Faculty of Informatics, University Sultan Zainal Abidin Malaysia (UniSZA), Gong Badak Campus, 21300 Kuala Terengganu, Malaysia
[email protected] 2 Faculty of Science and Technology Universiti Malaysia Terengganu(UMT), 21030 Mengabang Telipot, Kuala Terengganu, Malaysia
[email protected] Abstract. Tendering is an important issue that requires Decision Support System (DSS) attention as a decision to award tenders to certain competing applications could influence successful completion of a project. This paper presents a framework of DSS for a tendering process based on a combination of single criteria statistical model, weighted model and an extended AHP model known as Guided AHP (GAHP). This hybrid model allows single criteria tender prices which are considered abnormal to be excluded from further detail multicriteria GAHP evaluation. GAHP is proposed to minimize the possibility of inconsistent data entry and to improve evaluation accuracy and flexibility. The use of model integration takes the advantage of their strengths and complements each other’s weaknesses. Finally, a real organizational government tendering application is applied to demonstrate the potential of the proposed framework. Keywords: Decision support system, Multi Criteria Decision Making, AHP, Statistical model.
1 Introduction Tendering problem is one of the areas that requires DSS attention as a decision to award tender to certain competing applications could influence successful completion of a project [1,2,3,4,5,6,7,8,9]. There are a number of research done on tender evaluation. [10] proposed the use of AHP in the selection of the best discount in dealing with the tenders for public work contract. [11] developed a DSS for awarding the tender to the lowest bidder using Delphi and AHP. Statistical model has been used in the tender evaluation for work contracts in Malaysian Government procurement. The use of the statistical model in this case is to evaluate tender prices. [12] has done a study on contractors’ perceptions of the statistical model. The result of the study shows that the use of this model in tender evaluation process is agreed by many contractors. However, the contractors have suggested the need to consider the A. Özcan, J. Zizka, and D. Nagamalai (Eds.): WiMo/CoNeCo 2011, CCIS 162, pp. 406–413, 2011. © Springer-Verlag Berlin Heidelberg 2011
Application of Integrated Decision Support Model in Tendering
407
current market price criterion and other tender related criteria in the evaluation process. The assertion imposed by the contractors in the study done by Faridah has some similarities with the previous study [13]. Both of these studies emphasized the need for multi criteria factors (other non-price factors) to be taken into consideration with the tender price factor that is currently evaluated via statistical model. In this study, the integration of statistical, weighted and extended AHP known as Guided AHP framework is developed for effective tender application evaluation and selection. The proposed research aims to investigate, design and implement integrated model for tendering process implementable in web based environment. The main strength of the proposed approach is in the integration of model bases, where they formally address both single criterion and multi-criteria evaluation. In the single criterion, statistical model is adopted. To facilitate the multi-criteria evaluation AHP model [14] is used because it is suitable to be adopted when there is a finite set of attributes to be considered in the evaluation process [17,18,19,21,22,23]. Firstly, the AHP model is experimented to foresee for its possible deficiency or impediment before it is implemented in the selected case study. During the testing, it is found that the users have difficulties to enter consistent data into the decision matrix. In order to reduce this problem, AHP is extended to become Guided AHP (GAHP). The organization of this paper is as follows: Section 2 presents the framework for modeling the tendering and selection process. Section 3 we analyze single criteria statistical model. In Section 4 we describe criteria ranking operations. In Section 5 we proposed GAHP multi-criteria model operations. Finally, the concluding remarks is given in Section 6.
2 Framework for Integrated Model The process flow for DSS tendering process is shown in Fig. 1, which is a fourphased process beginning with pre-requisite analysis at Stage I to the forth stage analysis using multi criteria decision making (MCDM). Each stage of the process has to be executed sequentially. This work focuses on Stage II onwards as it requires the judgments or intuitions from the decision makers (DMs).
3 Statistical Analysis The second stage is statistical analysis which can be executed based on the tender price. All the prices that are considered in this stage are from the tender documents that have passed through Stage I evaluation. The tender price is evaluated to determine the freak prices (or Z-scores) and cut-off price (COP).
408
F. Ahmad and M.Y.M. Saman
Stage I Prerequisite Analysis Database
Statistical Model Base (contains standard deviation, coefficient of variation routines)
Stage II Statistical Analysis Stage III Weighted Model Operation
Storing criteria ranking data
Stage IV Enhanced AHP Analysis
GRAHP Model Base (contains GRAHP routines)
Fig. 1. Process Flow for Integrated Model Base Approach
Freak prices are the values that are considered either too low or too high and they have to be rejected from further analysis. The calculation for freak price or Z-score, Zx for each tender price, is given as: (1) IF -2.33 < Zx < 2.33 then Mark the tender price as non-freak
(2)
The calculation for Standard deviation, σ is: ∑
where
(3)
Xi = tender price (that has passed through Stage I evaluation) µ = average of all tender prices N = number of tenders (maximum number of tenders that are considered)
The calculation for Coefficient of Variation, CV for all tender prices is shown as: (4) COP is the lowest price level assumed to be acceptable for further evaluation. Any tender prices lower than these values are assumed to be extremely low and if chosen they cannot guarantee successful completion of a project. However, some prices lower than these values can still be considered, provided that the applicants have some supporting factors that contribute to the project success.
Application of Integrated Decision Support Model in Tendering
409
The calculation of COP is done based on the following process: IF the number of non-freak price tenders > 10 then IF 0.01 1 1. Append the counter value and a “” symbol to the file c. Append the tag/attribute data to the file with tag/attribute id as file name d. Append a “>” symbol to the same file e. Reset counter of this tag to zero 5. Else a. Increment count for this tag b. Go to step 4 until end-of-file 3.2 Compression and Indexing 3.2.1 LZW Compression LZW Compression technique is one of the most standard and widely used compression techniques. The strength of the technique lies in the fact that, when applied on a document containing content of similar type format, they provide a greater compression performance. This comes most handy as most XML documents contain content that are mostly of the same type. QUICX takes upon this advantage and implements a slightly modified version of the standard LZW compression technique.
Query and Update Support for Indexed and Compressed XML (QUICX)
421
The technique also involves Indexing the Dictionary created by the LZW compression. Its procedure involves choosing a prime number as the size of the dictionary and it has to be larger than 2^n, where n equals the number of bits for the codeword. The hash function used is (C” symbol. The Ideal time updater reads from the Update File and the Reference File and makes the updates permanent by modifying the container files in ideal time.
5 Performance Analysis In order to compare the performance of our algorithm to the existing XML compressors, a set of experiments has been run. The following are the Open Source Data Sets, which have been used to test and evaluate our system. DBLP presents the famous database of bibliographic information of computer science journals and conference proceedings. Mondial is a demographic XML document that contains details about various countries. Shakespeare represents the gathering of a collection of marked-up Shakespeare plays into a single XML file. It contains many long textual passages. Treebank is a large collection of parsed English sentences from the Wall Street Journal. It has a very deep, non-regular and recursive structure. Swissprot is a protein sequence database which strives to provide a high level of annotations, a minimal level of redundancy and high level of integration with other databases. The test machine was an Intel Pentium core 2 Duo @1.5 GHz, running Windows XP. The RAM capacity is 1 GB. Our System is implemented in Microsoft Visual Basic. Visual Basic 6.0 is used for building our system. We make use of native Microsoft Parser for parsing the XML file for our system. We initially considered two parsers namely the SAX parser and the JDOM parser. While the SAX parser parses the XML file and parses each stream; it is done by calling as an event. It is efficient for larger files as it does not use Pointer concept. JDOM parser parses the XML file and parses as DOM (Document Object Model) which is based on the hierarchical structure of the file. It considers the whole XML file as a single Tree and uses pointers to traverse faster across the Tree’s child
424
R. Senthilkumar and A. Kannan
nodes. It is efficient for smaller files as it uses Pointer concept. But while considering larger files, the use of Pointers require huge resources (Memory and Processor), hence it is not dependable. But Microsoft Parser, which is also an implementation of SAX parser, takes much less time than both these parsers which prompted us to makeshift this work into the windows platform and Visual Basic. Compression component of QUICX is currently implemented in C language for simplicity and can be easily migrated to any language. Java is used to update the QUICX database.
5.1 Performance Evaluation Criteria Performance of QUICX is evaluated in terms of storage space and support for querying and updation. We divide the storage space comparison into two parts and they are structure conversion compared with Queriable and Non-Queriable compressors and Data Comparison with Queriable and Non-Queriable compressors. In Table 3 the performance measure of various benchmark datasets for QUICX are given. 5.2 Performance Comparison We use two metric in our evaluation of the efficiency of XML compression tools: the compression ratio and the compression time. Table 3. Performance measures of various benchmark datasets for QUICX
Compression Ratio: We express the compression ratio as the ratio the size of the compressed document to the original document. For example, if a 10 MB file can be compressed to 2.5 MB, the file if 75% compressed. Higher compression ratios are obviously, better. Compression ratio =
1 -
Compressed file size ___________________ Original size
Compression time: Compression time represents the elapsed time during compression, expressed as the time from the start of program execution on a document until all the data are writer to disk. We do not separate CPU and I/O times.
Query and Update Support for Indexed and Compressed XML (QUICX)
425
Fig. 5 and Fig. 6 show the compression ratio using the three datasets (Shakespeare, Swissprot and DBLP) after the structure and data compression with the following Data Sources HUFFWORD, XPRESS, XQZIP, XBZIPINDEX, XBZIP, RFX and QUICX.
Fig. 5. Compression Ratio achieved by queriable compressor over the Benchmark datasets after structure Conversion
Fig. 6. Compression Ratio achieved by queriable compressor over the Benchmark datasets after Data Compression
426
R. Senthilkumar and A. Kannan
HUFFWORD [17] is a variant of classical HUFFMAN compressor, which is the typical storage scheme of (Web) search engines and Information retrieval tool and the average compression ratio of HUFFWORD is 43.06%.
Fig. 7. Average Compression Ratio
Fig. 8. QUICX after Structure and Data level compression
XPRESS[7] uses type inference engine to apply appropriate compression methods for data values, and its average CR is 45.67% . Thus the average compression ratio for XQZIP is of 36. The average CR of QUICX is of 44.41% at structure level, since QUICX eliminates all duplicate tags at structure level compression and uses LZW for
Query and Update Support for Indexed and Compressed XML (QUICX)
427
data level compression. LZW when applied on a document containing content of similar type/format provides a greater compression performance. Thus, the compression ratio of QUICX (SL) for Swissprot is 54.86 is better than that of other datasets as Swissprot has more redundant tags. Fig. 7 shows that the difference on the average compression ratios between the queriable compressors are close to each other and the difference between the best and the worst average compression ratios is less than 50%. Among all compressors, QUICX achieves the best average compression ratio. Fig.8 shows the compression ratio by QUICX after the structure and data compression on various bench mark datasets.
6 Conclusion In this work, we have proposed a new compact storage structure called QUICX for XML. It also provides sufficient support in the form of indexing for faster query processing which reduces the resource required such as space (memory) and processor requirements, increasing the processing speed. Our experimental results show that the proposed structure can handle query processing and updates efficiently. It is also shown that the proposed structure out performs other storage system and is significantly good for larger datasets. The proposed structure also allows direct update and querying without decompressing the entire document. In future, we have planned to extend QUICX with more features to accommodate multimedia data and query processing technique to process such storage techniques.
References 1. Arion, A., Bonifati, A., Manolescu, I., Pugliese, A.: XQueC: A Query-Conscious Compressed XML Database. ACM Transactions on Internet Technology 7(2), 1–32, Article 10(2007) 2. Farina, A., Ladra, S., Pedreira, O., Places, A.S.: Rank and Select for Succinct Data Structures. Electronic Notes in Theoretical Computer Science 236, 131–145 (2009) 3. Lee, C.-S., Haw, S.-C.: Extending path summary and region encoding for efficient structural query processing in native XML databases. The Journal of Systems and Software 82(6), 1025–1035 (2009) 4. Lam, F., Wong, R.K., Shui, W.M.: Querying and maintaining a compact XML storage. In: Proceedings of the 16th international conference on World Wide Web, Banff, Alberta, Canada, pp. 1073–1082 (May 2007) 5. Liefke, H., Suciu, D.: XMILL: An Efficient Compressor for XML Data. In: Proceedings of the 2000 ACM SIGMOD International Conference on Management of Data, Dallas, TX, USA, pp. 153–164 (2000) 6. Cheng, J., Ng, W.: XQzip: Querying compressed XML using structural indexing. In: Hwang, J., Christodoulakis, S., Plexousakis, D., Christophides, V., Koubarakis, M., Böhm, K. (eds.) EDBT 2004. LNCS, vol. 2992, pp. 219–236. Springer, Heidelberg (2004)
428
R. Senthilkumar and A. Kannan
7. Min, J.K., Park, M.J., Chung, C.W.: XPRESS: A queriable compression for XML data. In: Proceedings of the ACM SIGMOD International Conference on Management of Data, June 2003, pp. 122–133 (2003) 8. Levene, M., Wood, P.T.: XML structure compression. In: DEXA 2008, 19th International Conference on Database and Expert Systems Application pp. 243–247. IEEE Computer Society, Washington, DC,USA (2008) 9. Grimsmo, N.: Faster Path Indexes for Search in XML Data. In: Conference in Research and practice in Information Technology (CRPIT) – Database Technologies, vol. 75, pp. 127–135 (2008) 10. Tolani, P.M., Haritsa, J.R.: XGRIND: A query-friendly XML compressor. In: Proceedings of the 18th International Conference on Data Engineering (ICDE), February 26 -March 1, San Jose, USA, pp. 225–234 (2002) 11. Senthilkumar, R., Priyaa Varshinee, S., Manipriya, S., Gowrishankar, M., Kannan, A.: Query Optimization of RFX Compact Storage using Strategy List. In: 16th international conference on Advanced Computing & Communication (ADCOM), pp. 427–432 (December 2008) 12. Alkhatib, R., Scholl, M.H.: CXQU: A Compact XML Storage for Efficient Query and Update processing. In: 3rd International Conference on Digital Information Management (ICDIM), November 13 -16, pp. 605–612 (2008) 13. W3C Group. Guide to the W3C XML specification (XMLspec) DTD, version 2.1. (2004), http://www.w3.org/XML/1998/06/xmlspec-report.htm 14. Welch, T.: A Technique for High-Performance Data Compression. IEEE Computer 17(6), 8–9 (1984) 15. Cheny, J.: Compressing XML with Multiplexed Hierarchical PPM Models. IEEE Computer, 1–10 (2001) 16. Moura, E., Navarro, G., Ziviani, N., Baeza-Yates, R.: Fast and flexible word searching on compressed text. ACM Transactions on Information Systems 18(2), 113–139 (2000) 17. http://www.ebi.uniprot.org 18. DBLP, http://dblp.uni-trier.de/
Author Index
Abdelmalek, Abdelhafid 73 Abdu, Abdullahi Ibrahim 135 Ahmad, Fadhilah 406 Al-Asfoor, Muntasir 337 Aldabbas, Hamza 297 Aldabbas, Omer 297 Al-Majeed, Salah 337 Alwada’n, Tariq 297 Ananthapadmanabha, T. 57 Artin, Javad 145 Azer, Marianne Amir 65 Baccouche, Le¨ıla 123 Bag, Samiran 83 Bai, Yun 378, 388 Bandyopadhyay, Soma 288 Ben Ghezala, Henda 123 Ben Haj Hmida, Moez 348 Bouallegue, Ridha 185 Boucherkha, Samia 256
Janicke, Helge 297 Joseph C., Rijo 170 Kafaie, Somayeh 358 Kannan, A. 414 Kashefi, Omid 358 Kecman, Vojislav 110 Khandagre, Yogeshver 397 Kianersi, Mehrdad 236 Kushwaha, Ganesh Raj 397 Li, Qi 110 Linington, Peter Luo, Guangcun
325 95
Maiti, Souvik 288 Maleki, Amin Amini 246 Mozafari, Saadat Pour 228 Mukherjee, Amartya 170
Chaki, Rituparna 14, 24, 35 Chen, Aiguo 95 Chen, Hao 95 Chettibi, Saloua 1 Chikhi, Salim 1, 217
Nandi, Subrata 170, 197 Nandi, Sukumar 305
Dagdeviren, Orhan 315 Das, Ayan Kumar 24 da Silva, Erick Lopes 325 Djeghri, Mohamed Nadjib 256 Drira Rekik, Jihen 123 Dubey, Ashutosh K. 397 Dutta, Subhajit 288 Erciyes, Kayhan
Haddad, Rim 185 Hashemi, Reza 246 Hassan, Md. Mahedi 45 Hemnani, Khushboo 397
Patra, Sushovan 197 ¨ Payli, Re¸sat Umit 315 Policarpio, Sean 378, 388 Poo, Kuan Hoong 45 Puttamadappa, C. 57 Qin, Ke 95 Qu, Caihui 95
315 Rabieh, Khaled Mohamed 65 Rahimunnisa, K. 279 Rajeshkumar, K. 279 Ridjanovic, Dzenan 368 Roy, Debdutta Barman 14
Fasli, Maria 337 Feham, Mohamed 73 Fong, Simon 207, 266 Gardeshi, Mahmoud 236 Gherboudj, Amira 217 Ghorbannia Delavar, Arash
145
Saha, Amrita 83 Saha, Sujoy 170, 197
430
Author Index
Salamah, Muhammed 135, 159 Salman, Raied 110 Saman, M. Yazid M. 406 Sarı¸sın, G¨ ozde 159 Sarkar, Pinaki 83 Sarma, Abhijit 305 Sengupta, Munmun 288 Sengupta, Satadal 197 Senthilkumar, Radha 414 Shah, Vijay 197 Sharifi, Mohsen 358 Sheidaeian, Hamed 228 Sheldekar, Anirudh 170 Shrivastava, Nishant 397 Singh, Konsam Gojendra 197 Sinha, Ditipriya 35
Slimane, Zohra 73 Slimani, Yahya 348 Strack, Robert 110 Suma, A.P. 57 Surekha, T.P. 57 Sureshkumar, S. 279 Sushovan 170 Tajari, Mohammad Mahdi Taleb-Ahmed, Abdelmalik Test, Erick 110 Tiwari, Ruby 397 Yousefi, Hamed 236 Yousefi, Saleh 246 Zolfaghari, Behrouz
228
145 73