Wireless Communications IEEE Volume 18 Issue 1 February 2011 issue Volume 18 Issue 1 February 2011

A Publication of the IEEE Communications SocietyIn cooperation with IEEE Computer and VehicularTechnology Societies

®

WIRELESSIEEE

COMMUNICATIONS

February 2011, Vol. 18, No. 1• IMS EMERGENCY SERVICES: A PRELIMINARY STUDY• MIXING NETWORK CODING AND COOPERATION FOR RELIABLEWIRELESS COMMUNICATIONS

• SPECTRUM SENSING FOR COGNITIVE RADIO SYSTEMS• COOPERATIVE COMMUNICATION IN MULTIHOP COGNITIVE RADIO NETWORKSBASED ON MULTICARRIER MODULATION

• TOPOLOGICAL-BASED ARCHITECTURES FOR WIRELESS MESH NETWORK

IMS EMERGENCYSERVICES

February 2011 Wireless Cover 1 2/7/11 12:20 PM

1

WIRELESSIEEE

COMMUNICATIONSACCEPTED FROM OPEN CALL

6IMS EMERGENCY SERVICES:

A PRELIMINARY STUDYYI-BING LIN, MENG-HSUN TSAI, AND YUAN-KUANG TU

15MIXING NETWORK

CODING AND COOPERATION FORRELIABLE WIRELESS COMMUNICATIONS

FRANCESCO ROSSETTO AND MICHELE ZORZI

22SECURING UNDERWATER

WIRELESS COMMUNICATION NETWORKSMARI CARMEN DOMINGO

30SPECTRUM SENSING FOR

COGNITIVE RADIO SYSTEMS: TECHNICAL ASPECTS AND

STANDARDIZATION ACTIVITIES OF THEIEEE P1900.6 WORKING GROUP

KLAUS MOESSNER, HIROSHI HARADA, CHEN SUN,YOHANNES D. ALEMSEGED, HA NGUYEN TRAN,

DOMINIQUE NOGUET, RYO SAWAI, AND NAOTAKA SATO

38MULTICARRIER MODULATION ANDCOOPERATIVE COMMUNICATION IN

MULTIHOP COGNITIVE RADIO NETWORKSTAO LUO, FEI LIN, TAO JIANG, MOHSEN GUIZANI,

AND WEN CHEN

IEEE Wireless Communications • February 2011

FEBR U A RY 2011/VOL . 18, NO. 1

46CHALLENGES, OPPORTUNITIES, AND

SOLUTIONS FOR CONVERGED SATELLITE ANDTERRESTRIAL NETWORKS

TARIK TALEB, YASSINE HADJADJ-AOUL, AND TOUFIK AHMED

54INTERFERENCE COORDINATION FOR

OFDM-BASED MULTIHOPLTE-ADVANCED NETWORKS

KAN ZHENG, BIN FAN, JIANHUA LIU, YICHENG LIN, AND WENBO WANG

64DISTRIBUTED AUTOMATED

INCIDENT DETECTION WITH VGRIDBEHROOZ KHORASHADI, FRED LIU, DIPAK GHOSAL,

MICHAEL ZHANG, AND CHEN-NEE CHUAH

74TOPOLOGICAL-BASED ARCHITECTURES FOR

WIRELESS MESH NETWORKSAMIR ESMAILPOUR, NIDAL NASSER, AND TARIK TALEB

82SYNCHRONIZATION OF

MULTIHOP WIRELESS SENSOR NETWORKS ATTHE APPLICATION LAYER

ÁLVARO MARCO, ROBERTO CASAS, JOSÉ LUIS SEVILLANORAMOS, VICTORIAN COARASA, ÁNGEL ASENSIO,

AND MOHAMMAD S. OBAIDAT

MESSAGE FROM THE EDITOR-IN-CHIEF — 2SCANNING THE LITERATURE — 4

Cover image: Getty Images

LYT-TOC-Feb 2/7/11 11:41 AM Page 1


MESS A G E F ROM THE ED I T O R- I N -CH I E F

irst of all, I am happy to report that myterm as the Editor-in-Chief has beenextended for one more year. I also want

to report that we have been working onmoving into Manuscript Central, which willbe a great step toward a more efficientpaper handling process. As a new yearstarts, we also look forward to more excit-ing news on wireless communications, net-working, and their applications. You mayhave already read the first issue of IEEESpectrum this year. Among the top 11 iden-tified technologies (although it is not clearwhy it is 11 rather than the more commonlyused 10, perhaps to avoid a copyright law-suit from David Letterman?), smartphonesare ranked on top. Unfortunately, the col-umn does not spell out the drivers of futuresmartphone technologies: the deep integration technologies ofsmartphones with smart environments and human interactionsfor ubiquitous data collection, processing, data mining, com-

munications and networking, and decision mak-ing and control. The second technology on thetop 11 list is social networking, which has closeties to wireless technologies as most users havealready been doing their social networking withsmartphones or mobile devices, leading to theever more popular mobile social networking. Aswe can see, wireless technologies have alreadyrevolutionized the way we live and have trans-formed our society into a completely differentkind for everything imaginable. We, the engi-neers and scientists, are at center stage to uti-lize the technologies we have created to make asignificantly different society, good or bad, andit is up to us to reshape it and our living envi-ronments for better quality of life. We need youto invent new technologies, including wirelesstechnologies, and we need new ideas and new

innovations from you. Our magazine offers you the platformfor you to spread the word. We are searching for articles onnew innovations in the wireless area and solicit new special

CALL FOR MORE WIRELESS INNOVATIONS

Director of MagazinesAndrzej Jajszczyk, AGH U. of Sci. & Tech. Poland

Editor-in-ChiefYuguang Michael Fang, Univ. of Florida, USA

Associate Editor-in-ChiefDilip Krishnaswamy, Qualcomm Research Center

Senior AdvisorsHamid Ahmadi, AT&T Labs, USA

Abbas Jamalipour, University of Sydney, AustraliaThomas F. La Porta, Pennsylvania State Univ., USA

Mahmoud Naghshineh, IBM, USAMichele Zorzi, University of Padova, Italy

Advisory BoardDonald Cox, Stanford University, USA

David Goodman, Polytechnic University, USATero Ojanperä, Nokia, Finland

Kaveh Pahlavan, Worcester Polytech. Inst., USAMahadev Satyanarayanan, CMU, USAIEEE Vehicular Technology Liaison

Theodore Rappaport, Univ. of Texas, Austin, USAIEEE Computer Society Liaison

Mike Liu, Ohio State University, USATechnical Editors

Abouzeid Alhussein, Rensselaer Polytechnic Inst., USABenny Bing, Georgia Tech, USA

Azzedine Boukerche, Univ. of Ottawa, CanadaJyh-Cheng Chen, Natl. Chiao Tung Univ., Taiwan

Carla-Fabiana Chiasserini, Politecnico di Torino, ItalySunghyun Choi, National Seoul University, KoreaMischa Dohler, France Telecom R&D, France

Ekram Hossain, University of Manitoba, CanadaThomas Hou, Virginia Tech., USA

Nei Kato, Tohoku University, JapanPascal Lorenz, U. of Haute Alsace, France

Giacomo Morabito, U. di Catania, ItalyZhisheng Niu, Tsinghua University, China

Symeon Papavassiliou, Natl. Tech. Univ. Athens, GreeceVincent Poor, Princeton Univ., USA

Kui Ren, Illinois Institute of Tech., USAApostolis Salkintzis, Motorola, GreeceJohn Shea, University of Florida, USA

Sherman Shen, Univ. of Waterloo, CanadaRichard Wolff, Montana State Univ., USA

Junshan Zhang, Arizona State Univ.Qian Zhang, Hong Kong Univ. Science & Tech.,

Hong Kong

Department EditorsIndustrial Perspectives

Benny Bing, Georgia Tech, USAScanning the Literature

Yanchao Zhang, Arizona State Univ., USASpectrum Policy and Reg. Issues

Michael Marcus, Marcus Spectrum Solns., USA2011 Communications Society

Board of GovernorsOfficers

Byeong Gi Lee, PresidentMark Karol, VP–Technical ActivitiesKhaled B. Letaief, VP–Conferences

Sergio Benedetto, VP–Member RelationsLeonard Cimini, VP–PublicationsDoug Zuckerman, Past President

Stan Moyer, TreasurerJohn M. Howell, Secretary

Members-at-LargeClass of 2011

Robert Fish • Joseph EvansNelson Fonseca • Michele Zorzi

Class of 2012Stefano Bregni • V. Chan

Iwao Sasase • Sarah K. WilsonClass of 2013

Gerhard Fettweis • Stefano GalliRobert Shapiro • Moe Win

2011 IEEE OfficersMoshe Kam, President

Gordon W. Day, President-ElectRoger D. Pollard, Secretary

Harold L. Flescher, TreasurerPedro A. Ray, Past-President

E. James Prendergast, Executive DirectorNim Cheung, Director, Division III

IEEE Wireless Communications (ISSN 1536-1284) is pub-lished bimonthly by The Institute of Electrical and ElectronicsEngineers, Inc. Headquarters address: IEEE, 3 Park Avenue,17th Floor, New York, NY 10016-5997; tel: 212-705-8900; fax:212-705-8999; e-mail: [email protected]. Responsibilityfor the contents rests upon authors of signed articles and notthe IEEE or its members. Unless otherwise specified, theIEEE neither endorses nor sanctions any positions or actionsespoused in IEEE Wireless Communications.

Annual subscription: Member subscription: $40 per year;Non-member subscription: $250 per year. Single copy: $50.

Editorial correspondence: Manuscripts for considerationmay be submitted to the Editor-in-Chief: Yuguang MichaelFang, University of Florida, 435 New Engineering Building,P.O. Box 116130, Gainesville, FL 32611. Electronic submissionsmay be sent in postscript to: [email protected].

Copyright and reprint permissions: Abstracting is permit-ted with credit to the source. Libraries permitted to photocopybeyond limits of U.S. Copyright law for private use ofpatrons: those post-1977 articles that carry a code on the bottomof first page provided the per copy fee indicated in the code ispaid through the Copyright Clearance Center, 222Rosewood Drive, Danvers, MA 01923. For other copying,reprint, or republication permission, write to Director,Publishing Services, at IEEE Headquarters. All rights reserved.Copyright © 2011 by The Institute of Electrical and ElectronicsEngineers, Inc.

Postmaster: Send address changes to IEEE WirelessCommunications, IEEE, 445 Hoes Lane, Piscataway, NJ 08855-1331; or email to [email protected]. Printed in USA.Periodicals postage paid at New York, NY and at additionalmailing offices. Canadian GST #40030962. Return undeliver-able Canadian addresses to: Frontier, PO Box 1051, 1031Helena Street, Fort Eire, ON L2A 6C7.

Subscriptions: Send orders, address changes to: IEEE ServiceCenter, 445 Hoes Lane, Piscataway, NJ 08855-1331; tel.: 908-981-0060.

Advertising: Advertising is accepted at the discretion of the pub-lisher. Address correspondence to: Advertising Manager, IEEEWireless Communications, 3 Park Avenue, 17th Floor, NewYork, NY 10016.

Joseph Milizzo, Assistant PublisherEric Levine, Associate Publisher

Susan Lange, Online Production ManagerJennifer Porcello, Production Specialist

Catherine Kemelmacher, Associate Editor

WIRELESSIEEE

COMMUNICATIONS

®

YUGUANG MICHAEL FANG

F

LYT-EDIT NOTE-February 2/7/11 10:39 AM Page 2

IEEE Wireless Communications • February 2011 3

issues on hot topics that are of great interest to our reader-ship. Some of the topics we are particularly interested in arewireless technologies in cyber-physical systems, social net-works, smart grids, healthcare and medical systems, urbansensing and public safety, and cloud computing systems. Withthe upcoming electronic paper handling process, I hope wecan make the publication process much faster. For those whoare engaging in hot wireless innovations, I encourage you toorganize special issues to share your excitement with our gen-eral audiences.

Due to the unforeseen delay in getting articles for theplanned special issue, we have decided to accommodate arti-cles from our open call to cut the long queue of acceptedpapers. We have selected 10 articles in this issue; their briefsummaries follow.

“IMS Emergency Services: A Preliminary Study,” by Yi-Bing Lin, Meng-Hsun Tsai and Yuan-Kuang Tu, presents astudy on how to support emergency call and walkie-talkie ser-vices over IP Multimedia Subsystem in wireless cellular sys-tems.

“Mixing Network Coding and Cooperation for ReliableWireless Communications,” by Francesco Rossetto andMichele Zorzi, gives an overview of how to take advantage ofboth cooperation and network coding to improve the perfor-mance and error correction capabilities of radio networks andhighlights the main challenges for future research.

“Securing Underwater Wireless Communication Net-works,” by Mari Carmen Domingo, reviews some importantsecurity design issues specific to underwater wireless commu-nication networks and discusses possible research challenges.

“Spectrum Sensing for Cognitive Radio Systems: TechnicalAspects and Standardization Activities of IEEE 1900.6 Work-ing Group,” by Klaus Moessner, Hiroshi Harada, Chen Sun,Yohannes D. Alemseged, Ha Nguyen Tran, DominiqueNoguet, Ryo Sawai, and Naotaka Sato, overviews the techni-cal issues on spectrum sensing for cognitive radio systems andthe related IEEE standardization activities for sensing infor-mation exchange, particularly focused on activities from theIEEE P1900.6 working group.

“Cooperative Communication in Multihop CognitiveRadio Networks Based on Multicarrier Modulation,” by TaoLuo, Fei Lin, Tao Jiang, and Mohsen Guizani, studies themulticarrier modulation schemes for multihop cognitive radionetworks and shows that filtered multitone modulation per-forms better than orthogonal frequency-division multiplexing(OFDM) in terms of mutual interference elimination, syn-chronization, and transmission efficiency. Moreover, theauthors combine cognitive radio capability with cooperativediversity and come up with three efficient cooperative diversi-ty cognitive models.

“Challenges, Opportunities, and Solutions for ConvergedSatellite and Terrestrial Networks,” by Tarik Taleb, YassineHadjadj-Aoul and Toufik Ahmed, investigates some impor-tant design issues related to interworking operations betweenthe satellite and terrestrial domains in order to support a widevariety of services for users with a variety of roles (consumer,producer, or manager of communication and media), and sug-gests some possible solutions and their potential.

“Interference Coordination in OFDM-Based MultihopCellular Networks toward LTE-Advanced,” by Kan Zheng,

Bin Fan, Yicheng Lin, and Wenbo Wang, presents anoverview of the interference coordination strategies forOFDM-based multihop cellular networks and proposes sever-al static or semi-static interference coordination schemesbased on the framework of Long Term Evolution (LTE)-Advanced networks with multihop relaying to improve cover-age and increase the data rate over cell edge areas.

“Distributed Automated Incident Detection with Vgrid,”by Behrooz Khorashadi, Fred Liu, Dipak Ghosal, Chen-NeeChuah, and Michael Zhang, studies an ad hoc distributedautomated incident detection algorithm for highway trafficusing vehicles that are equipped with wireless communica-tions, processing, and storage capabilities. By requesting vehi-cles with such capability to broadcast beacon informationcontaining their speed, location, and lane information, thedetection algorithm can make better decisions and yield bet-ter performance.

“Topological-Based Architectures for Wireless Mesh Net-work,” by Amir Esmailpour, Nidal Nasser, and Tarik Taleb,provides an overview on architectural design for wireless meshnetworks, summarizes the state-of-the-art research findings,and calls for further research on this topic.

“Synchronization of Multihop Wireless Sensor Networks atthe Application Layer,” by Álvaro Marco, Roberto Casas,José Luis Sevillano, Victorián Coarasa, Ángel Asensio, andMohammad S. Obaidat, proposes a method for accurate syn-chronization of large multihop networks, which operates atthe application layer while minimizing message exchange.

I hope you enjoy reading these articles. I also wish you aproductive 2011!

BIOGRAPHYYUGUANG MICHAEL FANG [F’08] ([email protected]) received a Ph.D. degreein systems engineering from Case Western Reserve University in January1994 and a Ph.D. degree in electrical engineering from Boston Universityin May 1997. He was an assistant professor in the Department of Electricaland Computer Engineering at New Jersey Institute of Technology from July1998 to May 2000. He then joined the Department of Electrical and Com-puter Engineering at the University of Florida in May 2000 as an assistantprofessor, got an early promotion to associate professor with tenure inAugust 2003, and to full professor in August 2005. He held a University ofFlorida Research Foundation (UFRF) Professorship from 2006 to 2009, aChangjiang Scholar Chair Professorship with Xidian University, Xi’an,China, from 2008 to 2011, and a Guest Chair Professorship with TsinghuaUniversity, China, from 2009 to 2012. He has published over 250 papersin refereed professional journals and conferences. He received the NationalScience Foundation Faculty Early Career Award in 2001 and the Office ofNaval Research Young Investigator Award in 2002, and is the recipient ofthe Best Paper Award from the IEEE International Conference on NetworkProtocols (ICNP) in 2006 and the recipient of the IEEE TCGN Best PaperAward at the IEEE High-Speed Networks Symposium, IEEE GLOBECOM in2002. He is also active in professional activities. He is a member of ACM.He is currently serving as the Editor-in-Chief for IEEE Wireless Communica-tions (2009–present) and serves/has served on several editorial boards oftechnical journals including IEEE Transactions on Mobile Computing(2003–2008, 2011–present), IEEE Transactions on Communications(2000–present), IEEE Transactions on Wireless Communications(2002–2009), IEEE Journal on Selected Areas in Communications(1999–2001), IEEE Wireless Communications (2003–2009), and ACM Wire-less Networks (2001-present). He served on the Steering Committee forIEEE Transactions on Mobile Computing (2008–2010). He has been activelyparticipating in professional conference organizations such as serving asSteering Committee Co-Chair for QShine (2004-2008), Technical ProgramVice-Chair for IEEE INFOCOM’2005, Technical Program Area Chair for IEEEINFOCOM (2009–2012), Technical Program Symposium Co-Chair for IEEEGLOBECOM 2004, and member of the Technical Program Committee forIEEE INFOCOM (1998, 2000, 2003–2008).

MESS A G E F ROM THE ED I T O R- I N -CH I E F

LYT-EDIT NOTE-February 2/7/11 10:39 AM Page 3


SC A N N I N G T H E L I T E R AT U R E

EDITED BY YANCHAO ZHANG

Using Classification to Protect theIntegrity of Spectrum Measure-ments in White Space NetworksO. Fatemieh, A. Farhadi, R. Chandra, andC. Gunter, in the 18th Annual Network& Distributed System Security Sympo-sium (NDSS), San Diego, CA, February2010

The emerging paradigm for using thewireless spectrum more efficiently isbased on enabling secondary users toexploit white space frequencies that arenot occupied by primary users. A keyenabling technology for forming net-works over white spaces is distributedspectrum measurements to identify andassess the quality of unused channels.This spectrum availability data is oftenaggregated at a central base station ordatabase to govern the usage of spec-trum. This process is vulnerable tointegrity violations if the devices aremalicious and misreport spectrum sens-ing results. This paper presents CUSP, anew technique based on machine learn-ing that uses a trusted initial set of signalpropagation data in a region as input tobuild a classifier using support vectormachines. The classifier is subsequentlyused to detect integrity violations. Usingclassification eliminates the need forarbitrary assumptions about signal propa-gation models and parameters or thresh-olds in favor of direct training data.Extensive evaluations using TV transmit-ter data from the FCC, terrain data fromNASA, and house density data from theU.S. Census Bureau for areas in Illinoisand Pennsylvania show that CUSP iseffective against attackers of varyingsophistication, while accommodatingregional terrain and shadowing diversity.

Privacy-Preserving RegressionModeling of Participatory Sensing DataH. Ahmadi, N. Pham, R. Ganti, T.Abdelzaher, S. Nath, and J. Han, inthe 8th ACM Conference on EmbeddedNetworked Sensor Systems (SenSys),Zurich, Switzeland, November 2010

Many participatory sensing applicationsuse data collected by participants to con-struct a public model of a system or phe-nomenon. For example, a healthapplication might compute a model relat-ing exercise and diet to amount of weightloss. While the ultimately computedmodel could be public, the individual

input and output data traces used to con-struct it may be private data of partici-pants (e.g., their individual food intake,lifestyle choices, and resulting weight).This paper proposes and experimentallystudies a technique that attempts to keepsuch input and output data traces pri-vate, while allowing accurate model con-struction. This is significantly differentfrom perturbation-based techniques inthat no noise is added. The main contri-bution of the paper is to show a certaindata transformation at the client sidethat helps keeping the client data privatewhile not introducing any additionalerror to model construction. The authorsparticularly focus on linear regressionmodels which are widely used in partici-patory sensing applications. They use thedata set from a map-based participatorysensing service to evaluate their scheme.The service in question is a green naviga-tion service that constructs regressionmodels from participant data to predictthe fuel consumption of vehicles on roadsegments. They evaluate the proposedmechanism by providing empirical evi-dence that: i) an individual data trace isgenerally hard to reconstruct with anyreasonable accuracy, and ii) the regres-sion model constructed using the trans-formed traces has a much smaller errorthan one based on additive data-pertur-bation schemes.

Reliable Clinical Monitoring Using Wire-less Sensor Networks: ExperiencesIn A Step-Down Hospital UnitO. Chipara1, C. Lu, T. Bailey, and G.Roman, in the 8th ACM Conference onEmbedded Networked Sensor Systems(SenSys), Zurich, Switzeland, Novem-ber 2010

This paper presents the design, deploy-ment, and empirical study of a wirelessclinical monitoring system that collectspulse and oxygen saturation readingsfrom patients. The primary contributionof this paper is an in-depth clinical trialthat assesses the feasibility of wirelesssensor networks for patient monitoringin general hospital units. The authorspresent a detailed analysis of the systemreliability from a long-term hospitaldeployment over seven months involving41 patients in a step-down cardiologyunit. The network achieved high reliabil-ity (median 99.68 percent, range95.21–100 percent). The overall reliabili-ty of the system was dominated by sens-

ing reliability of the pulse oximeters(median 80.85 percent, range 0.46–97.69percent). Sensing failures usuallyoccurred in short bursts, although longerperiods were also present due to sensordisconnections. The authors show thatthe sensing reliability could be signifi-cantly improved through oversamplingand by implementing a disconnectionalarm system that incurs minimal inter-vention cost. A retrospective data analy-sis indicated that the system providedsufficient temporal resolution to supportthe detection of clinical deterioration inthree patients who suffered from signifi-cant clinical events including transfer tointensive care units.

CodeOn: Cooperative Popular ContentDistribution for Vehicular NetworksUsing Symbol Level Network Cod-ingM. Li, Z. Yang, and W. Lou, IEEE Jour-nal on Selected Areas in Communica-tions (JSAC), special issue on VehicularCommunication Networks, vol. 29, no. 1,January 2011 Driven by both safety concerns and com-mercial interests, one of the key servicesoffered by vehicular networks is popularcontent distribution (PCD). The funda-mental challenges to achieve high speedcontent downloading come from thehighly dynamic topology of vehicular adhoc network (VANET) and the lossynature of the vehicular wireless commu-nications. This paper introducesCodeOn, a novel push-based PCDscheme where contents are activelybroadcasted to vehicles from road sideaccess points and further distributedamong vehicles using a cooperativeVANET. In CodeOn, we employ arecent technique, symbol level networkcoding (SLNC), to combat the lossywireless transmissions. Through exploit-ing symbol level diversity, SLNC isrobust to transmission errors andencourages more aggressive concurrenttransmissions. In order to fully enjoy thebenefits of SLNC, we propose a suite oftechniques to maximize the downloadingrate, including a prioritized and localizedrelay selection mechanism where theselection criteria are based on the use-fulness of vehicle-possessed contents,and a lightweight medium access proto-col that naturally exploits the abundantconcurrent transmission opportunities.We also propose additional mechanisms

LYT-SCANNING-February 2/7/11 10:40 AM Page 4


to reduce the protocol overhead withoutsacrificing the performance.

Free Side Channel: Bits over Inter-ferenceK. Wu, H. Tan, Y. Liu, J. Zhang, Q. Zhang,and L. Ni, in the 16th Annual InternationalConference on Mobile Computing andNetworking (MobiCom), Chicago, Illinois,September 2010

Interference is a critical issue in wire-less communications. In a typical multi-ple-user environment, different usersmay severely interfere with each other.Coordination among users therefore isan indispensable part of interferencemanagement in wireless networks. It isknown that coordination among multi-ple nodes is a costly operation, taking asignificant amount of valuable commu-nication resources. In this paper, theauthors have an interesting observationthat by generating intended patterns,some simultaneous transmissions (i.e.,“interference”) can be successfullydecoded without degrading the effec-

tive throughput in the original transmis-sion. As such, an extra and “free” coor-dination channel can be built.

Based on this idea, the authors pro-pose a DC-MAC to leverage this “free”channel for efficient medium access ina multiple-user wireless network. Theytheoretically analyze the capacity of thischannel under different environmentswith various modulation schemes.

Enabling High-Bandwidth VehicularContent DistributionU. Shevade, Y. Chen, L. Qiu, Y. Zhang,V. Chandar, M. Han, H. Song, and Y.Seung, in the 6th International Conferenceon emerging Networking EXperiments andTechnologies (CoNEXT), Philadelphia, PA,November 2010

This paper presents VCD, a novel sys-tem for enabling high-bandwidth con-tent distribution in vehicular networks.In VCD, a vehicle opportunisticallycommunicates with nearby access points(APs) to download the content of inter-est. To fully take advantage of such

transient contact with APs, the authorsproactively push content to the APs thevehicles are likely to visit in the nearfuture. In this way, vehicles can enjoythe full wireless capacity instead ofbeing bottlenecked by Internet connec-tivity, which is either slow or evenunavailable. The authors develop a newalgorithm for predicting the APs thatwill soon be visited by the vehicles. Theythen develop a replication scheme thatleverages the synergy among (i) Internetconnectivity (which is persistent but haslimited coverage and low bandwidth),(ii) local wireless connectivity (whichhas high bandwidth but transient dura-tion), (iii) vehicular relay connectivity(which has high bandwidth but highdelay), and (iv) mesh connectivityamong APs (which has high bandwidthbut low coverage). The authors demon-strate the effectiveness of the VCD sys-tem using trace-driven simulation andEmulab emulation based on real taxitraces. They further deploy VCD in twovehicular networks, one using 802.11band the other using 802.11n, to demon-strate its effectiveness.

SC A N N I N G T H E L I T E R AT U R E

Are you enthusiastic? Have you performed quality reviews for technical periodicals?

Demonstrated solid technical accomplishments? Have a reputation for high ethical standards and for reliability?

You may be ready ...

The IEEE Communications Society is looking for volunteers who are interested in becoming part of a prestigious Communications Society editorial board.

Duties include: A commitment to handle at least two manuscripts per month; arrange for three reviews or more in a timely fashion; and the ability to make firm and fair decisions.

Qualifications: Subject matter expertise, editing experience, technical program committee experience; references, representative papers.

Apply at: www.comsoc.org/editor

The decision to appoint an editor rests with the Editor-in-Chief of the journal/magazine. Please note that it will not be possible to send individual acknowledgments to all applicants.

Open Call from the IEEE Communications Society

www.comsoc.org/editor

LYT-SCANNING-February 2/7/11 10:40 AM Page 5

IEEE Wireless Communications • February 20116 1536-1284/11/$25.00 © 2011 IEEE

Fixed network

GW

(3) MGC

AC C E P T E D F R O M OP E N CALL

INTRODUCTIONIP Multimedia Subsystem (IMS) supports IP-based multimedia services. IMS was originallydesigned by the Third Generation PartnershipProject (3GPP) to deliver Internet services overgeneral packet radio service (GPRS) in 3G net-works such as Universal Mobile Telecommunica-tions System (UMTS). IMS was later updated tosupport other access networks including wirelessLAN, CDMA2000, and fixed line. For the pur-pose of this article, Fig. 1 illustrates a simplifiedIMS network architecture (the reader is referredto [1] for detailed descriptions of IMS).

The IMS (Fig. 1b) connects to both mobileand fixed telecommunications networks (Fig. 1a)for fixed mobile convergence (FMC). IMS is notintended to standardize applications or services.Instead, it provides a standard approach forvoice/multimedia application access from userequipment (UE; 1, Fig. 1) in wireline and wire-less networks. This goal is achieved by havinghorizontal control that isolates the access net-works from the service and application networks(Fig. 1c).

In the IMS, the transport of user data is sepa-rated from that for control signals, where IETFprotocols such as Session Initiation Protocol(SIP) [2] are used to ease the integration withthe Internet. For example, the call session con-trol function (CSCF, 5, Fig. 1) is a SIP server,which is responsible for call control. The mediagateway control function (MGCF, 3, Fig. 1) con-

trols the connection for media channels in amedia gateway (MGW, 4, Fig. 1). The MGWconnects toward the legacy fixed and mobile net-works to provide user data transport. The homesubscriber server (HSS, 2, Fig. 1) is the masterdatabase containing all user-related mobile sub-scription and location information. In 2004Chunghwa Telecom deployed the first commer-cial IMS in Taiwan to provision commercialtelecommunications services such as voice, video,and Internet-based multimedia services. The ini-tial capacity was 125,000 subscribers. CurrentIMS capacity can accommodate about 500,000subscribers in daily commercial operation.

During Typhoon Morakot in August 2009,Taiwan experienced serious damage from flood-ing and mudslides (Fig. 2, left), and the rescuemissions solely relied on GSM and satellite com-munications that offer basic services (Fig. 2, right)such as emergency call without location tracking.From Typhoon Morakot, we learned that it isdesirable to accommodate emergency services inIMS with a 3G network including emergency calland push-to-talk over cellular (PoC).

A GSM user in Taiwan can make an emer-gency call by dialing 110, 112, or 119. However,the existing GSM emergency call service onlyidentifies the location of the caller at the time ofcall setup and does not track the user’s locationduring the call. In Typhoon Morakot many peo-ple waiting for rescue could not be accuratelylocated through their phone calls, whichincreased the difficulty of the rescue missions.

PoC is a walkie-talkie-like service defined inOpen Mobile Alliance (OMA) specifications [3,4]. In 2004 Chunghwa Telecom first launchedthis service in Asia using 2.5G technology. Ourexperience with 2.5G PoC included long PoCcall setup time (13 s) and handoff time (2.5–3 sfor reconnecting a PoC client when it movedfrom one base station to another). These prob-lems have been resolved by IMS-based PoCestablished on 3G networks. For example, the3G PoC call setup time is less than 6 s. Althoughit is not clear if PoC will be a successful residen-tial service, it has proven useful for business cor-porations and government organizations such asthe National Security Bureau in Taiwan.

YI-BING LIN, NATIONAL CHIAO TUNG UNIVERSITYMENG-HSUN TSAI, NATIONAL CHENG KUNG UNIVERSITY

YUAN-KUANG TU, CHUNGHWA TELECOM

ABSTRACT

Emergency call and walkie-talkie are two ser-vices utilized in emergency situations. DuringTyphoon Morakot in 2009, we experienced thedeficiency of emergency call service that cannotcontinuously track callers in real time andwalkie-talkie communications where a speakermay not be granted the permission to talk. Theseissues can be resolved by the emergency call andpush-to-talk over cellular services in IP Multime-dia Subsystem. This article conducts a prelimi-nary study on how these two services can beeffectively exercised in IMS.

IMS EMERGENCY SERVICES:A PRELIMINARY STUDY

Deficiencies in communications during emergenciescan be resolved bythe emergency calland the Push-to-talkover Cellular servicesin IP MultimediaSubsystem network.The authors conducta preliminary studyon how these twoservices can be effectively exercisedin IMS.

LIN LAYOUT 2/7/11 12:04 PM Page 6


In the past three years, we have studied emer-gency call [5] and PoC [6]. During TyphoonMorakot in 2009, emergency calls and walkie-talkies were important means of communicationsin rescue missions. Clearly, it is desirable to sup-port emergency services in IMS for disastrousevents. Therefore, this article conducts a prelimi-nary investigation on IMS emergency call and PoC.

LOCATION TRACKING FOR EMERGENCY CALL

An important feature of emergency call is thatthe system can track the location of a calling UE(1, Fig. 1) during the conversation. To supportIMS emergency call, three network nodes aredeployed. When the UE originates an emergen-cy call, the call is established by a special CSCFcalled an emergency-CSCF (E-CSCF, 5, Fig. 1),which dispatches the call to the nearest publicsafety answering point (PSAP) based on thelocation information of the UE. The PSAP (7,Fig. 1) is an IMS application server that process-es emergency calls according to the types ofemergency events. For example, in a fire eventthe PSAP connects the UE (the caller) to thefire department (8, Fig. 1). The PSAP interactswith the E-CSCF by using SIP, and the voiceconversation path is set up through the MGW(4, Fig. 1) to the UE by using Real-Time Trans-port Protocol (RTP) [7]. The gateway mobilelocation center (GMLC, 6, Fig. 1) supports alocation service (LCS) [8]. Through SignalingSystem Number 7 (SS7) Mobile Application Part(MAP) [1], the GMLC interacts with the HSSand mobile network to obtain the accurate loca-tion of UE. The GMLC provides the locationinformation to the PSAP and E-CSCF.

The LCS merits further discussion. This ser-vice utilizes one or more positioning methodsbetween the mobile network (Fig. 1a) and UE todetermine the location of the UE [9]. The cell-ID-based positioning method determines theUE’s position based on the coverage of serviceareas (SAs). An SA includes one or more cells(base stations). The observed time difference ofarrival (OTDOA) and uplink time difference ofarrival (U-TDOA) positioning methods utilizetrilateration to determine the UE’s positionbased on the time differences between downlinkand uplink signal arrivals, respectively. TheAssisted Global Positioning System (A-GPS)method speeds up GPS positioning by down-loading GPS information through the mobilenetwork. In Chunghwa Telecom A-GPS is uti-lized for location-based services.

Without loss of generality, we consider thecell-ID-based method. After emergency callsetup, the PSAP may need to monitor the UE’slocation in real time. In the 3GPP 23.167 specifi-cation [10], the UE’s location is tracked throughpolling, where the PSAP periodically queries theUE’s location. For description purposes, werefer to the 3GPP 23.167 approach as the loca-tion polling scheme. In this scheme, if the UEdoes not change its location between twoqueries, the second query is wasted (this is calledredundant polling). On the other hand, if the UEhas visited several locations between two loca-tion queries, the PSAP may lose track of the UEin this time period (this is called mistracking). To

resolve these issues, the active location reportingscheme was proposed in [5]. This scheme reportsthe UE’s location upon change of its SA. Thissection describes location polling and activelocation reporting, and comments on their per-formance.

EMERGENCY CALL SETUPFigure 3 illustrates IMS emergency call setupmessage flow with the following steps [10]:Step A.1 The UE establishes IP connectivity to

the IMS through the mobile network [1].Step A.2 The UE sends a SIP INVITE message

to the E-CSCF. This message includes thesupported positioning methods of the UE(cell-ID-based in our example).

Step A.3 The E-CSCF uses the received infor-mation to select a GMLC and sends the Emer-gency Location Request message to theGMLC.

Steps A.4 and A.5 The GMLC exchanges the SS7MAP_SEND_ROUTING_INFO_FOR_LCS andacknowledgment message pair with the HSSto identify the mobile network node responsi-ble for connection to the UE. In UMTS thisnode is a serving GPRS support node (SGSN).

Step A.6 The GMLC sends the SS7 MAP_PRO-VIDE_SUBSCRIBER_LOCATION message.

Step A.7 The mobile network and UE exercisethe cell-ID-based positioning procedure toobtain the location estimate information ofthe UE (i.e., the SA identity of the UE).

Step A.8 The mobile network returns the SA iden-tity to the GMLC through SS7 MAP_PRO-VIDE_SUBSCRIBER_LOCATION_ack message.

Step A.9 The GMLC selects a suitable PSAPaccording to the SA of the UE and replies theEmergency Location Response message (withthe selected PSAP address) to the E-CSCF.

Figure 1. A simplified IMS network architecture (dashed lines: control signal-ing; solid lines: user data/control signaling): a) fixed and mobile telecom net-works; b) IP Multimedia Subsystem; c) service and application networks.

Fixed network

Mobile network

(7) PSAP

(8) Fire department

Police department (6) GMLC

(9) PoC server

(4) MGW

(5) CSCF

(3) MGCF (2) HSS

(1) UE

(b)

(a)

(c)



Steps A.10–A.12 The E-CSCF forwards the SIPINVITE to the PSAP to set up the call. ThePSAP and the UE exchange the 200 OK andthe SIP ACK messages through the E-CSCF.After the PSAP has received the SIP ACKmessage, the emergency call is established.

Step A.13 The GMLC sends the location infor-mation obtained at step A.8 to the PSAP afterthe call has been established, where the PSAPaddress is resolved by the GMLC at step A.9.

LOCATION POLLINGUE may move during an emergency call, and thePSAP needs to monitor the UE’s location in realtime. In location polling, the PSAP periodicallyqueries the UE’s location. In each query, the fol-lowing steps are executed (Fig. 4) [8]:Step B.1 The PSAP sends the Location Informa-

tion Request message to the GMLC.Steps B.2–B.6 These steps retrieve the UE’s SA

identity, which are similar to steps A.4–A.8 inFig. 3.

Step B.7 The GMLC returns the SA identity ofthe UE to the PSAP.

Steps B.8–B.10 When the emergency call isterminated, the E-CSCF exchanges the Emer-gency Location Release and Response mes-sage pair with the GMLC to terminatelocation tracking.

ACTIVE LOCATION REPORTINGTo resolve redundant polling and mistrackingissues in location polling, the active locationreporting scheme was proposed in [5], whichreports the UE’s location upon change of its SA.This scheme introduces a new locationEstimate-Type initiateActiveReport (to trigger active loca-tion reporting) in the MAP_PROVIDE_SUBSCRIBER_LOCATION message (at step A.6).Since the IP connectivity exists during the IMSemergency call, the UE is in the cell-connectedstate and is tracked by the mobile network at thecell level [1]. Therefore, the mobile network candetect when the UE moves from one base stationto another, and report the new SA identity to theGMLC. In this approach the GMLC maintains aUE-PSAP mapping table to store the (UE,PSAP) pair at step A.9. The GMLC does not

need to query the HSS to identify the mobilenetwork node responsible for connection to theUE (i.e., steps B.2 and B.3 are eliminated). Theactive location reporting scheme is illustrated inFig. 5 with the following steps:Step C.1 When the UE moves to a new SA, the

mobile network detects this movement at thecell tracking mode and then triggers the posi-tioning procedure.

Step C.2 After the positioning procedure is exe-cuted, the UE’s SA identity is obtained.

Step C.3 The mobile network sends the SS7MAP_SUBSCRIBER_LOCATION_REPORTmessage with the SA identity to the GMLC.

Step C.4 From the UE-PSAP mapping table, theGMLC retrieves the PSAP address of the UEstored at step A.9 and then sends the updatedlocation information to the PSAP.When the emergency call is terminated, the

following steps are executed:Step C.5 When the IMS call is released, the UE

moves from the cell-connected mode to theidle mode, and the mobile network no longertracks the movement of the UE.

Step C.6 The E-CSCF sends the EmergencyLocation Release message to the GMLC toterminate location tracking.

Step C.7 The GMLC returns the EmergencyLocation Response message to the E-CSCFand then deletes the (UE, PSAP) mappingfrom the UE-PSAP table.Note that steps C.1 and C.2 in active location

reporting automatically detects the movement ofUE, which is different from steps B.1–B.5 inlocation polling.

PRELIMINARY PERFORMANCE EVALUATIONIt is clear that redundant polling creates extranetwork traffic without providing useful locationinformation. Furthermore, mistracking mayresult in wrong positioning in case of emergencysituations. These issues are resolved by activelocation reporting. However, it is desirable toevaluate the performance of location polling tojustify the modifications to the existing locationtracking procedure in active location reporting.

Suppose that the SA residence time has aGamma distribution with mean 1/μ and variance

Figure 2. Telecommunications services in Typhoon Morakot: (left) deploying temporary cables in floodedareas; (right) GSM/satellite communications through a vehicular base station.

Redundant pollingcreates extra network traffic without providinguseful location infor-mation. Furthermore,mis-tracking mayresult in wrong positioning in case of emergency situations. Theseissues are resolvedby Active LocationReporting.



Vm (other distributions have shown similarresults [5]). The inter-query interval is a fixedperiod 1/λ in location polling. Several outputmeasures are studied. Let α be the mistrackingprobability. An SA crossing is mistracked ifthere is no query between this and the next SAcrossings (i.e., the system does not know that theuser has visited this SA). Let β be the probabilitythat redundant queries exist between two SAcrossings. The larger the α or β values, the worsethe performance of location polling.

Figure 6a shows intuitive results that as thepolling frequency λ increases, α decreases and βincreases. We describe two effects of Vm:• Effect 1: When the SA residence time intervals

become more irregular (i.e., Vm increases),more SA residence time intervals without anyquery are observed.

• Effect 2: When Vm increases, if a query arrivesin an SA residence time interval, more thanone query will tend to arrive in this interval.Effect 1 implies that as Vm increases, more

SA crossings are mistracked, and we have a non-trivial observation that α increases as Vm increas-

es. Similarly, when λ is large (e.g., λ ≥ 5 μ),effect 1 is significant, and β is a decreasing func-tion of Vm. The impact of Vm on β is more sub-tle when λ = μ. In this case β increases and thendecreases as Vm increases, which implies thatwhen Vm is small, effect 2 is more significant. Onthe other hand, when Vm is large, effect 1 domi-nates. An important observation is that when0.1/μ2 < Vm < 10/μ2, both α and β values arenon-negligibly large, and poor performance oflocation polling cannot be ignored.

Besides mistracking and redundant pollingprobabilities, we would also like to investigatethe following output measures:• Ti: The expected period in which the PSAP

cannot correctly track the UE’s location (i.e.,Ti is the period between an SA crossing andwhen the next query arrives). In this periodthe system does not know the user’s correctlocation.

• NR: The expected number of redundantqueries between two SA crossings (i.e., NR isthe number of queries issued within two con-secutive SA crossings).

Figure 3. IMS emergency call setup.

PSAP GMLC

A.3. Emergency Location Request

A.2. SIP INVITE

A.4. MAP_SEND_ROUTING_INFO_FOR_LCS

A.5. MAP_SEND_ROUTING_INFO_FOR_LCS_ack

A.6. MAP_PROVIDE-SUBSCRIBER_LOCATION

A.8. MAP_PROVIDE-SUBSCRIBER_LOCATION_ack

A.9. Emergency location response

A.10. SIP INVITE

A.12. SIP ACK

A.13. Location information

A.11. 200 OK

A.12. SIP ACK

A.11. 200 OK

E-CSCF HSS

Mobilenetwork

UE

A.1. IMS connection establishment

A.7. Positioning

Report initial location

Establish emergency call

In the PoC serviceseveral predefined

group members participate in one

PoC session. SincePoC utilizes half-duplex

communications,only one PoC

member speaks at a time,

and others listen.



The larger the above output values, the worsethe performance of location tracking. Figure 6bplots Ti against λ and Vm (the solid curves),where Ti is measured by 1/μ. Ti is a decreasingfunction of λ and is not affected by Vm. Wenotice that for λ < 10 μ, the PSAP will mistrackUE more than 5 percent of the time, which maynot be acceptable in emergency situations.

It is clear that when λ > μ, NR⋅=⋅ (λ/μ) – 1.

Let N*R be the number of redundant queries

between two SA crossings under the conditionthat there is at least one redundant query inthis interval. Figure 6b plots N*

R (the dashedcurves). Clearly, N*

R is an increasing function ofλ. When Vm < 1/μ2, N*

R⋅=⋅ NR. When Vm > 1/μ2,

N*R significantly increases as Vm increases. This

phenomenon implies that when the UE’s move-ment pattern becomes irregular, the PSAP willreceive many redundant (useless) locationreports after it has received a correct locationupdate.

PUSH-TO-TALK OVER CELLULAR

In the PoC service several predefined groupmembers participate in one PoC session. SincePoC utilizes half-duplex communications, onlyone PoC member speaks at a time, and otherslisten. When a PoC member attempts to speak,

he/she presses the push-to-talk button of his/herUE (Fig. 1 (1)) to ask for permission. This UEwith the PoC application installed is called thePoC client. The PoC server (9, Fig. 1) is responsi-ble for handling PoC session management (cre-ate or delete a PoC session). It arbitrates speakpermission through the talk burst control mecha-nism [4]. SIP and Session Description Protocol(SDP) [11] are utilized for session establishment.After the PoC session is established, each of thePoC clients has built an RTP session with thePoC server through the MGW (Fig. 1 (4)). Thetalk burst control messages between the PoCclients and the PoC server are carried out byReal-Time Transmission Control Protocol(RTCP) packets [7].

In SIP a new parameter, tb_grant, is added inSDP’s attribute field so that the PoC server canarbitrate the speak permission during a session.If tb_grant = 1, the PoC client is granted thepermission to speak. Otherwise (i.e., tb_grant =0), the PoC client is not permitted to talk.

In emergency situations, important messagesmay not be delivered via walkie-talkie-like com-munications if the message sender cannot obtainpermission to talk. Therefore, a mechanism isdesirable to guarantee that a PoC client has afair chance to talk. This issue can be resolved byPoC with queueing option.

Figure 4. Location polling.

PSAP GMLC E-CSCF HSS

Mobilenetwork

UE

B.5. Positioning

Report current location

Call setup

B.8. Emergency call release

B.1. Location Information Request

B.7. Location information

B.2. MAP_SEND_ROUTING_INFO_FOR_LCS

B.3. MAP_SEND_ROUTING_INFO_FOR_LCS_ack

B.4. MAP_PROVIDE_SUBSCRIBER_LOCATION

B.6. MAP_PROVIDE_SUBSCRIBER_LOCATION_ack

B.9. Emergency location release

B.10. Emergency location response

In emergency situa-tions, important mes-sages may not bedelivered in walkie-talkie like communi-cations if themessage sender can-not obtain the per-mission to talk.Therefore, a mecha-nism is desirable toguarantee that a PoCclient has a fairchance to talk. Thisissue can be resolvedby PoC with queuingoption.



TALK BURST CONTROL MECHANISM

The talk burst control mechanism is implement-ed by finite state machines (FSMs) on both theserver and client sides. Figure 7 illustrates sim-plified talk burst control FSMs for PoC server(called FSMG) and PoC client (called FSMU).For every PoC session, there is one FSMG in thePoC server and an FSMU in each of the PoCclients. Two timers are defined in FSMG. TimerT2 is used to determine whether the PoC clientspeaks too long. If a PoC client talks longer thanthe T2 period, he/she is asked to release the per-mission. Timer T3 is used to gracefully terminatethe talk burst. After the PoC client gives up per-mission, the transient RTP packets it generatedin T2 are continuously forwarded by the PoCserver in the T3 period. When T3 expires, thePoC server allows the next PoC client to talk.

Session Initiation — For the PoC clients and PoCserver involved in a PoC session, their FSMs areinitialized at the start-stop state. The session ini-tiator (a PoC client) starts a PoC session bysending a SIP INVITE message to the PoC serv-er. The PoC server then broadcasts SIP INVITEmessages with tb_grant = 0 to the invited PoCclients (i.e., PoC group members other than thesession initiator). Each of the invited PoC clientsanswers with a SIP 200 OK message, and itsFSMU enters has no permission (transition 2, Fig.7b). This PoC client is not permitted to speak.

After receiving the first SIP 200 OK messagefrom each of the invited PoC clients, the PoCserver replies a SIP 200 OK with tb_grant = 1 to

the session initiator, and FSMG enters TB_Taken(transition 1, Fig. 7a). This state means thatsome PoC client (the session initiator in thiscase) has obtained the permission. FSMU of thesession initiator enters has permission (transition1, Fig. 7b), and the PoC client is allowed tospeak. The session initiator becomes the permit-ted PoC client (i.e., the PoC client allowed tospeak), and the invited PoC clients become listen-ing PoC clients (i.e., the PoC clients not permit-ted to speak). At this moment, the sessioninitiator speaks, and all invited PoC clients listen.

Permission Releasing — After finishing the talk, thepermitted PoC client X releases the permissionby sending the TB_Release message to thePoC server, and its FSMU enters pendingTB_Release (Transition 6 in Fig. 7b). In thisstate, client X stops sending media packets andwaits for the response from the PoC server.FSMG enters pending TB_Release after the PoCserver has received the TB_Release message(transition 2, Fig. 7a). In this state the PoC serv-er keeps forwarding the transient media packetsissued from client X before the TB_Releasemessage. When the last transient media packethas been processed or T3 expires, FSMG entersTB_Idle (transition 3, Fig. 7a). In this state noPoC client is granted permission to speak. ThePoC server broadcasts the TB_Idle message toall PoC clients. FSMU of client X enters has nopermission upon receipt of the TB_Idle mes-sage (transition 7, Fig. 7b). A listening PoCclient remains in has no permission when itreceives the TB_Idle message (Transition 14 in

Figure 5. Active location reporting.

PSAP GMLC E-CSCF

Mobilenetwork

UE

C.2. Positioning

C.3. MAP_SUBSCRIBER_LOCATION_REPORT

C.6. Emergency location release

C.4. Location information

C.7. Emergency location response

C.1. Change of dervice areaReport current location

Call setup

C.5. Emergency call release

A queueing option isprovided in the talk

burst control mechanism. If thisoption is selected,

then the ungrantedrequests are buffered

in the queue at thePoC server

instead of beingdenied.



Fig. 7b). At this point, all PoC clients can com-pete for permission to speak.

Permission Requesting — To obtain permission, alistening PoC client X sends the TB_Requestmessage to the PoC server. Client X becomes arequesting PoC client (for speak permission), andits FSMU enters pending TB_Request (transition3, Fig. 7b). This state indicates that client X iswaiting for arbitration from the PoC server. Ifsome other PoC client has been granted permis-sion, the PoC server sends the TB_Deny messageto client X, and FSMU of client X moves back tohas no permission (transition 4, Fig. 7b). If thePoC server grants permission to client X, it sendsthe TB_Granted message to client X and theTB_Taken message to other PoC clients. FSMGenters TB_Taken (transition 4, Fig. 7a), andFSMU of client X enters has permission (transi-tion 5, Fig. 7b). When a listening PoC clientreceives TB_Taken, its FSMU remains at has nopermission, and it is not allowed to request per-mission. After client X has become the permittedPoC client, T2 at the PoC server is started. Thistimer is used to monitor if this client X speakstoo long (and therefore should be revoked).

Permission Revoking — If permitted client X speakslonger than the T2 period, the PoC server willsend the TB_Revoke message to reclaim thepermission and start the T3 timer. Upon receiptof the TB_Revoke message, FSMU of client Xenters pending TB_Revoke (transition 8, Fig. 7b).FSMG enters pending TB_Revoke (transition 5,Fig. 7a). In this state the PoC server keeps for-warding transient media packets of client X untilT3 expires. Then FSMG enters TB_Idle (transi-tion 6, Fig. 7a). The PoC server sends theTB_Idle message to all PoC clients. FSMU ofclient X enters has no permission (transition 9,Fig. 7b) upon receipt of the TB_Idle message.

When a listening PoC client receives TB_Idle,its FSMU remains at has no permission. At thispoint, all PoC clients can compete for permis-sion to speak.

Permission Queuing — A queueing option is provid-ed in the talk burst control mechanism. If thisoption is selected, the ungranted requests arebuffered in the queue at the PoC server insteadof being denied. In this option, after the permit-ted PoC client finishes talking, the PoC servergrants the next request from the queue. Thestate queued in FSMU (dark oval, Fig. 7b) indi-cates that a request of the client is buffered inthe PoC server and will be granted later.

After PoC client X has obtained permission,the PoC server may receive the TB_Requestmessage from another requesting PoC client Y.With the queueing option, FSMG is in theTB_Taken state, and FSMU of client Y is in thepending TB_Request state. The PoC server buffersthe TB_Request message in the queue andreplies with the TB_Queued message to client Y.FSMG stays in TB_Taken, and FSMU of client Yenters queued (transition 10, Fig. 7b). Client Y iscalled a queued PoC client. If a queued PoCclient is not patient, it will send the TB_Releasemessage to the PoC server to cancel the request,and its FSMU moves back to has no permission(transition 11, Fig. 7b). In this case, the PoCserver removes the corresponding request fromthe queue. The period between when a PoCclient enters the queued state and when it entersthe has no permission state is called the patienttime. After the permission is released (orrevoked), FSMG will enter TB_Idle, and FSMUof the permitted PoC client will enter pendingTB_Release (transition 6, Fig. 7b) or pendingTB_Revoke (Transition 8, Fig. 7b). If the queue isnot empty, the PoC server grants permission tothe next queued request instead of sending the

Figure 6. Emergency performance measures: a) α and β; b) Ti (unit: 1/μ) and N*R.

Vm (unit: 1/ 2)

10-2

0.2

0.0

0.4

0.6

0.8

1.0

10-3 10-1 100 101 102 103

Vm (unit: 1/μ2)

(a) (b)

10-1

100

101

102

10-2

103

10-3 103 102 101 100 10-1 10-2

Solid: α;Dashed: βλ=μλ=5μλ=10μ

Solid: Ti;Dashed: N*

Rλ=μλ=5μλ=10μ



TB_Idle message. The next queued PoC clientwill receive the TB_Granted message from thePoC server, and its FSMU enters has permission(transition 12, Fig. 7b). This queued PoC clientbecomes the next permitted PoC client. At thesame time, other PoC clients receive theTB_Taken message from the PoC server. ClientX becomes a listening PoC client, and its FSMUenters has no permission (transition 7 or 9, Fig.7b). FSMU of every listening PoC client remainshas no permission, and FSMU of every queuedPoC client remains queued.

Session Termination — When a PoC client leavesthe PoC session, its FSMU moves back to start-stop (transition 13, Fig. 7b). The PoC sessionremains active for other PoC clients. After allPoC clients have left the PoC session, the ses-sion is implicitly terminated. FSMG moves backto start-stop (transition 7, Fig. 7a).

PERFORMANCE EVALUATIONBased on [4], we investigate the performance ofthe PoC talk burst control mechanism withqueueing (approach Q) and without queueing(approach NQ). Define the revoking timer as TR= T2 + T3. Let the permission request arrivalsof a PoC client form a Poisson process with rateλ. The speak time is a random variable withmean 1/μ. The patient time is a random variablewith mean 1/ω. Both speak and patient times areassumed to be Exponentially distributed. Otherdistributions show similar results, and the readeris referred to [6] for details. Let M be the num-ber of PoC clients in a PoC group. In ChunghwaTelecom’s commercial PoC service, M is limitedto 20 (i.e., at most 20 members can participate ina PoC session). In this study we assume that 5 ≤M ≤ 40. Two output measures are considered:• PD: The denied probability that the request of

a PoC client is not granted because this clientis not patient in approach Q or the PoC serv-er rejects the request in approach NQ.

• W: The expected waiting time between when aPoC client requests permission to speak andwhen it is granted permission under the condi-

tion that the PoC client is not immediatelygranted permission. The W measure excludesthe waiting times of immediately grantedrequests (which are 0) so that we can answerthe question “If you have to wait, how longwill you wait?”Under the conditions that λ = 0.01μ and TR

= 3/μ, Fig. 8a indicates that for the same PD per-formance, approach Q can support at least twiceas many clients as approach NQ. For example, tomaintain PD = 0.038, approach NQ can only sup-port M = 5, while approach Q can support M =10 (for ω = μ) and M = 40 (for ω = 0.1μ).

Figure 8b illustrates the expected waitingtime performance (WQ for approach Q and WNQfor approach NQ) before a PoC client is grantedpermission to speak (if he/she is not immediatelyaccepted by the PoC server). To make a faircomparison, we set ω = 0 for approach Q sothat WQ will not be shortened by impatience.For approach NQ, when a PoC client fails toobtain permission, it keeps requesting until per-mission is granted. Therefore, WNQ is the periodbetween its first try and when it is granted per-mission. Figure 8b plots WQ and WNQ (measuredby 1/μ) against M. The figure indicates that WQis much shorter than WNQ. Furthermore, even ifthe number of PoC participants is large (e.g., M= 40), WQ is reasonably short (less than 2.5/μ).

CONCLUSIONS

This article has conducted a preliminary investi-gation of two emergency services for IMS: emer-gency call and push-to-talk over cellular (PoC).In the IMS emergency call study we observe thatthe existing location tracking mechanism (calledlocation polling) might mistrack the caller andcause unnecessary signaling overhead. Wedescribe a modified mechanism called activelocation reporting, which may potentiallyenhance the performance of location polling.Based on this conclusion, most location-basedservices in Chunghwa Telecom have utilized anactive-location-reporting-like mechanism. In thefuture this mechanism will be provided in IMS.

Figure 7. Talk burst control finite state machines for: a) PoC server (FSMG); b) PoC client (FSMU).

10

3

Start-stop

TB_Taken

TB_Idle

(a) (b)

pendingTB_Release pending

TB_Revoke

7

4

5

1

2

3 6

Any State

has no permission

pending TB_Request pending TB_Revoke

pending TB_Release 12

5

has permission

Start-stop 13

2 14

1

9

7

11

6 8

4

Any State

queued

Our study indicatesthat PoC with queue-ing can comfortablyaccommodate twice

as many participantsas PoC without

queueing, where thewaiting time for a

PoC participant to begranted

permission is reasonable short.



For PoC, to guarantee that everyone canalways obtain permission to talk in emergencysituations, the queueing option of PoC should beselected. Our study indicates that PoC withqueueing can comfortably accommodate twice asmany participants as PoC without queueing,where the waiting time for a PoC participant tobe granted permission is reasonably short. In thefuture, we will also consider the trade-offbetween priority and fairness in the queueingmechanism.

In summary, our preliminary study providesguidelines to deploy emergency call and PoC ina commercial IMS network.

ACKNOWLEDGMENTY.-B. Lin’s work was supported in part by NSC97-2221-E-009-143-MY3, NSC 098001228630,NSC 97-2219-E-009-016, Intel, Chunghwa Tele-com, ITRI, NCTU joint research center, andMoE ATU plan. M.-H. Tsai’s work was support-ed by NSC 100-2218-E-006-015-MY2.

REFERENCES[1] Y.-B. Lin and A.-C. Pang, Wireless and Mobile All-IP Net-

works, Wiley, 2005.[2] IETF RFC 3261, “SIP: Session Initiation Protocol,” 2002.[3] OMA, “Push to Talk over Cellular (PoC) — Architecture,”

OMA-AD-PoC-V2 0 1-20080226-C Candidate Version2.0, Feb. 26, 2008.

[4] OMA, “Push to Talk over Cellular (PoC) — User Plane,”OMA-TS-PoC-V2 0 1-20080226-C Candidate Version2.0, Feb. 26, 2008.

[5] M.-H. Tsai, Y.-B. Lin, and H.-H.Wang, “Active LocationReporting for Emergency Call in UMTS IP MultimediaSubsystem,” IEEE Trans. Wireless Commun., vol. 8, no.12, 2009, pp. 5837–43.

[6] M.-H. Tsai and Y.-B. Lin, “Talk Burst Control for Push-to-Talk over Cellular,” IEEE Trans. Wireless Commun., vol.7, no. 7, 2008, pp. 2612–18.

[7] IETF RFC 3550, “RTP: A Transport Protocol for Real-TimeApplications,” 2003.

[8] 3GPP TS 23.271, “Functional Stage 2 Description ofLocation Services (LCS),” v. 7.9.0, 2007.

[9] 3GPP TS 25.305, “Stage 2 Functional Specification ofUser Equipment (UE) Positioning in UTRAN,” v. 7.4.0,2007.

[10] 3GPP TS 23.167, “Internet Protocol (IP) based IP Multi-media Subsystem (IMS) Emergency Sessions,” v. 7.11.0,2008.

[11] IETF RFC 4566, “SDP: Session Description Protocol,”2006.

BIOGRAPHIESYI-BING LIN [M‘95, SM‘95, F‘03] ([email protected]) is thedean and chair professor of the College of Computer Science,National Chiao Tung University, Taiwan. He is also an adjunctresearch fellow of the Institute of Information Science,Academia Sinica, Nankang, Taipei, Taiwan, and a consultantprofessor of Beijing Jiaotong University, China. His currentresearch interests include wireless communications andmobile computing. He has published over 240 journal articlesand more than 200 conference papers. He is a co-author ofthe books Wireless and Mobile Network Architecture withImrich Chlamtac (Wiley), Wireless and Mobile All-IP Networkswith Ai-Chun Pang (Wiley), and Charging for Mobile All-IPTelecommunications with Sok-Ian Sou (Wiley). He is an ACMFellow, an AAAS Fellow, and an IET (IEE) Fellow.

MENG-HSUN TSAI [S‘04, M‘10] ([email protected])received B.S., M.S., and Ph.D. degrees from NCTU, Hsinchuin 2002, 2004, and 2009, respectively. He joined theDepartment of Computer Science and Information Engi-neering, National Cheng Kung University, Tainan, Taiwan,as an assistant professor in 2010. His current researchinterests include design and analysis of personal communi-cations services networks, mobile computing, and perfor-mance modeling.

YUAN-KUANG TU [M‘90] joined Chunghwa Telecom in 1981,working in a variety of R&D and management positions inthe Telecommunication Laboratories, and served as VicePresident in 2006. In 2007 he was assigned senior manag-ing director of the Corporate Planning Department inHeadquarters, in charge of strategic planning and newbusiness development. In May 2009 he was promoted topresident of Chunghwa Telecommunication Laboratories.He has published over 60 journal papers and holds 20patents. His professional fields include optical communica-tions, broadband networks, and mobile networks.

Figure 8. Effects on PD, WQ, and WNQ (TR = 3/μ): a) PD (λ = 0.01μ); b) WQ and WNQ (unit: 1/μ), (ω = 0).

M

(a)

10

0.05

0.00

0.10

0.15

0.20

0.25

0.30

5 15 20 25 30 35 40M

(b)

15

101

100

102

10-1

10 20 25 30 35 5 40

Solid: Q;Dashed: NQω=μω=0.1μ

Solid: WQ;Dashed: WNQλ=0.02μλ=0.01μλ=0.005μ



R

Traditional routing

R


INTRODUCTIONIn the past decade network coding (NC) [1] andcooperation [2] have emerged as two techniquesthat may deeply impact wireless networks. Theytackle different issues and show complementaryproblems. In the former, packets are allowed tobe combined together at intermediate nodes soas to reduce the redundancy needed to reliablydeliver them. In other words, nodes collaborateto reduce the overhead. In the latter family ofstrategies, nodes cooperate by relaying the samedata from different terminals so as to protectinformation with spatial diversity and improvethe reliability and overall performance. Theyboth aim to deliver information as efficiently aspossible, but by different means. We remark thatin this article cooperation is considered in thedomain of physical layer relaying, in the sense of[2], with focus on the improvement of error cor-rection capabilities. Hence, although in NetworkCoding nodes also “cooperate” (i.e., they shareinformation to reach a common goal), the defini-tion of cooperation we adopt in this article doesnot include Network Coding as a special case.

It comes at no surprise that there has been arecent surge in the efforts to bring these twoapproaches together. However, three mainphilosophies stand out in this pursuit. In the firstone, network coding is tightly coupled with chan-nel coding (given the ties of channel coding toboth the discussed systems). In the second

approach (spurred by the groundbreaking con-cept of physical layer network coding [PNC] [3,14]), packets are linearly combined on the wire-less channel by means of collisions. Finally, inthe third branch, NC is made more robust bymeans of techniques drawn from multiple-inputmultiple-output (MIMO).

The purpose of this article is to summarizethe main discoveries in this field and point outthe first design guidelines that are emerging. Weoutline the results for channel/network coding,PNC, and MIMO_NC, respectively. The con-cluding remarks are then drawn.

HYBRID CHANNEL/NETWORK CODING

Before any discussion, it is insightful to introducethe reference topology of this article, because vir-tually all interesting aspects and issues arise in itsanalysis. Such a scenario is depicted in Fig. 1: aset of P sources R1, …, SP send their data to a setof destinations D1, …, DK. They are supported byM relays S1, …, RM, who send redundancy basedon the signals received from S1, …, SP. Weremark that these three groups of nodes (sources,relays and destinations) are not necessarily dis-joint. For instance, some sources may be destina-tions for other sources. In this model, P sourcesgenerate one new frame each and a total of N =P + M transmissions are performed. Note thatNC is known to bring substantial benefits whendifferent flows must be routed through the sameterminals, as is indeed the case in Fig. 1, sincethe relays are shared by all sources. Such a sce-nario is rather general and applies to a variety ofcontexts. For instance, it is representative of theuplink of a cellular system, where there is only K=1 destination D1 (i.e., the base station). Hybridchannel/NC has chiefly been applied in this envi-ronment, whose underlying goal is to improve theefficiency of conventional cooperative protocolsso that fewer relays are needed to achieve a cer-tain degree of reliability; conversely, higher per-formance can be attained with the same amountof redundancy. Instead, in physical layer NC thetopology is another instance of Fig. 1 with multi-ple destinations, but again the flows are con-strained to go through a common set of relays.

Channel coding is intimately related to coop-

FRANCESCO ROSSETTO, DLR (GERMAN AEROSPACE CENTER)MICHELE ZORZI, UNIVERSITY OF PADOVA

ABSTRACT

Cooperation and network coding are twopowerful techniques for wireless networks. Dur-ing the past few years a surge of research activityhas tried to combine the best of these twophilosophies to improve the performance anderror correction capabilities of radio networks.This article gives an overview of the literature inthe field, summarizes the main achievementsand design guidelines found so far, and finallyhighlights the main challenges yet to be solved.

MIXING NETWORK CODING AND COOPERATIONFOR RELIABLE WIRELESS COMMUNICATIONS

The first author performed the work while at the Universityof Padova, Italy.

Cooperation and Network Coding aretwo powerful tech-niques for wirelessnetworks. During thepast four years asurge of researchactivity has tried tocombine the best ofthese two philoso-phies to improve theperformance anderror correction capa-bilities of radio net-works.

ROSSETTO LAYOUT 2/7/11 10:35 AM Page 15


eration, especially coded cooperation. On theother hand, NC is akin to fountain codes,although their applications are often very differ-ent. It is then very natural that channel codingplays a crucial role in bringing together theseworlds [4–8]. In the first and most widely ana-lyzed approach, the transmissions are structuredinto two rounds [5, 6]. In the first part, eachsource sends its own information unit (separatelyencoded from the others). The destination andrelays try to decode these frames, and the coop-erators concatenate the decoded informationunits to generate additional redundancy, which istransmitted in the second round. Hence, a largerchannel code that takes as input all decodedinformation units is created, thus obtaining spa-tial diversity. This effectively turns NC into acoded cooperative protocol based on Type IIhybrid automatic repeat request (HARQ), so theobserved benefits are due to the joint channelcoding across nodes.

This method of combining NC and channelcoding is by far the most studied in this areabecause of its natural connections to these twotechniques and relative ease of implementation.However, at least two important exceptions arise.Yang et al. [7] study a P = 2, N = 3 uplink (twosources, one shared cooperator), whose relay willforward the log likelihood ratios (LLRs) of thenetwork coded packets received from the sources.Note that these quantities can be computed alsoif one or both of the information units are cor-rupted; hence, the relay can also cooperate with

unreliable information about the original data,while in the previous approach [5, 6] redundancyis generated only on correctly decoded frames.However, this comes at the price of additionalbandwidth consumption, since the LLRs ratherthan the bits themselves are sent.

In the third approach [8], there are no addi-tional relays (i.e., N = P = 2); nonetheless, adiversity order of 2 is achievable. Such a result ispossible because each source knows its owninformation, and this leads to an encodingscheme in which each partner transmits the alge-braic superposition of its local and relayed infor-mation. Decoding at the destination is thencarried out by iterating between the codewordsfrom the two partners. Reference [8] is impor-tant also because it highlighted the so-calledCooperative Dilemma: if a terminal has to trans-mit two packets at the same time (e.g., becausethe first packet is a frame of its own and the sec-ond is a relayed one), it can either sum theirGalois symbols and then modulate them, orapply superposition coding by summing theirmodulated signals. In the latter case, the nodemust strike a trade-off between the power toallocate between the first and the second packet.This can be very detrimental, because subtract-ing power from the relayed packet will reducethe effectiveness of cooperation, while subtract-ing power from the endogenous frame will makeit harder to decode that packet both by theintended destination and by any possible relay.The scheme in [8] uses Galois arithmetic ratherthan superposition coding, because multiplexingtwo bit streams by means of Boolean additiondoes not require to divide the transmitted powerbetween the two frames. The importance of thisdetail will be apparent also in Physical LayerNetwork Coding, where it will be shown thatsuperposition coding can imply a loss of 3 dBwith respect to Galois arithmetic.

Table 1 summarizes the performance of thesethree strategies. The shown parameters are P, N,the coding rate, the achieved diversity order (i.e.,the slope of the bit error rate [BER] curve vs.signal-to-noise ratio [SNR] in a log-log scaleunder Rayleigh fading), and the scenario forwhich these schemes are designed. If N > P,there are more transmitters than informationunits. It can be noticed that diversity is achievedin all schemes by means of an additional relay.The only exception is [8], for the aforemen-tioned reasons. As an additional remark, [9] hasshown that separate channel and network codingcannot achieve full diversity order N. In order toattain this goal, lower channel code rates areneeded. This is indirectly confirmed by all theprevious network-channel coding systems, sincetheir code rate is relatively low (at most 1/2).

To summarize, the studies on hybrid chan-nel/network coding have delineated a few impor-tant points:• The performance gap between separate and

joint channel/network coding is about 2–3 dB[5, 6, 8].

• The usage of incorrectly decoded frames togenerate additional redundancy leads to band-width expansion [7].

• The achievable diversity order increases as thecode rate is decreased [8, 9].

Figure 1. The reference scenario for hybrid channel network coding.

SP

S1

R1

RM

DK

D1

Table 1. Comparison of cooperative channel/network coding schemes.

Reference P N rate Diversity order Notes

[5] (JNCC) 2 3 1/3 2 Third node is a relay

[6] (ANCC) P 2P 1/2 ≤ P

[7] 2 3 1/3 2 Third node is a relay

[8] 2 2 1/3 2 No relays



In this research area, significant results havebeen achieved for idealized settings. Nonethe-less, some issues are still open. First of all, allefforts focus on metrics taken from digital com-munication theory (e.g., diversity order, BER) orinformation theory (e.g., capacity regions), whilenetworking issues have not received much atten-tion so far.

Another problem is the lack of evaluation ofthese techniques in realistic environments, eitherbecause the systems are impractical (and mainlyshow performance bounds, rather than evaluat-ing real world protocols or network layer issuesspurred by this techniques) or the communica-tion protocols are too idealized. In conclusion,the design and performance of these systems inmore realistic scenarios including even a mini-mal protocol stack is yet to be investigated indetail, and appears to be a promising area ofresearch.

PHYSICAL LAYER NETWORK CODING

Another approach that has aroused much excite-ment has been so-called analog NC or physicallayer NC [3, 14]. Let us refer to Fig. 2, whichillustrates the well-known two-way relaychannel.1 Nodes A and B must exchange twopackets (X from A and Y from B) through relayR. If the channels are assumed to be error-free,this frame exchange would entail four transmis-sions in traditional routing: first X is sent fromA to R, then Y from B to R, X from R to B, andfinally Y from R to A. Two phases can be distin-guished: first, A and B must send their packetsto R (multiple access phase), and then R has tosend these frames to the destinations (broadcastphase). Digital NC (DNC) [1] works only on thebroadcast phase, by sending an invertible func-tion of X and Y. Nodes A and B can recover allinformation since they know X and Y, respec-tively. Such strategy reduces the number ofrequired transmissions from 4 to 3. Physicallayer Network Coding takes off from the obser-vation that R need not know X and Y separate-ly, but it is enough to decode a function of them(e.g., their sum).2 Hence, in PNC multiple accessphase, A and B are allowed to transmit simulta-neously. The channel additivity (i.e., the colli-sion, in networking jargon) naturally computes alinear combination of the two packets and thisfunction will then be broadcast to A and B. Theclear advantage is that just 2 slots are requiredto deliver the frames, rather than 3 in DNC or 4in traditional routing.

Of course, several details need to be filled inthis picture, and the vast majority of the researchhas focused on the two-way relay channel as asimple and neat starting point. We shall assumethroughout this section that this is the case ofinterest. Three main approaches have emerged[3, 10–17]:

•Amplify and Forward PNC (AF-PNC): Alsoknown as Analog Network Coding (ANC).According to this idea [10, 11, 14, 16], the relaydecodes neither X, Y nor their sum. Hence, itdeals only with the analog signals that hºave col-lided during the packet reception. The resultingsignal is amplified and broadcast, and the twoextreme nodes decode the intended packet after

subtracting the frame that they sent (which isknown, of course). The advantage of this scheme,compared to the other flavors of PNC, is its sim-plicity at the relay. On the other hand, the relayalso amplifies the noise with which the final des-tinations (A and B) must cope. And last but notleast, A and B need to carry out non-trivialinterference cancellation that requires channelinformation. This sophisticated signal processinghas actually been implemented, but it requiresconsiderable effort and works only in some spe-cific proof-of-concept scenarios [10]. It is yetunknown how to generalize it to less particularsettings at affordable complexity.

•Decode and forward PNC type 1 (DF-PNC1): Also known as compute and forward, thisstrategy [3, 12, 15] suggests decoding the sum ofX and Y at the relay, but at neither X nor Yindividually. This strategy is inherently morecomplicated than ANC, because the sum of thesignals from A and B is less structured than, forinstance, the modulation to which X and Ybelong. All these authors try to solve this issuewith lattices, which are mathematical subsets ofRN such that the sum of any two elements ofthese subsets still belongs to the lattice. Undersimplifying assumptions on the channel coeffi-cients between A, R and B, this scheme workswell, because it is not subject to error amplifica-tion, and can effectively support higher rates

Figure 2. Different protocols for the two way relay channel. The subscript to apacket is the time slot in which this frame is sent.

x1

x4 y3

y2

A R

Traditional routing

B

x1 y2

A R

x⊕y3

x+y1

f(x+y)/g(x,y)2

Digital network coding

Analog network coding

B

A R B

1 Note that the two-way relaychannel is a special case ofFig. 1, with two sources, onerelay, and two destinations.We also remark that eachsource is the destination forthe other source.

2 This statement is certain-ly true for the two-wayrelay channel. However, Rcan no longer computeany other function of Xand Y except for the one ithas received over the air.This can be a problem innetworks where nodesshould keep on producinginnovative packets,because in this case R can-not compute, for instance,any linear combination ofX and Y independent ofwhat has been received.Whether such a problemoffsets the potential gainsof PNC in large networksis yet to be explored.



than ANC. However, lattice-based schemes areextremely sensitive to timing and carrier syn-chronization; that is, A and B must be symboland phase synchronous at R, which may be hardto achieve.

•Decode and forward PNC type 2 (DF-PNC2): In this approach [13] the relay decodes X andY out of their superimposed signal by means ofmultiuser detection. While ANC and DF-PNC 1base their transmission on a complex linear com-bination of the transmitted signals X and Y, DF-PNC 2 separately decodes the two superimposedframes. After that, R transmits a linear combina-tion of the digital packets. This approach is notaffected by noise propagation (which besetsANC), but requires to decode more informationthan DF-PNC 1. As we will show, this matter canheavily affect the system performance.

In order to compare these three approaches,we have computed the achievable capacities forANC, DF-PNC 1 and DF-PNC 2 for the twoway relay network with AWGN channels (seealso [17]). Let us assume, for simplicity, thatboth A-R and B-R links are subject to the sameSNR?. The maximum rate for each flow is theminimum between the capacity of the multipleaccess phase and the capacity of the broadcastphase. The four protocols to be compared areDNC, ANC, DF-PNC 1 and DF-PNC 2, andtheir capacities can be computed as follows:

1. DNC

2. ANC [10, 11, 13, 14, 16]

3. DF-PNC 1 [12, 15]

4. DF-PNC 2 [13]

The results are plotted in Fig. 3. As can benoticed, DF-PNC 1 is the overall winner, becauseit can successfully suppress the noise at the relay(hence it is not affected by error propagation,unlike ANC). On the other hand, the necessityfor DF-PNC 2 to decode both X and Y inducesa big performance loss also with respect to DNC.

As a final remark, despite these works, someimportant matters have not been dealt with yet.For instance, all these ideas work for frequencyflat channels, and it is unknown how to modifythem in a frequency selective channel. In addi-tion, tight symbol synchronization is oftenassumed (i.e., the packets must be symbol- andoften also phase-synchronized at the relay). Anexception is [10], which was able to overcomesome of these problems, although in some verysimple, standard topologies where these issuesare not as troublesome as in a generic, randomnetwork. Finally, almost all the aforementionedpapers study information-theoretic based met-rics, such as the achievable rate regions, while anactual MAC protocol that could work in a vari-ety of settings has not yet been designed beyondthe first steps of [10, 16].

In spite of these open issues, physical layernetwork coding has the inherent and paramountmerit of truly merging Network Coding and themultiple access channel: the benefits of NetworkCoding for the multicast/broadcast channel havereceived immense attention by the NC commu-nity and this situation is well understood. How-ever, it is less clear how NC impacts multipleaccess schemes, whereas physical layer networkcoding provides an original and new point ofview on this issue, although many challenges arestill ahead.

MIMO NETWORK CODING

Network coding is inherently a MIMO scheme.It codes together multiple information units (theinputs) to yield multiple coded packets (the out-puts), and all techniques to recover the informa-tion units (e.g., the Gaussian eliminationprocedure) are effectively vector detection algo-rithms. Recovering a vector of received informa-tion units from a vector of received samples isone of the key issues in MIMO, and some papershave investigated how ideas drawn from MIMOcan be used to improve some aspects of NC [9,18, 19]. For instance, MIMO is well known toprovide diversity and power gains and hence it isrobust to errors and noise. Given this property,MIMO may also be able to retrieve informationeven if some of the antennas are subject tostrong fading. Such features are especially desir-able in NC, as the loss of a coded packet maydelay the whole decoding process. Since thisarea is still rather unexplored, it has been left asthe last in this article.

CDF PNC− = + +⎛⎝⎜

⎞⎠⎟ 2

1

41 2

1

21min log( ), log( )Λ Λ

CDF PNC− = +⎛⎝⎜

⎞⎠⎟ 1

1

2

1

2log Λ

CANC = ++

⎛

⎝⎜

⎞

⎠⎟

1

21

3 1

2

logΛΛ

CDNC = +1

31log( )Λ

Figure 3. Comparison of DNC and the three main approaches to PNC.

SNR (dB) -5 -10

0.5

0

Max

thr

ough

put

(bit

s/s/

Hz)

1

1.5

2

2.5

3

3.5

4

0 5 10 15 20 25

DNCANCDF-PNC 1DF-PNC 2



The assumption of having channel state infor-mation at the receiver (coherent detection) ornot (noncoherent detection) constitutes one ofthe basic distinctions in MIMO signal process-ing. Network coding approaches based on coher-ent MIMO processing have found more directapplication to cooperation so far; thus, we focuson them.

In [9], usage of coherent MIMO signal pro-cessing for NC detection is investigated. Theunderlying principle is to use the channel stateestimates and NC coefficients to perform jointdemodulation and NC decoding based on thereceived analog signals. The scheme is madepractical by using efficient MIMO algorithmssuch as Sphere Decoding, which leads to near-optimal performance at affordable complexity.Such a method offers some relevant advantagesover conventional NC, of which the most impor-tant is the ability to use corrupted or redundantframes (i.e., linearly dependent on the alreadyreceived coded packets). This ability is definitelyinteresting in wireless NC, since the loss of evena single coded packet might lead to a rank defi-cient network coding matrix and thus precludedecoding. Instead, by leveraging the possibilityto use also corrupted frames, such an event isindeed far less likely. In addition, the jointdemodulation-decoding process always yields aBER no larger than conventional NC, henceimproving the error correction capabilities. Onthe other hand, [9] remarked that conventionalNC encoding cannot perform well in a wirelessenvironment when network and channel codingare separated. For instance, if N nodes transmitone coded packet each out of the same pool of Pinformation units, the diversity order at anyreceiver will be at most N – P +1, even withmaximum likelihood detection. This is in starkcontrast to MIMO, where the diversity orderwith N receive antennas, M transmit antennas,and M independent, spatially multiplexed trans-mitted streams with maximum likelihood detec-tion is N. Such a problem has been implicitlyrecognized by [8], where it is shown that evenwith N = P = 2, a diversity order of 2 can beachieved, but only if channel and network codingare jointly designed.

A straightforward application of this princi-ple is Phoenix [19], so far one of the two hybridcooperative/network coding protocols that havebeen tested in a full-fledged discrete event sim-ulator or in a testbed, the other being the Ana-log Network Coding protocol of [10]. Phoenix isa Carrier Sense Multiple Access (CSMA) basedprotocol that reduces the bandwidth inefficiencyof cooperative retransmissions. Let us considerFig. 4, where node A has unsuccessfully trans-mitted a packet X to node D. In a cooperativeprotocol, one of A and D’s neighbors is electedas relay (called R), and has to perform a retrans-mission on behalf of A. In a standard coopera-tive protocol, R retransmits a copy of X; henceR helps A, but receives no direct reward. If Rsent a coded packet, which was a linear combi-nation X ⊕ Y of X and one of R’s packets, calledY, R would also deliver its own traffic, albeit atthe price of a slightly higher packet error rate.However, standard NC would not be able torecover X and Y, because A’s coded packet (the

first version of X) is corrupted and hence unus-able by ordinary NC algorithms. Instead, if Demploys MIMO_NC, D can jointly decode Xand X ⊕ Y, and could potentially recover bothX and Y. Phoenix encourages nodes to helpeach other, since they can perform a retransmis-sion and pursue their own interest. Cooperativebehavior is especially useful in multihop net-working, because a route composed by multiplehops needs to pursue higher link reliability thansingle-hop communication in order to achievesatisfactory performance. An example is report-ed in Fig. 5 [19]. This picture reports the aggre-gate throughput for two-hop and four-hoproutes delivered by an ad hoc network with 25nodes and 1 Mb/s link data rate, where all ter-minals transmit data to a gateway, possiblythrough a multihop path; three protocols arecompared: IEEE 802.11 (carrier sense multipleaccess, CSMA), a conventional decode-and-for-ward cooperative protocol (cooperative CSMA,CCSMA), and the hybrid NC/cooperative ver-sion (Phoenix). Phoenix gains as much as 18percent over CCSMA for two-hop traffic andeven more for four-hop traffic. In addition, thechance to send a new packet along with aretransmission is especially useful to relievecongestion in bottlenecks. For instance, if manyflows converge to the same node (e.g., a gate-way), any packet loss would delay the traffic ofall the flows converging there. Total aggregatethroughput can gain as much as 25 percent overCCSMA [19]. Phoenix is able to relieve this net-work-level problem by its cooperative/NCretransmissions.

For completeness, we also briefly summarizethe main results on Network Coding techniquesinspired by non-coherent MIMO [18]. Reference[18] points out that if the NC coefficient matrixis unknown to the receiver, information isbrought not by the information units themselves,but by the vector space they span. Such a prop-erty is strongly related to noncoherent MIMOdetection, since in that case as well informationis not brought by the transmitted symbols but bythe subspace they span. This parallel enables todeeply understand the limits and possibilities oferror correction techniques for NC and to designbetter codes. Many concepts of classic channelcoding theory (such as the sphere packing,sphere covering, and singleton bounds) areextended to non-coherent NC, also thanks tothis parallel with MIMO.

Figure 4. Reference topology for NC/cooperative protocols [19].

x

x/x⊕y R

S D

y

x



THE CHALLENGES AHEADThe quest for highly efficient and practical hybridNC/cooperative protocols started about fouryears ago and has already energized a whole sec-tor of the research community. In this relativelyshort period of time, three main approaches tothis cross-layer area have emerged that havelooked at this field with different perspectives,yet interactions among them have shed light onsome deep and important issues on the topic.These are some of the lessons that have beenlearned together with the challenges they open:

What will the role of PHY be in this area?All these roads suggest that interaction

between NC and cooperation must have thephysical layer (PHY) as a pivotal element. More-over, the three PHYs adopted by the describedapproaches are radically different from eachother, since each of them must be designedaccording to the different needs and philosophybehind every technique. Hence, the adoption ofone of these systems implies a clear choice onthe physical layer.

No new PHY can be effective without a prop-er medium access control (MAC) or networklayer, and MIMO networking is a recent andtelling example. For the time being, MAC layerissues or opportunities brought by hybrid NC/cooperative techniques have gone largely unad-dressed. The only notable exceptions are Phoenix[19], where an 802.11-like protocol usesMIMO_NC [9] to “hide” packet losses, andANC [10]. In addition, can this PHY be support-ed by conventional off-the-shelf radios, or shoulda brand new chip be designed?

Can the NC architecture remain the same fora wireless cooperative protocol?

•The encoding phase of NC must be changedin order to exploit the full gain of NC in thewireless environment [9]. Work in thechannel/network coding area and MIMO_NChave shown that separation of network codingand channel coding implies non-negligible per-formance degradations.

•Combining these two techniques entails morecomplexity in the system, and the question ofwhat is the trade-off between performance, redun-dancy, and complexity has only been partly inves-tigated and is still largely uncharted territory.

What have we learnt from physical layer NC?•Physical layer NC is a very interesting and

potentially beneficial idea. One of the basic find-ings is that in spite of exploiting the “analog”nature of the waveforms, especially for the possi-bility of “combining” signals in the air, it is morebeneficial to retain the use of digital algebra,because of its ability to suppress noise much bet-ter than amplify-and-forward protocols. Betterperformance can be achieved by decoding afunction of the transmitted information units(DF-PNC 1) rather than decoding each singlepacket (DF-PNC 2) or not decoding anythingand amplifying the aggregate received signal(AF-PNC).

•Several practical issues must be answeredbefore actual deployment in a generic randomnetwork is viable. So far, PNC requires tightsymbol synchronization, the same modulationfor colliding signals, and flat fading channels.

In conclusion, cooperation and network cod-ing have shown opposite features and problems,and their cutting-edge union may enhance theirvirtues while minimizing their problems. Thepotential for very high performance is unques-tionable, as the preliminary but clear indicationsin information and digital communication theoryhave shown.

ACKNOWLEDGMENTSFrancesco Rossetto is supported in part by theSpace Agency of the German Aerospace Centerand teh Federal Ministry of Economics andTechnology based on the agreement of the Ger-man Bundestag with the support code 50YB0905and by the European Commission within theSeventh Framework Programme (FP7) undertheSAPHYRE Project (C.A. #2480011). Theauthors would like to thank Elena Fasolo,Andrea Munari, and Bobak Nazer for insightfuldiscussions on these topics. They would also liketo thank the Associate Editor and anonymousreviewers for their insightful comments.

REFERENCES[1] R. Ahlswede et al., “Network Information Flow,” IEEE Trans.

Info. Theory, vol. 46, no. 4, July 2000, pp. 1204–16.[2] J. Laneman, Cooperation in Wireless Networks: Principles

and Applications, Springer, 2006, “Cooperative Diversity:Models, Algorithms, and Architectures,” pp. 163–88.

[3] S. Zhang, S. C. Liew, and P. P. Lam, “Physical-Layer NetworkCoding,” ACM MOBICOM, Los Angeles, CA, Sept. 2006.

[4] P. Larsson and N. Johansson, “Multi User ARQ,” VTCSpring, Melbourne, Australia, May 2006.

[5] C. Hausl and P. Dupraz, “Joint Network-Channel Codingfor the Multiple Access Relay Channel,” IEEE IWWAN,New York, NY, June 2006.

[6] X. Bao and J. Li, “Adaptive Network Coded Cooperation(ANCC) for Wireless Relay Networks: Matching Code-on-Graph with Network-on-Graph,” IEEE Trans. WirelessCommun., vol. 7, no. 2, Feb. 2008, pp. 574–83.

[7] S. Yang and R. Koetter, “Network Coding over a NoisyRelay: a Belief Propagation Approach,” IEEE ISIT, Nice,France, 9 Jan. 2007.

[8] L. Xiao et al., “A Network Coding Approach to Cooper-ative Diversity,” IEEE Trans. Info. Theory, vol. 53, no.10, Oct. 2007, pp. 3714–22.

[9] E. Fasolo, F. Rossetto, and M. Zorzi, “Network Codingmeets MIMO,” NetCod 2008, Hong Kong, China, Jan.2008.

Figure 5. Aggregate throughput for two- and four-hop traffic [19].

Load [pk/s] 3 2

2

0

Agg

rega

te t

hrou

ghpu

t [k

b/s]

4

6

8

10

12

4 5 6 7 8 9 10 11 12

CSMA, 2 hopsCSMA, 4 hopsCCSMA, 2 hopsCCSMA, 4 hopsPHOENIX, 2 hopsPHOENIX, 4 hops



[10] S. Katti, S. Gollakota, and D. Katabi, “Embracing Wire-less Interference: Analog Network Coding,” ACM SIG-COMM 2007, Kyoto, Japan, 27–31 Aug. 2007.

[11] S. Katti et al., “Joint Relaying and Network Coding inWireless Networks,” IEEE ISIT 2007, Nice, France, 24–28June 2007.

[12] B. Nazer and M. Gastpar, “Computation over MultipleAccess Channels,” IEEE Trans. Info Theory, vol. 53, no.10, Oct. 2007, pp. 3498–516.

[13] B. Rankov and A. Wittneben, “Spectral Efficient Proto-cols for Half-Duplex Fading Relay Channels,” IEEE JSAC,vol. 25, no. 2, Feb. 2007, pp. 379–89.

[14] P. Popovski and H. Yomo, “Bi-directional Amplificationof Throughput in a Wireless Multi-Hop Network,” IEEEVTC Spring, Melbourne (Australia), May 2006.

[15] –, “The Anti-Packets Can Increase the AchievableThroughput of a Wireless Multi-Hop Network,” IEEEICC, Istanbul, Turkey, June 2006.

[16] Z. Ding et al., “On the Study of Network Coding withDiversity,” IEEE Trans. Wireless Commun., vol. 8, no. 3,Mar. 2009, pp. 1247–59.

[17] Y. Hao et al., “Achievable Rates for Network Codingon the Exchange Channel,” IEEE MILCOM, Orlando, FL,Oct. 2007.

[18] R. Koetter and F. R. Kschischang, “Coding for Errorsand Erasures in Random Network Coding,” IEEE Trans.Infor. Theory, vol. 54, no. 8, Aug. 2008, pp. 3579–91.

[19] E. Fasolo et al., “Phoenix: A Hybrid Cooperative-Net-work Coding Protocol for Fast Failure Recovery in AdHoc Networks,” IEEE SECON 2008, San Francisco, CA,16–20 June 2008.

BIOGRAPHIESFRANCESCO ROSSETTO [S’06, M’09] ([email protected])received his Laurea (equivalent to M.S.) and Ph.D. degreesin telecommunications engineering in 2005 and 2009,respectively, from the University of Padova, Italy. In2004–2005 he studied electrical engineering at the Univer-sity of California, San Diego (UCSD) under a student

exchange program. In 2008 he was on leave at UCSD,working for the MURI project, a multiuniversity initiativefor the development of multihop MIMO networks. Since2009 he has been with the DLR (German Aerospace Cen-ter), Munich. His research interests include satellite com-munication, network coding, and cross-layer design. Hiscorporate experience includes a summer internship in 2006at Ericsson Eurolabs, Aachen, Germany, working on HARQfor 3G/LTE cellular networks.

MICHELE ZORZI [F’07] ([email protected]) received his Laureaand Ph.D. degrees in electrical engineering from the Uni-versity of Padova in 1990 and 1994, respectively. Duringacademic year 1992–1993 he was on leave at UCSD,attending graduate courses and doing research on multipleaccess in mobile radio networks. In 1993 he joined the fac-ulty of the Dipartimento di Elettronica e Informazione,Politecnico di Milano, Italy. After spending three years withthe Center for Wireless Communications at UCSD, in 1998he joined the School of Engineering of the University ofFerrara, Italy, where he became a professor in 2000. SinceNovember 2003 he has been on the faculty of the Informa-tion Engineering Department at the University of Padova.His present research interests include performance evalua-tion in mobile communications systems, random access inmobile radio networks, ad hoc and sensor networks, ener-gy constrained communications protocols, and underwatercommunications and networking. He was Editor-In-Chief ofIEEE Wireless Communications from 2003 to 2005, is cur-rently Editor-In-Chief of IEEE Transactions on Communica-tions, and serves on the Editorial Board of the WileyJournal of Wireless Communications and Mobile Comput-ing. He was also guest editor for special issues in IEEE Per-sonal Communications (“Energy Management in PersonalCommunications Systems”) and IEEE Journal on SelectedAreas in Communications (“Multimedia Network Radios”and “Underwater Wireless Communications and Network-ing”). He is a Member-at-Large of the Board of Governorsof the IEEE Communications Society.




INTRODUCTIONUnderwater wireless communication networks(UWCNs) are constituted by sensors andautonomous underwater vehicles (AUVs) thatinteract to perform specific applications such asunderwater monitoring (Fig. 1) [1]. Coordinationand sharing of information between sensors andAUVs make the provision of security challeng-ing. The aquatic environment is particularly vul-nerable to malicious attacks due to the high biterror rates, large and variable propagationdelays, and low bandwidth of acoustic channels.Achieving reliable intervehicle and sensor-AUVcommunication is especially difficult due to themobility of AUVs and the movement of sensorswith water currents.

The unique characteristics of the underwateracoustic channel, and the differences betweenunderwater sensor networks and their ground-based counterparts require the development ofefficient and reliable security mechanisms.

This article discusses security in UWCNs. It isstructured as follows. The following sectionexplains the specific characteristics of UWCNsin comparison with their ground-based counter-parts. Next, the possible attacks and counter-measures are introduced. Subsequently, securityrequirements for UWCNs are described. Later,the research challenges related to secure timesynchronization, localization, and routing aresummarized. Finally, the article is concluded.

CHARACTERISTICS ANDVULNERABILITIES OF UWCNS

Underwater sensor networks have some similari-ties with their ground-based counterparts such astheir structure, function, computation and ener-gy limitations. However, they also have differ-ences, which can be summarized as follows.

Radio waves do not propagate well underwa-ter due to the high energy absorption of water.Therefore, underwater communications arebased on acoustic links characterized by largepropagation delays. The propagation speed ofacoustic signals in water (typically 1500 m/s) isfive orders of magnitude lower than the radiowave propagation speed in free space.

Acoustic channels have low bandwidth. Thelink quality in underwater communication isseverely affected by multipath, fading, and therefractive properties of the sound channel. As aresult, the bit error rates of acoustic links areoften high, and losses of connectivity arise [1].

Underwater sensors move with water cur-rents, and AUVs are mobile. Although certainnodes in underwater applications are anchoredto the bottom of the ocean, other applicationsrequire sensors to be suspended at certain depthsor to move freely in the underwater medium.

The future development of geographical rout-ing is very promising in UWCNs due to its scala-bility and limited signaling properties. However,it cannot rely on the Global Positioning System(GPS) because it uses radar waves in the 1.5GHz band that do not propagate in water.

Since underwater hardware is more expen-sive, underwater sensors are sparsely deployed.

Underwater communication systems havemore stringent power requirements than terrestri-al systems because acoustic communications aremore power-hungry, and typical transmission dis-tances in UWCNs are greater; hence, highertransmit power is required to ensure coverage [1].

The above mentioned characteristics ofUWCNs have several security implications.UWCNs suffer from the following vulnerabili-ties. High bit error rates cause packet errors.Consequently, critical security packets can belost. Wireless underwater channels can be eaves-dropped on. Attackers may intercept the infor-

MARI CARMEN DOMINGO, BARCELONA TECH UNIVERSITY

ABSTRACTUnderwater wireless communication net-

works are particularly vulnerable to maliciousattacks due to the high bit error rates, largeand variable propagation delays, and low band-width of acoustic channels. The unique charac-teristics of the underwater acousticcommunication channel, and the differencesbetween underwater sensor networks and theirground-based counterparts require the develop-ment of efficient and reliable security mecha-nisms. In this article, a complete survey ofsecurity for UWCNs is presented, and theresearch challenges for secure communicationin this environment are outlined.

SECURING UNDERWATER WIRELESSCOMMUNICATION NETWORKS

The authors presenta complete survey of security for Underwater WirelessCommunication Networks, and theyoutline the researchchallenges for securecommunication inthis environment.

DOMINGO LAYOUT 2/7/11 10:26 AM Page 22


mation transmitted and attempt to modify ordrop packets. Malicious nodes can create out-of-band connections via fast radio (above the watersurface) and wired links, which are referred to aswormholes. Since sensors are mobile, their rela-tive distances vary with time. The dynamic topol-ogy of the underwater sensor network not onlyfacilitates the creation of wormholes but it alsocomplicates their detection [2]. Since power con-sumption in underwater communications is high-er than in terrestrial radio communications, andunderwater sensors are sparsely deployed, ener-gy exhaustion attacks to drain the batteries ofnodes pose a serious threat for the network life-time.

ATTACKS ON UWCNS ANDCOUNTERMEASURES

Both intervehicle and sensor-AUV communica-tions can be affected by denial-of-service (DoS)attacks. Next, we summarize typical DoS attacks,evaluate their dangers, and indicate possibledefenses to muffle their effects.

JAMMINGA jamming attack consists of interfering with thephysical channel by putting up carriers on thefrequencies neighbor nodes use to communicate.Since underwater acoustic frequency bands arenarrow (from a few to hundreds of kilohertz),UWCNs are vulnerable to narrowband jamming.Localization is affected by the replay attackwhen the attacker jams the communicationbetween a sender and a receiver, and laterreplays the same message with stale information(an incorrect reference) posing as the sender(Fig. 2).

Since jamming is a common attack in wirelessnetworks, some of the solutions proposed fortraditional wireless networks can be applied.Spread spectrum is the most common defenseagainst jamming [3]. Frequency hopping spreadspectrum (FHSS) and direct sequence spreadspectrum (DSSS) in underwater communicationsare drawing attention for their good perfor-mance under noise and multipath interference.These schemes are resistant to interference fromattackers, although not infallible. An attackercan jam a wide band of the spectrum or followthe precise hopping sequence when an FHSSscheme is used. A high-power wideband jam-ming signal can be used to attack a DSSSscheme.

Underwater sensors under a jamming attackshould try to preserve their power. When jam-ming is continuous, sensors can switch to sleepmode and wake up periodically to check if theattack is over. When jamming is intermittent,sensors can buffer data packets and only sendhigh-power high-priority messages to report theattack when a gap in jamming occurs.

In ground-based sensor networks, other sen-sors located along the edge of the area underattack can detect the jamming signal as higher-than-normal background noise and report intru-sion to outside nodes. That will cause any furthertraffic to be rerouted around the jammed region[3]. However, this solution cannot be applied to

UWCNs, since nodes underwater are usuallysparsely deployed, which means there would notbe enough sensors to delimit the jammed regionaccurately and reroute traffic around it. Anothersolution proposed for ground-based sensor net-works against jamming is to use alternative tech-nologies for communication such as infrared oroptical [3]. However, this solution cannot beapplied either, since optical and infrared wavesare severely attenuated under water.

WORMHOLE ATTACKA wormhole is an out-of-band connection creat-ed by the adversary between two physical loca-tions in a network with lower delay and higherbandwidth than ordinary connections. This con-nection uses fast radio (above the sea surface) orwired links (Fig. 3) to significantly decrease thepropagation delay. In a wormhole attack themalicious node transfers some selected packetsreceived at one end of the wormhole to theother end using the out-of-band connection, andre-injects them into the network [4]. The effect

Figure 1. Underwater sensor network with AUVs.

AUV1

AUV4

AUV2

AUV3

Event 1

Event 2

Event 3

Sink

Figure 2. Replay attack.

Sink

Malicious node

Underwater sensor node

Underwater sensor node

Message with stale

information



is that false neighbor relationships are created,because two nodes out of each other’s range canerroneously conclude that they are in proximityof one another due to the wormhole’s presence.This attack is devastating. Routing protocolschoose routes that contain wormhole linksbecause they appear to be shorter; thus, theadversary can monitor network traffic and delayor drop packets sent through the wormhole.Localization protocols can also be affected bythese attacks when malicious nodes claim wronglocations and mislead other nodes.

One proposed method for wormhole detec-tion in ground-based sensor networks consists ofestimating the real physical distance betweentwo nodes to check their neighbor relationship[4]. If the measured distance is longer than thenodes’ communication range, it is assumed thatthe nodes are connected through a wormhole.However, accurate distance estimation dependson precise localization (geographical packetleashes, wormhole detection using position infor-mation of anchors), tight clock synchronization(temporal packet leashes), or use of specifichardware (directional antennas) [4]. In underwa-ter communications accurate localization andtime synchronization are still challenging.

The authors in [2] propose a distributedmechanism named Distributed Visualization ofWormhole (Dis-VoW) to detect wormholeattacks in three-dimensional underwater sensornetworks. In Dis-VoW, every sensor collects thedistance estimations to its neighbors using theround-trip time of acoustic signals; after thesedistances are broadcast by each sensor to itsneighbors, every node is able to construct the

local network topology (virtual layout) withintwo hops using multidimensional scaling (MDS).Since a wormhole contracts the virtual layout atcertain regions, some nodes far away appear tobe neighbors, and these contradictions can bedetected visualizing the virtual layout. A worm-hole indicator variable is defined to compute thedistortion in angles; the distortion in edgelengths is computed as the difference betweenthe measured distances among neighboring sen-sors and the lengths of the reconstructed con-nections. In [5] a suite of protocols is proposedto enable wormhole-resilient secure neighbordiscovery with high probability in underwatersensor networks. This solution is based on thedirection of arrival (DoA) estimation of acousticsignals, which depends on the relative locationsof signal transmitters and receivers, and cannotbe manipulated.

SINKHOLE ATTACKIn a sinkhole attack, a malicious node attemptsto attract traffic from a particular area toward it;for example, the malicious node can announce ahigh-quality route. Geographic routing andauthentication of nodes exchanging routinginformation are possible defenses against thisattack, but geographic routing is still an openresearch topic in UWCNs.

HELLO FLOOD ATTACKA node receiving a HELLO packet from a mali-cious node may interpret that the adversary is aneighbor; this assumption is false if the adver-sary uses high power for transmission. Bidirec-tional link verification can help protect againstthis attack, although it is not accurate due tonode mobility and the high propagation delaysof UWCNs. Authentication is also a possibledefense.

ACKNOWLEDGMENT SPOOFINGA malicious node overhearing packets sent toneighbor nodes can use this information to spooflink layer acknowledgments with the objective ofreinforcing a weak link or a link located in ashadow zone. Shadow zones are formed whenthe acoustic rays are bent and sound waves can-not penetrate. They cause high bit error ratesand loss of connectivity [1]. This way, the routingscheme is manipulated. A solution to this attackwould be encryption of all packets sent throughthe network.

SELECTIVE FORWARDINGMalicious nodes drop certain messages insteadof forwarding them to hinder routing. InUWCNs it should be verified that a receiver isnot getting the information due to this attackand not because it is located in a shadow zone.Multipath routing and authentication can beused to counter this attack, but multipath rout-ing increases communication overhead.

SYBIL ATTACKAn attacker with multiple identities can pretendto be in many places at once. Geographic rout-ing protocols are also misled because an adver-sary with multiple identities can claim to be inmultiple places at once (Fig. 4).

Figure 3. Underwater network with a wormhole link.

Wormhole link

Sink

Distributed underwater

sensor nodes



Authentication and position verification aremethods against this attack, although positionverification in UWCNs is problematic due tomobility.

SECURITY REQUIREMENTS

In UWCNs the following security requirementsshould be considered.

AUTHENTICATIONAuthentication is the proof that the data wassent by a legitimate sender. It is essential in mili-tary and safety-critical applications of UWCNs.Authentication and key establishment are strong-ly related because once two or more entities ver-ify each other’s authenticity, they can establishone or more secret keys over the open acousticchannel to exchange information securely; con-versely, an already established key can be usedto perform authentication. Traditional solutionsfor key generation and update (renewal) algo-rithms should be adapted to better address thecharacteristics of the underwater channel. In [6],a key generation system is proposed that requiresonly a threshold detector, lightweight computa-tion, and communication costs. It exploitsreciprocity, deep fades (strong destructive inter-ference), randomness extractor, and robustsecure fuzzy information reconciliators. Thisway, the key is generated using the characteris-tics of the underwater channel and is secureagainst adversaries who know the number ofdeep fades but not their locations.

CONFIDENTIALITYConfidentiality means that information is notaccessible to unauthorized third parties. There-fore, confidentiality in critical applications suchas maritime surveillance (Fig. 5) should be guar-anteed.

INTEGRITYIt ensures that information has not been alteredby any adversary. Many underwater sensor appli-cations for environmental preservation, such aswater quality monitoring [7], rely on the integrityof information.

AVAILABILITYThe data should be available when needed by anauthorized user. Lack of availability due todenial-of-service attacks would especially affecttime-critical aquatic exploration applicationssuch as prediction of seaquakes.

RESEARCH CHALLENGES

The security issues and open challenges forsecure time synchronization, localization, androuting in UWCNs are summarized in the fol-lowing sections.

SECURE TIME SYNCHRONIZATIONTime synchronization is essential in many under-water applications such as coordinated sensingtasks. Also, scheduling algorithms such as time-division multiple access (TDMA) require precisetiming between nodes to adjust their sleep-wake-up schedules for power saving. For example, in

water quality monitoring [7], sensors aredeployed at different depths because the chemi-cal characteristics of water vary at each level.The design of a delay-tolerant time synchroniza-tion mechanism is very important to accuratelylocate the water contaminant source, set up thesleep-wakeup schedules among neighboringnodes appropriately, and log the water qualitydata correctly into the annual database withaccurate timing information.

Achieving precise time synchronization isespecially difficult in underwater environmentsdue to the characteristics of UWCNs. For thisreason, the time synchronization mechanismsproposed for ground-based sensor networks can-not be applied, and new mechanisms have been

Figure 4. Sybil attack.

Maliciousnode C

Sink

Nodes A and Bwant to send

their datatowards the sink.

Maliciousnode C

broadcastsadvertisementmessages of

invented non-existent

positions ofnodes (yellow

nodes)

A and B select theinvented non-existentpositions of malicious

node C to forward theirmessages. Node C

overhears them.

B

B

A

A

a)

Sink

Sink

b)

c)

Maliciousnode C

Maliciousnode C

BA



proposed. Tri-Message [8] is a time synchroniza-tion protocol designed for high-latency networkswith a synchronization precision that increaseswith distance. A multilateration algorithm is pro-posed in [9] for localization and synchronizationin 3D underwater acoustic sensor networks. It isassumed that a set of anchors, several buoys onthe ocean surface, already know their locationsand time without error. A group of nearby sen-sors receives synchronization packets containingthe coordinates and packet transmit times fromat least five anchor nodes and performs multilat-eration to obtain their own locations. The sen-sors learn the time difference betweenthemselves and each anchor node by comparingtheir local times at which they received the timesynchronization packet with the transmit timeplus propagation delays; these nodes subse-quently become new anchor nodes and there-

after broadcast new synchronization packets to alarger range, and so on. MU-Sync [10] is a clus-ter-based synchronization protocol that esti-mates the clock skew by performing the linearregression twice over a set of local time informa-tion gathered through message exchanges. Thefirst linear regression enables the cluster head tooffset the effect of long and varying propagationdelay; the second regression enables the cluster-head to obtain the final estimated skew and off-set.

None of the aforementioned time synchro-nization schemes [8–10] consider security,although it is critical in the underwater environ-ment. Time synchronization disruption due tomasquerade, replay and message manipulationattacks, can be addressed using cryptographictechniques [11]. However, countering other pos-sible attacks such as delays (deliberate delayingthe transmission of time synchronization mes-sages) [11] and DoS attacks requires the use ofother strategies. The countermeasures againstdelay attacks proposed in [11] for ground-basedsensor networks are not applicable to UWCNs.They are based on the detection of outliers(malicious time offsets), but they do not distin-guish between deliberate alterations and abnor-mal values resulting from long and variablepropagation delays or node mobility. Moreover,they do not support global synchronization inmulti-hop sensor networks.

A correlation-based security model for waterquality monitoring systems has been proposed in[7] to detect outlier timestamps due to insiderattacks. The authors prove that the acousticpropagation delays between two sensors inneighboring depth levels fit an approximatelynormal distribution, which means that the times-tamps between them should correlate. However,this correlation is lost if a captured inside nodeis sending falsified timestamps. With properdesign of a timestamp sliding window scheme,insider attacks are detected. Each sensor shouldobtain timestamp readings from multiple sensorsand calculate the correlation coefficient for eachneighbor’s timestamp, obtaining a window ofcoefficients. If a coefficient of the window ofdata is below a threshold, it is an outlier value. Ifthe abnormal percentage of data in one window(outlier percentage) is consistently (10 consecu-tive windows) higher than a predeterminedthreshold, the corresponding neighbor is flaggedas a malicious node generating insider attacks.However, identifying a neighbor node as mali-cious is difficult, because sometimes timestampscan be corrupted due to propagation delay varia-tions caused by the channel rather than deliber-ately. Because of wave motion, the signalmultipath components undergo time-varyingpropagation delays. Node mobility due to watercurrents also modifies the propagation delays. Inorder to better distinguish between unintendedand malicious timestamp alterations, the authorsin [12] improve the proposed scheme by using asa second step a statistical reputation and trustmodel to detect outlier timestamps, and identifynodes generating insider attacks. It is based onquantitative measurements and on the assump-tion that identifying an insider attacker requireslong-term behavior observations.

Figure 5. Intruder submarine detection.

Distributed underwater

sensor nodes Intruder

submarine Surface sink Data path

Data transmitted to the on-shore command

center



The following open research issues for securetime synchronization need to be addressed:• Because of the high and variable propagation

delays of UWCNs, the time required to syn-chronize nodes should be investigated.

• Efficient and secure time synchronizationschemes with small computation and commu-nications costs need to be designed to defendagainst delay and wormhole attacks.

SECURE LOCALIZATIONLocalization is a very important issue for datatagging. Sensor tasks such as reporting the occur-rence of an event or monitoring require localiza-tion information. Localization can also help inmaking routing decisions. For example, theunderwater sensors in [13] learn the location andspeed of mobile beacons and neighbors duringthe localization phase; the position and motionof mobile beacons are used by the routing proto-col to choose the best relay for a node to for-ward its data.

Localization approaches proposed forground-based sensor networks do not work wellunderwater because long propagation delays,Doppler effect, multipath, and fading cause vari-ations in the acoustic channel. Bandwidth limita-tions, node mobility, and sparse deployment ofunderwater nodes also affect localization estima-tion. Proposed terrestrial localization schemesbased on received signal strength (RSS) are notrecommended in UWCNs, since non-uniformacoustic signal propagation causes significantvariations in the RSS. Time of arrival (ToA) andtime difference of arrival (TDoA) measurementsrequire very accurate time synchronization(which is a challenging issue), and angle ofarrival (AoA) algorithms are affected by theDoppler shift.

Localization schemes can be classified into:Range-based schemes (using range and/or

bearing information): The location of nodes inthe network is estimated through precise dis-tance or angle measurements.

•Anchor-based schemes: Anchor nodes aredeployed at the seabed or sea surface at loca-tions determined by GPS. The propagation delayof sound signals between the sensor [9] or AUVand the anchors is used to compute the distanceto multiple anchor nodes.

•Distributed positioning schemes: Positioninginfrastructure is not available, and nodes com-municate only with one-hop neighbors and com-pute their locations using multilateration.Underwater sensor positioning (USP) has beenproposed in [14] as a distributed localizationscheme for sparse 3D networks, transformingthe 3D underwater positioning problem into a2D problem using a distributed non-degenera-tive projection technique. Using sensor depthinformation, the neighboring reference nodesare mapped to the horizontal plane containingthe sensor to be localized. After projecting thereference nodes, localization methods for 2Dnetworks such as bilateration or trilateration canbe used to locate the sensor.

•Schemes that use mobile beacons/anchors:They use mobile beacons whose locations arealways known. Scalable localization with mobili-ty prediction (SLMP) has been proposed in [15]

as a hierarchical localization scheme. At thebeginning, only surface nodes know their loca-tions, and anchor nodes can be localized bythese surface buoys. Anchor nodes are selectedas reference nodes because of their known loca-tions; with the advance of the location processmore ordinary nodes are localized and becomereference nodes. During this process, everynode predicts its future mobility pattern accord-ing to its past known location information. Thefuture location is estimated based on this pre-diction.

Range-free schemes (not using range or bear-ing information): They have been designed assimple schemes to compute only coarse positionestimates. A range-free scheme proposed in [16]estimates the location of a sensor within a cer-tain area.

None of the aforementioned localizationschemes [9, 13–16] was designed with security inmind. Some localization-specific attacks (replayattack, Sybil attack, wormhole attack) have pre-viously been described.

Open research issues for secure localizationare:• Effective cryptographic primitives against

injecting false localization information inUWCNs need to be developed.

• It is necessary to design resilient algorithmsable to determine the location of sensorseven in the presence of Sybil and wormholeattacks.

• Techniques to identify malicious or compro-mised anchor nodes and to avoid false detec-tion of these nodes are required.

• Secure localization mechanisms able to handlenode mobility in UWCNs need to be devised.

SECURE ROUTINGRouting is essential for packet delivery inUWCNs. For example, the Distributed Under-water Clustering Scheme (DUCS) [17] does notuse flooding and minimizes the proactive routingmessage exchange.

Routing is specially challenging in UWCNsdue to the large propagation delays, the lowbandwidth, the difficulty of battery refills ofunderwater sensors, and the dynamic topologies.Therefore, routing protocols should be designedto be energy-aware, robust, scalable and adap-tive.

Many routing protocols have been proposedfor underwater wireless sensor networks. How-ever, none of them has been designed withsecurity as a goal. Routing attacks can disablethe entire network’s operation. Spoofing, alter-ing, or replaying routing information affectsrouting. Important routing attacks (selectiveforwarding, sinkhole attack, Sybil attack, worm-hole attack, HELLO flood attack, acknowledg-ment spoofing) have been previously described.Although the attacks against routing in UWCNsare the same as in ground-based sensor net-works, the same countermeasures are notdirectly applicable to UWCNs due to their dif-ference in characteristics. Proposed broadcastauthentication methods would cause high com-munication overhead and latency in UWCNs.Multipath routing would cause high communi-cation overhead as well.

Routing is speciallychallenging in

UWCNs due to thelarge propagationdelays, low band-

width, difficulty of battery

refills of underwatersensors, and dynam-ic topologies. There-

fore, routingprotocols should be

designed to be energy-aware,

robust, scalable and adaptive.



Open research issues for secure routing are:•There is a need to develop reputation-based

schemes that analyze the behavior of neighborsand reject routing paths containing selfish nodesthat do not cooperate in routing. The properfunctioning of these schemes is challengingbecause they do not work well in mobile environ-ments, the time required to detect compromisednodes increases substantially in UWCNs due tothe long propagation delays, and they must beadapted to tolerate short-term disruptions.

•Quick and powerful encryption and authen-tication mechanisms against outside intrudersshould be devised for UWCNs because the timerequired for intruder detection is high due to thelong and variable propagation delays, and rout-ing paths containing undetected malicious nodescan be selected in the meantime for packet for-warding.

•Sophisticated mechanisms should be devel-oped against insider attacks such as selective for-warding, Sybil attacks, HELLO flood attacks,and acknowledgment spoofing.

•There is a need to develop new techniquesagainst sinkholes and wormholes, and improveexisting ones. With Dis-VoW [2] a wormholeattack can still be concealed by manipulating thebuffering times of distance estimation packets.The wormhole-resilient neighbor discovery in [5]is affected by the orientation error between sen-sors.

CONCLUSIONS

In this article we have discussed security inUWCNs, underlining the specific characteristicsof these networks, possible attacks, and counter-measures. The main research challenges relatedto secure time synchronization, localization, androuting have also been surveyed. These researchissues remain wide open for future investigation.

ACKNOWLEDGMENTThis work was supported by the Spanish Min-istry of Education and Science under projectTSI2007-66637-C02-01.

REFERENCES[1] I. F. Akyildiz, D. Pompili, and T. Melodia, “Underwater

Acoustic Sensor Networks: Research Challenges,” AdHoc Net., vol. 3, no. 3, Mar. 2005.

[2] W. Wang et al., “Visualization of Wormholes in Under-water Sensor Networks: A Distributed Approach,” Int’l.J. Security Net., vol. 3, no. 1, 2008, pp. 10–23.

[3] A. D. Wood and J. A. Stankovic, “A Taxonomy forDenial-of-Service Attacks in Wireless Sensor Networks,”chapter in Handbook of Sensor Networks: CompactWireless and Wired Sensing Systems, M. Ilyas and I.Mahgoub, Eds., CRC Press, 2004.

[4] L. Buttyán and J.-P. Hubaux, Security and Cooperationin Wireless Networks: Thwarting Malicious and SelfishBehaviour in the Age of Ubiquitous Computing, Cam-bridge Univ. Press, 2008.

[5] R. Zhang and Y. Zhang, “Wormhole-Resilient SecureNeighbor Discovery in Underwater Acoustic Networks,”Proc. IEEE INFOCOM, 2010.

[6] Y. Liu, J. Jing, and J. Yang, “Secure Underwater Acous-tic Communication Based on a Robust Key GenerationScheme,” Proc. ICSP, 2008.

[7] F. Hu, S. Wilson, and Y. Xiao, “Correlation-Based Securi-ty in Time Synchronization of Sensor Networks,” Proc.IEEE WCNC, 2008.

[8] C. Tian et al., “Tri-Message: A Lightweight Time Syn-chronization Protocol for High Latency and Resource-Constrained Networks,” Proc. IEEE ICC, 2009.

[9] C. Tian et al., “Localization and Synchronization for 3DUnderwater Acoustic Sensor Networks,” in UbiquitousIntelligence and Computing, LNCS, Springer, 2007, pp.622–31.

[10] N. Chirdchoo, W.-S. Soh, and K. Chua, “MU-Sync: ATime Synchronization Protocol for Underwater MobileNetworks,” Proc. WUWNet, 2008.

[11] H. Song, S. Zhu, and G. Cao, “Attack-Resilient TimeSynchronization for Wireless Sensor Networks,” Ad HocNet., vol. 5, no. 1, 2007, pp. 112–25.

[12] F. Hu et al., “Vertical and Horizontal SynchronizationServices with Outlier Detection in Underwater AcousticNetworks,” Wireless Commun. Mobile Comp., vol. 8,no. 9, 2008, pp. 1165–81.

[13] M. Erol and S. Oktug, “A Localization and RoutingFramework for Mobile Underwater Sensor Networks,”Proc. IEEE INFOCOM, Apr. 2008.

[14] W. Cheng et al., “Underwater Localization in Sparse 3DAcoustic Sensor Networks,” Proc. IEEE INFOCOM, 2008.

[15] Z. Zhou, J.-H. Cui, and A. Bagtzoglou, “Scalable Local-ization with Mobility Prediction for Underwater SensorNetworks,” Proc. IEEE INFOCOM, 2008.

[16] Y. Zhou et al., “A Range-free Localization Scheme forLarge Scale Underwater Wireless Sensor Networks,” JShanghai Jiaotong Univ. (Science), vol. 14, no. 5, 2009,pp. 562–68.

[17] M. C. Domingo and R. Prior, “Design and Analysis of aGPS-Free Routing Protocol for Underwater WirelessSensor Networks in Deep Water,” Proc. UNWAT, 2007.

BIOGRAPHYMARI CARMEN DOMINGO ([email protected])received her Lic. and Ph.D. degrees in electrical engineeringfrom Barcelona Tech University, Spain, in 1999 and 2005,respectively. She currently works as an assistant professorin the Electrical Engineering Department of the same uni-versity. Her current research interests are in the area ofnetwork security, sensor, and wireless networks.

The proper function-ing of these schemesis challengingbecause they do notwork well in mobileenvironments, thetime required todetect compromisednodes increasessubstantially inUWCNs due to thelong propagationdelays, and theymust be adapted totolerate short-termdisruptions.



Cognitive engine/data archive

Sensor A


INTRODUCTIONThe need for advanced radio systems that canaccommodate more users, provide higherthroughput, and support higher mobility is grow-ing. This has led to an increased use of spectrumresources. However, radio spectrum is a limited

resource, and the spectrum for each service hasbeen allocated mostly through fixed spectrumassignment. Hence, spectrum scarcity is immi-nent. On the other hand, recent studies revealthat some frequencies allocated for some radioaccess technologies (RATs) are underutilized [1].

Dynamic spectrum access (DSA) promises toprovide an efficient way to utilize spectrumresources through enabling technologies such ascognitive radio (CR). To achieve DSA, a cogni-tive radio system (CRS) must acquire radio envi-ronment knowledge. This will be done in manycases through spectrum sensing. Based on thesensing information, the system analyzes thespectrum usage and makes decisions on spectrumaccess [2, 3]. Apparently, the sensing functionrealized by spectrum sensors plays a fundamentalrole in the DSA process of a CRS. It is mostimportant that the sensing information providedby spectrum sensors can be conveyed to otherentities or units in the CRS, which in IEEEP1900.6 is defined as a client of the sensors.

Recently proposed advanced radio systemsbased on sensing technology combine sensingand cognitive engines (CEs) that use the sensingresults in proprietary transceiver and CR archi-tectures (e.g., those being worked on withinIEEE P802.22 [4]). This model of proprietarydevelopment usually reduces innovation and, inthe past in cases like closed operating systems,has limited the opportunities for integrating newcomponent technologies for better system per-formance. In the case of CR, proprietary inter-faces also limit the degree and capability atwhich different types of sensors and clients caninteroperate. In order to make the developmentof spectrum sensors independent of the evolu-tion and development of advanced wireless com-munication systems, and to ensure the

KLAUS MOESSNER, UNIVERSITY OF SURREYHIROSHI HARADA, CHEN SUN, YOHANNES D. ALEMSEGED, AND HA NGUYEN TRAN, NATIONAL

INSTITUTE OF INFORMATION AND COMMUNICATIONS TECHNOLOGYDOMINIQUE NOGUET, CEA-LETI

RYO SAWAI AND NAOTAKA SATO, SONY CORPORATION

ABSTRACTThe evolution of future wireless communica-

tion systems imposes a strong requirement on theefficiency of spectrum usage, which is expected tobe leveraged by interacting and cooperating cog-nitive radios forming wider cognitive radio sys-tems. Dynamic spectrum access is a potentialmeans to improve spectrum usage. A key step inrealizing DSA is to obtain spectral occupancyinformation provided by spectrum sensors. Sub-sequently, nodes in a CRS analyze spectrumusage to find unused spectrum (often referred toas white spaces). Then the system makes a deci-sion on the best opportunities considering regula-tory policy, transceiver capacity, and so on. Insuch an operation, sensing information exchangeplays a fundamental and key role in enabling effi-cient DSA. This article presents the technicalissues and IEEE standardization activities relatedto sensing information exchange. In particular,the IEEE P1900.6 working group activities aimedat standardizing logical interfaces and data struc-tures required for exchange of sensing relatedinformation between sensors and their clients arediscussed. By explaining the objective, use cases,reference model, data structure, data representa-tion, and generic procedures developed so far,the article presents the main technical aspects ofthe IEEE P1900.6 sensing interface and its use-fulness in CRSs.

SPECTRUM SENSING FOR COGNITIVE RADIOSYSTEMS: TECHNICAL ASPECTS AND

STANDARDIZATION ACTIVITIES OF THEIEEE P1900.6 WORKING GROUP

The authors presentthe technical issuesand IEEE standardization activities related tosensing informationexchange. In particular, IEEE P1900.6 working group activities are discussed.

MOESSNER LAYOUT 2/7/11 10:42 AM


compatibly and coexistence of different sensorsfrom various manufacturers that implement avariety of sensing techniques, a standard is need-ed which defines the interface between spectrumsensors and their clients (i.e., the functional enti-ties that exploit sensing information obtainedfrom the sensors).

As the IEEE DYSPAN Standards Committee(formerly Standards Coordinating Committee 41— SCC41), one of the active working groups(WGs) within the IEEE P1900.6 WG was estab-lished in 2008 to address the above need to stan-dardize the information exchange betweenspectrum sensors and their clients in radio com-munication systems [5]. The logical interface andsupporting data structures used for informationexchange are defined in an abstract mannerwithout constraining the sensing technology,client design, or data link between sensor andclient. By defining the spectrum sensing inter-faces and data structures for DSA and otheradvanced radio communications systems, thestandard will facilitate interoperability betweenindependently developed devices and thus allowfor separate evolution of spectrum sensors andother system functions.

This article introduces the current status ofIEEE P1900.6 standardization activities based onpublished contributions and the P1900.6 draftstandard, which is subject to changes along thecourse of development. The next section presentsthe technical aspects of the IEEE P1900.6 stan-dard in terms of scope and purpose, use cases,reference model, data structure and representa-tion, and generic procedures, respectively. Thefollowing section gives the future direction of theWG. The final section concludes the article.

THE P1900.6 APPROACH

This section provides the technical scope, usecases, and requirements considered in the WG.

SCOPE OF IEEE P1900.6 STANDARDThe scope of the standard, as stated in the pro-ject authorization request (PAR), is to define theinformation exchange between spectrum sensorsand their clients in radio communication systems[6]. The logical interface and supporting datastructures used for information exchange aredefined abstractly without constraining the sens-ing technology, client design, or data link betweensensors and their clients. According to the scopeof the standard, the WG defines the informationexchange between spectrum sensors and theirclients in radio communication systems.

Figure 1 shows a scenario where sensinginformation is exchanged among sensors andtheir clients. The clients include the CE, dataarchive (DA), and sensors. The CE is defined asthe portion of the CRS containing the policy-based control mechanism and the cognitive con-trol mechanism, which must have knowledgeabout the current state and a set of attainablestates, and may have knowledge about the costassociated with (state) transitions of the recon-figurable radio platform. The DA is defined as alogical entity where sensing information obtainedfrom spectrum sensors or other informationsources, and regulatory and policy information

are stored systematically. The sensor can play aclient role of another sensor; that is, a sensorcan receive sensing information from one ormore sensors and forward it to CE or DA.

Figure 2 gives an abstract view of the abovedescribed scenario and illustrates different inter-faces between spectrum sensors and their clientsthat are within the scope of the P1900.6 standard.

The CE/DA-S interface is used for exchang-ing sensing information between a CE or DAand a sensor. As an example, the CE/DA-S inter-face is used in scenarios where a given CE orDA obtains sensing information from one orseveral sensors, or a given sensor provides sens-ing information to one or several CEs or a DA.

The S-S interface is used for exchanging sens-ing information between sensors; this is neededin cases where one sensor may not be able toobtain all required information, or in scenarioswhere sensors A and B exchange sensing infor-mation for collaborative or cooperative sensing.

The CE-CE/DA interface is used for exchang-ing sensing information between CEs or betweena CE and a DA. As an example, the CE-CE/DAinterface is used in scenarios where CEs A andB exchange sensing information for collaborativeor cooperative sensing. The CE-CE/DA inter-face is also used in scenarios where a CE obtainssensing information as well as policies (e.g.,sensing and access rules) and regulatory infor-mation from a DA.

USE CASESAt the current stage the WG has defined 31 usecases within the scope and purpose of the stan-dard. These use cases represent the situations

Figure 1. CRS scenario where the IEEE P1900.6 interface can be employed forsensing information exchange among different entities.

IEEE P1900.6 logical interface

Cognitive engine

Standalone spectrum sensor

Embedded spectrum sensor

Data archive

Primary user

Secondary user

Secondary system base station

Primary userbase station



where the upcoming P1900.6 interface will beemployed in a CRS. In this article we give oneuse case as an example. A complete descriptionof the use cases is given in [7].

CRS with distributed sensors: The CRS ana-lyzes spectrum usage and makes a decision onspectrum access based on the sensing informa-tion provided by sensors located at differentphysical positions.

For example, these sensors can be deployedin certain service areas such as a city hot spotarea or a university campus area to form a spec-trum sensor network. A CR terminal (e.g., acognitive mobile phone) can initiate the sensingfunction of the sensing network once it entersthe service area of the sensor network.

System Model — Different applications that imple-ment CR may have different ways of exchangingsensing information and of using sensing infor-mation interfaces. This subsection provides anabstract system model that is not dependent ona particular topology and spectrum usage. Basedon the system model, the use cases are classifiedin the following subsection.

The system model consists of three scenarios.Figure 3a shows a single CE/DA and single sen-sor scenario, which is denoted the 1:1 scenario.In this scenario, a single sensor provides sensinginformation to one CE/DA. This is the simplestscenario and is considered as a reference. Figure3b shows a single CE/DA and multiple sensorsscenario, which is denoted the 1:N scenario. Nnumber of sensors provide sensing informationto a CE/DA. In other words, one CE/DA canaccess sensing information from N sensors. In

cooperative sensing and collaborative sensing,the CE/DA can access multiple sensors andmake DSA based on sensing information fromdistributed sensors. Also, multiple sensors canexchange information and provide the CE/DAwith an improved sensing result. Figure 3c showsan M number of CEs/DAs and single sensor sce-nario, which is denoted the M:1 scenario, wheremultiple CEs/DAs access sensing information ofone sensor. In other words, one sensor providessensing information to multiple CEs/DAs. In asituation where the sensor can be accessed bymultiple CEs/DAs, these CEs/DAs can shareusage of the sensor. Or if one of the sensors iscapable of accessing a sensor, other CEs/DAscan relay sensing information from a sensor tothe CE/DA. For the case of multiple CEs/DAsand multiple sensors, the system can be decom-posed into a combination of the first three sce-narios. Note that this is understood within thecontext of model description and does not implydecomposition on the implementation level.

Use Case Classification and Analysis — Differentapplications that implement CR may have differ-ent ways of exchanging sensing information andof using sensing information interfaces. To ana-lyze the use cases and extract sensing require-ments, the WG classified various use cases basedon how the sensing information and interfaceare utilized. For this classification, the systemmodel described earlier is used.

From the spectrum usage perspective thereare two usage models: long-term and short-term.The long-term spectrum usage model refers tothe situation where spectrum is used over a rela-tively long period. In most cases the spectrum isused as a substitute primary user service. In theshort-term spectrum usage model spectrum isused over short periods (e.g., autonomous spec-trum access). Within each usage model, usecases are further classified into three scenarios(i.e., 1:1, 1:N, and M:1) based on the systemmodel given earlier. Such an approach clarifieswhich interfaces are involved during the infor-mation exchange. Furthermore, it assists in gen-erating the sensing parameters for each use case.Table 1 presents the classification of selected usecases.

SENSING REQUIREMENTSIn many scenarios involving CR, a communica-tion device needs to capture the current usage ofthe spectrum before establishing its own commu-nication. This behavior is referred to as detectingfree bands, which means identifying frequencybands that are free of already established com-munications. Band BL can be considered free ifthe signal received in this band BL is only madeof noise. On the other hand, if signals are detect-ed in the presence of noise, the band is declaredoccupied. Thus, the function the detector has toperform is that of detecting signals in the pres-ence of noise, which can be stated as the follow-ing hypothesis:

H0: r(t) = n(t) H1: r(t) = hs(t) + n(t),

where H0 is the hypothesis that BL is a freeband, and H1 assumes that BL is occupied. n(t) is

Figure 2. Scope of the P1900.6 standard.

CE/DA-S interface between cognitive engine or data archive and sensor to exchange sensing information and sensing control information.

*The client role can be taken by a cognitive engine, sensor, or data archive.

Cognitive engine/data archive Cognitive engine

Sensor A Sensor B

P1900.6 interfaces*

S-S interface between sensor and sensor to exchangesensing information and sensing control information.

CE-CE/DA interface between cognitive engine and cognitive engine or data archive to exchange sensing information and sensing control information.



noise, and s(t) is a telecommunication signal sentover communication channel h. The major issuein asserting a band is free is that the sensor mustbe able to detect signals at very low SNRs. Manytechniques have been studied for signal detec-tion and recently adapted to the CR case. Thesetechniques can be classified according to the apriori knowledge of the sensor on the signal todetect.

Simple energy detectors (EDs) have limitedperformance but exhibit low complexity and no apriori knowledge is needed, whereas featuredetectors (e.g., cyclostationarity detectors) out-perform EDs but at the cost of higher complexi-ty and the additional requirement of someknowledge about the waveform to detect. A sur-vey of spectrum sensing techniques has been car-ried out for the P1900.6 WG and is captured in[8]. This survey highlights the fact that the choiceof a sensing technique very much depends onscenarios and hardware resources available. Forinstance, in TV white space scenarios, the prima-ry users (PUs) are known, and the sensor canutilize a priori knowledge about the PU toimprove its detection capability, whereas in moreambitious scenarios such as technology agnosticspectrum usage, a priori knowledge is almostimpossible to consider.

Bearing in mind this variety of cases andsolutions, it is not possible for a group likeP1900.6 to focus on specific sensing techniquesin order to define the sensor interface require-ments. The aim of the WG is rather to considermore abstract features that can be generalized toa large number of systems, while using the spe-cific sensing scheme analysis to cross-check howthese requirements stand against practical imple-mentation, and to see if some specific cases leadto additional requirements.

REFERENCE MODEL

To give the industry a guideline on how to imple-ment the logical interface, the WG defines a ref-erence model, shown in Fig. 4. In general, theP1900.6 logical interfaces in P1900.6 definedentities (i.e., sensor, CE, and DA) have the ref-erence model shown in the figure. Sensors andtheir clients can have all service access points(SAPs) or a subset of the three SAPs (i.e., appli-cation SAP [A-SAP], communication SAP [C-

SAP], and measurement SAP [M-SAP]); theirlogical positions and boundaries are depicted inFigs. 4 and 5. They utilize distinct SAPs to real-ize the logical interface.

The P1900.6 services in the reference modelrefer to functional blocks inside sensors andclients required to realize logical interfacesbetween sensors and clients. In particular, thesefunctional blocks are responsible for generatingdata structures defined by the P1900.6 standardfor exchanging sensing and sensing control infor-mation.

The A-SAP defines a set of generic primitivesand data structures to control the P1900.6 entityand/or obtain the sensing results for applicationpurposes. For example, this SAP is used for aP1900.6 entity to utilize sensing information forits purpose (e.g., policy investigation and analysisof spectrum usage). The A-SAP may providefunctions to set up a configuration of P1900.6entities (e.g., sensors and cognitive engine), toconfigure these for collaborative sensing, to startthe data acquisition and processing (e.g., policyprocessing), and to obtain the results of P1900.6processing in order to configure the radio fre-quency (RF) interface accordingly.

The M-SAP is used by P1900.6 entities toaccess P1900.6 compliant services provided bythe station’s hardware and/or firmware to con-trol the spectrum measurement module, such ascollocated physical spectrum measurement mod-ule (i.e., analog-to-digital/digital-to-analog con-verter [ADC/DAC], filtering, signal condition,etc.), and acquire measurement data from these.For example, a station (terminal) utilizes its RFinterface during idle times for spectrum mea-surement and provides RF spectrum data to col-located P1900.6 sensor entities that registered atthe local M-SAP. The M-SAP shall be instantiat-ed by the station’s P1900.6 compliant measure-ment function.

The C-SAP is used for sensing information(sensing message, sensor message, control mes-sage, and regulatory information) exchangebetween sensors and their clients. The client rolecan be taken by sensor, CE, and DA. It abstractscommunication mechanisms from P1900.6 enti-ties by providing a set of generic primitives andmapping these primitives on transport protocols.A P1900.6 compliant message transport servicehas to be provided by the station in order to

Figure 3. IEEE P1900.6 system model: a) 1:1 scenario; b) 1:N scenario; c) M:1 scenario.

Cognitive engine/Data archive

Sensor

(a) (b)

P1900.6 interfaces

Cognitive engine/Data archive

Sensor A Sensor B

P1900.6 interfaces

(c)

Cognitive engine/ Data archive

Sensor

P1900.6 interfaces

Cognitive engine

Many techniqueshave been studied

for signal detectionand recently adapted

to the CR case.These techniques can

be classified according to the

a priori knowledgeof the sensor on the

signal to detect.



locate a remote P1900.6 peer entity and establisha communication link with this entity. Messageexchange then takes place as defined by thisstandard, and it is the responsibility of the trans-port service to map these message transfers to asuitable transport, network or link layer commu-nication. For example, a P1900.6 entity CE cantake the role of a client to a remote P1900.6entity sensor by sending a request to configurethe sensor for delivering measurement data tothe requestor utilizing its C-SAP functionality.The C-SAP shall be instantiated by the P1900.6compliant message transport service. The ser-vices available at these SAPs realize the logicalinterfaces (i.e., CE/DA-S interface, S-S interface,and CE/DA-CE interface) between differentP1900.6 entities.

Figure 5 provides alternative views of thereference model. The figure shows how theabove mentioned P1900.6 services and SAPsare involved in the sensing informationexchange between sensors and their clients.Figure 5a shows a situation where the CE/DAand sensor exchange sensing information. Thecontrols/applications at the CE/DA generatesensing information requests (desired frequen-cy band to sense, desired sensing performance,etc.). The P1900.6 services receive theserequests at the A-SAP and convert theserequests into a message format defined by thecommonly understood interface. The messageis sent to the C-SAP. The communication sub-system provides communication services andsends the requests as messages to a sensor that

Table 1. Classification of use cases.

Spectrumusagemodels

Scenariosof systemmodel

Use cases and application examples

Lon

g t

erm

sp

ectr

um

usa

ge

1:1

Load sharing to reduce blocking at peak traffic times: two networks that exhibit peak loads at different times andat different locations can relieve network congestion by offloading their peak traffic to one another using CR.

Introduction of new users or services: The use of CR to find and utilize spectrum that is underutilized at aspecific time and location would allow the introduction of new users or services without significant delay.

Worldwide mobility: A CR radio could do automatic RAT switching to avoid missed calls in the event that auser travels to a different country or location.

1:N

Emergency services: CR allows all of the emergency personnel to find common usage channels throughout thedisaster area which can be operated without interference from any of the surviving legacy communications

Dynamic spectrum assignment: In the dynamic spectrum assignment use case, frequency bands aredynamically assigned to the RANs in order to optimize radio resource usage and improve quality-of-service.

M:1

Self management of uncoordinated spectrum: CR provides an effective means of self coordination so as toavoid interference with other networks while providing useful throughput, for instance in unlicensed bandswhere central coordination might not be practical.

Dynamic spectrum assignment: In the dynamic spectrum assignment use case, frequency bands aredynamically assigned to the RANs in order to optimize radio resource usage and improve quality-of-service.

Sho

rt t

erm

sp

ectr

um

usa

ge

1:1 Add capacity for emergency: Dynamic spectrum access, or the ability for CRs to identify unused or underutilizedspectrum, could be used in emergency scenario and provide a means for expanding capacity when needed.

1:N

Network extension for coverage: This can be achieved by automatically reconfiguring the CRs to include a repeatercapability to extend network coverage to areas where radios are otherwise cut off from their infrastructure.

Dynamic spectrum sharing: In the dynamic spectrum sharing use case, frequency bands assigned to RANs arefixed. However, a particular frequency band can be shared by several RANs to optimize radio resource usageand improve quality-of-service.

M:1

Policy violation: This use case is focuses on the usage of spectrum sensing in CR for the investigation of policyviolations.

Cognitive relay: In cognitive wireless network, multiple cognitive terminals form a wireless ad hoc network ormesh network. Such network can be considered as having multiple peer to peer links. Each radio in the peer topeer link detects the behavior of the radio frequency environment in multiple channels and provides efficientcommunication links and paths.

Priority based sensing information provision: when multiple clients contend to access the same sensinginformation source at a time, clients that have a high priority level can access the sensing information first.

Distributed radio resource usage optimization: In this use case, frequency bands assigned to RANs are fixed.Also, reconfiguration of RANs is not considered in this use case. Instead, reconfigurable terminals with orwithout multi-homing capability are considered.


also has a communication subsystem through aradio channel (wireless LAN, ultrai wideband[UWB], etc.). The P1900.6 services at the sen-sor unpack the messages based on the com-monly understood data structure and obtainthe requests sent from the CE/DA. Then thesesensing requests are sent to the spectrum mea-surement through the M-SAP. Thehardware/firmware performs the spectrummeasurement and produces sensing results.The P1900.6 services at the sensor obtain thesensing results from the M-SAP according tothe commonly understood data structure.Through the C-SAP, the sensor sends the sens-ing results to the client CE/DA. The P1900.6services at the CE/DA unpack the sensingresults based on the commonly understooddata structure and provide the sensing infor-mation to the controls/applications through theA-SAP.

Similarly, Fig. 5b provides an alternative viewof the reference model for the situation where twosensors exchange sensing information. Sensinginformation received from one sensor can be pro-cessed at another sensor (the client). For example,the sensing information from two sensors can becombined to provide improved sensing informa-tion. Finally, Fig. 5c provides an alternative viewof the reference model for the situation where oneCE can access one sensor and share sensing infor-mation with another CE/DA.

DATA STRUCTUREBased on the use case analysis, reference model,and sensing techniques the WG has listed fourtypes of information (i.e., sensing information,sensor information, control information, andpolicy information). A sensing message relates tomeasurement data (e.g., sensing results) fromsensors and clients. A control message relates tothe control of sensing activity of sensors orclients, such as target performance and sensingduration. A sensor message relates to sensorspecification, sensor capability, or sensor identi-ty. Regulatory information is defined as sensingand access rules derived from radio spectrumregulation, such as required sensing durationand sensitivity levels.

Of course, all these messages depend signifi-cantly on the regulatory requirements. Forinstance, if TV white space detection assumesthat digital TV channels of 6 MHz must bedetected with a sensitivity of –114 dBm, theclient must ensure that the sensor can providethis performance level. Then, when all messagesare identified, an important task is to define howthese messages are represented in order to bewell adapted to real case implementation. Forinstance, the message format is to be kept rea-sonably complex while generic enough to sup-port all cases investigated.

Each parameter of these four types of infor-mation is defined by giving a short textualdescription, parameter ID, name, unit, parame-ter type, and size, as well as range and resolu-tion. The textual description explains thepurpose of the parameter. The parameter’sname provides a unique identification of theparameter in human readable form such as startfrequency, detection threshold, and sensor location,

whereas the numerical ID is given to unambigu-ously identify the parameter in the process ofinformation exchange between P1900.6 logicalentities. The size of a parameter is defined asthe number of elements contained in the param-eter. Data types include primitive types such asinteger and Boolean, as well as complex andderived types such as enumeration and struc-tured. Data unit is used to describe the physicalunits of the data.

For example, bandwidth is defined as anunbounded parameter with the unit frequency. Itis defined by the lower and upper frequencies,thus having the array (frequency) type and size2. Here, frequency is also a defined parameterand used as a data type of the elements in thisparameter.

DATA REPRESENTATIONTo represent various sensing related informationdefined in the IEEE P1900.6 WG while allowingfor extendibility of uses to add more sensinginformation, the object-oriented Unified Model-ing Language (UML) is employed. Sensing relat-ed information is categorized into four classes:control information, sensor information, sensinginformation, and regulatory requirement.

The control information class contains theinformation to control the transport of sensinginformation. It also contains the information toset the spectrum measurement, such as measure-ment objective, measurement profile, and mea-surement performance. The sensor informationclass contains information to describe spectrumsensors. For example, this class contains the sen-sor’s physical profile, such as antenna profile andlocation profile, and manufacturer’s information.Information that describes the spectrum mea-surement capability is also included in this class.The sensing information class contains mainlythe measurement data such as measured valuesof signals, channels, and RATs. It also containsthe information related to the measurement thathas been carried out, such as where, when, andhow the measurement has been done. Finally,the regulatory requirement class contains differ-ent regulatory information related to differentprimary signals in different regions and countries.

Figure 4. Reference model of P1900.6 standard.


Control/application

Spectrummeasurement

IEEE P1900.6 service

Communication subsystem

App

licat

ion

SAP

(A-S

AP)

Mea

sure

men

t SA

P(M

-SA

P)

Communication SAP (C-SAP)



USE OF THE INTERFACES:GENERIC PROCEDURE

This subsection gives a generic procedure and itsusage examples. The generic procedure can beseen as a basic building block for a more com-plex procedure required for the exchange ofsensing related information between one or mul-tiple sensors and their clients.

To express the procedure, the notions of ser-vice user and service provider model as specifiedby ITU-T X.210 [9] are used as a reference. Theservice provider refers to an abstraction of thetotality of those entities that provide an P1900.6service to the service user. The P1900.6 serviceprovider includes the P1900.6 service (as indicat-ed in the reference model in the previous sec-tion), and the associated SAPs that instantiatethe logical interface for exchange of sensingrelated information between the P1900.6 serversand clients. The service user refers to the IEEEP1900.6 logical entities that shall implement oruse the P1900.6 logical interface to exchange

sensing information. For example, the P1900.6logical entity that shall use the logical interfaceto obtain sensing related information plays theclient role.

Here we give a simple use case where the CEobtains sensing information from a logical entitysensor to exemplify the procedure. Both the CEand sensor are P1900.6 service users. They areusing the services provided by the P1900.6 serviceprovider to realize information exchange, wherethe CE is playing the client role and the sensor isplaying the server role of a client/server model:1. When a CE needs to obtain sensing informa-

tion, it issues a request to the P1900.6 serviceprovider.

2. Then the P1900.6 service provider forwardsthe request along with its parameters towarda sensor. The parameters from the P1900.6logical interface form the service data unit(SDU), which shall be communicated betweenCE and sensor as the payload of a protocoldata unit (PDU) used by the protocol layersbeneath. Upon reception of an SDU, theP1900.6 service provider generates an indica-tion toward the sensor. The actual procedureand protocol used in the communication sub-system is out of the scope of this standard.Upon receiving the indication, the sensor per-forms further processing to consider or rejectthe request.

3. In response to the indication, the sensor issuesa response to the P1900.6 service provider tosend the results (e.g., the sensing information)as requested by the CE.

4. The P1900.6 service provider generates a con-firmation toward the CE while providing theSDU as originated by the sensor. For this typeof synchronous communication flow, the con-firmation indicates that the P1900.6 requesthas been completed, and the CE may nowcheck if it has obtained valid results.Note that in the above communication mes-

sage flow, only the service request from theclient (i.e., CE) is depicted. In this case the sen-sor is stated as acceptor while the CE is stated asrequestor. The requestor and acceptor roleschange depending on who is initiating the ser-vice request.

FUTURE DIRECTION

The P1900.6 draft standard passed the sponsorballot in late 2010 and is currently with the Stan-dards Board for approval. DYSPAN workinggroup 6 continues its work, and is currentlydeveloping a new project aiming to complementthe definitions of the current draft standard.

CONCLUSION

This article presents the activities of the IEEEP1900.6 WG, which is currently defining aninterface of sensing information exchangebetween sensors and their clients. By explainingthe scope, use cases, and system model referencemodels we have shown how the interface can beused in various use cases of CRSs. The interfacedefined in this standard will ensure the interop-erability of sensors and other system compo-nents of the CRS. Thus, the standard will bringFigure 5. Alternative views of the reference model.

(a)

A-SAP

Control/application

Control/application

P1900.6 service

Communication subsystem(data bus, wired or wireless channel)

C-SAP

M-SAP

Spectrummeasurement

P1900.6 service

C-SAP

Client (CE/DA) Sensor

(c)

A-SAP

Control/application

P1900.6 service

Communication subsystem(data bus, wired or wireless channel)

C-SAP

A-SAP

Control/application

P1900.6 service

C-SAP

Client (CE) CE/DA

(b)

M-SAP

Spectrummeasurement

P1900.6 service

Communication subsystem (data bus, wired or wireless channel)

C-SAPA-S

AP

M-SAP

Spectrummeasurement

P1900.6 service

C-SAP

Client (Sensor) Sensor



freedom to manufacturers to develop their ownadvanced sensing techniques.

ACKNOWLEDGMENTThe authors would like to express their sinceregratitude to all the participants of the IEEEP1900.6 WG.

REFERENCES[1] G. Stap le and K. Werbach, “The End of Spectrum

Scarcity,” IEEE Spectrum, vol. 41, no. 3, Mar. 2004, pp.48–52.

[2] J. Mitola III and G. Q. Maguire Jr., “Cognitive Radio:Making Software Radio More Personal,” IEEE Pers.Commun., vol. 6, no. 4, 1999.

[3] S. Haykin, “Cognitive Dynamic Systems,” Proc. IEEE, vol.94, no. 11, Nov. 2006, pp. 1910–11.

[4] IEEE P802.22/ DRAFTv2.0, “Draft Standard for WirelessRegional Area Networks Part 22: Cognitive WirelessRAN Medium Access Control (MAC) and Physical Layer(PHY) Specifications: Policies and Procedures for Opera-tion in the TV Bands.”

[5] IEEE SCC41; http://www.scc41.org/.[6] IEEE P1900.6 WG; http://grouper.ieee.org/groups/scc41/

6/index.htm.[7] IEEE Std. 1900.6/D0.6, “IEEE Draft Standard for Spec-

trum Sensing Interfaces and Data Structures for Dynam-ic Spectrum Access and Other Advanced RadioCommunication Systems,” Feb. 2010.

[8] D. Noguet et al., “Sensing Techniques for CognitiveRadio — State of the Art and Trends,” IEEE SCC41 —P1900.6 Working Group,” White Paper, Apr. 2009;http://grouper.ieee.org/groups/scc41/6/documents/white_papers/P1900.6_WhitePaper_Sensing_final.pdf.

[9] ITU Rec. X.210, “Information Technology (Open SystemInterconnection) Basic Reference Model: Conventionsfor the Definition of OSI Services.”

BIOGRAPHIESKLAUS MOESSNER ([email protected]) is a professorialresearch fellow in the Centre for Communication SystemsResearch at the University of Surrey, United Kingdom. Heearned his Dipl-Ing (FH) at the University of Applied Sci-ences, Offenburg, Germany, an M.Sc. from Brunel Universi-ty, and his Ph.D. from the University of Surrey. His researchinterests include dynamic spectrum allocation, cognitiveradio networks reconfiguration management, service plat-forms, and adaptability of multimodal user interfaces. He ischair of the IEEE SCC41 WG6 defining interfaces and datastructures for DSA.

HIROSHI HARADA [M] ([email protected]) is director of theUbiquitous Mobile Communication Group, National Insti-tute of Information and Communications Technology(NICT), Yokosuka, Japan. He is also director of NICT Singa-pore Wireless Communication Laboratory. He joined theCommunications Research Laboratory, Ministry of Postsand Communications (currently NICT), in 1995. His researchinterests include software-defined radio, cognitive radio,dynamic spectrum access networks, and broadband wire-less access systems on the microwave and millimeter-wavebands. He currently serves on the Board of Directors of theSDR Forum, and has been Chair of the IEEE StandardsCoordinating Committee 41 (IEEE SCC41; IEEE P1900) since2009 and Vice Chair of IEEE P1900.4 since 2008. Moreover,he was Chair of the Institute of Electronics, Information,and Communication Engineers (IEICE) Technical Committeeon Software Radio from 2005 to 2007. He was the recipi-ent of the Achievement Award and made a Fellow fromIEICE in 2006 and 2009, respectively, and received theAchievement Award from the Association of Radio Indus-tries and Businesses (ARIB) in 2009 on the topic of SDRand cognitive radio research and development.

CHEN SUN [S’02, M‘05] ([email protected]) received a B.E.degree in electrical engineering from Northwestern Polytech-nical University, Xi’an, China, in 2000 and a Ph.D. degree inelectrical engineering from Nanyang Technological Universi-ty, Singapore, in 2005. From August 2004 to May 2008 hewas a researcher with ATR Wave Engineering Laboratories,Kyoto, Japan. In June 2008 he joined the Ubiquitous MobileCommunications Group, NICT, Yokosuka, Japan, as an ExpertResearcher working on cognitive radio. His research interests

include spectrum sensing, dynamic spectrum access, smartantennas, and cooperative communications. He is a votingmember of IEEE Standards Coordinating Committee 41. Heis also a voting member of the IEEE 1900.6 Working Groupand serves as the Technical Editor.

YOHANNES D. ALEMSEGED [S’06, M‘08] ([email protected])received a B.Sc. degree in electrical and electronic technol-ogy from Nazareth Technical College (currently Adama Uni-versity), Nazareth, Ethiopia, in 1997, an M.Sc. degree inelectrical engineering from Addis Ababa University,Ethiopia, in 2002, and a Ph.D. degree from Graz Universityof Technology, Austria, in 2008. He is currently an expertresearcher with NICT, Yokosuka, Japan. His research inter-ests include digital signal processing for communications,spectrum-sensing algorithms for cognitive radios, and low-complexity ultra-wideband transceivers. He is a votingmember of IEEE Standards Coordinating Committee 41 andserves as the Committee Secretary. He is also a votingmember of the IEEE 1900.6 Working Group.

HA NGUYEN TRAN [M‘08] ([email protected]) received hisB.E. and M.E. degrees in electronics and information engi-neering and a Ph.D. degree in information science and tech-nology from the University of Tokyo, Japan, in 2000, 2002,and 2005, respectively. He is currently an expert researcherwith the New Generation Wireless CommunicationsResearch Center, NICT, Yokosuka, Japan. His research inter-ests include wireless networking and cognitive radio. He is avoting member of IEEE Standards Coordinating Committee41. He is also a voting member of the IEEE 1900.6 WorkingGroup and serves as the Working Group Secretary.

DOMINIQUE NOGUET joined CEA-LETI in 1998 where he hascarried out digital communication hardware architectureresearch. His research activities cover reconfigurable andflexible radio, and more recently cognitive radio. Since2001 he has had several national and international projectmanagement positions in the field of wireless communica-tion. He currently leads flexible radio activities (WPRC)within the European Network of Excellence NEWCOM++and is the technical manager of QoSMOS, a major EU pro-ject on cognitive radio. He received a best paper awardand the best Ph.D. award from INPG. He has authored orco-authored about 40 papers in peer reviewed journals andconferences. He is currently head of the Digital Architec-tures and Prototypes group at LETI, where he also leadscognitive radio activities.

RYO SAWAI ([email protected]) received his B.E., M.E.and Ph.D. degrees in electrical and electronic engineeringfrom Chuo University, Tokyo, Japan, in 1998, 2000, and2002, respectively. From July 1999 to September 1999 hejoined the Nokia student exchange program in Helsinki,Finland. From October 1999 to May 2002 he was a studenttrainee with NICT, Yokosuka, Japan. In April 2002 he joinedthe Ubiquitous Technology Laboratories, Sony Corporation,Japan. He is currently a researcher at System TechnologyLaboratories, Sony Corporation, Japan. His research inter-ests include advanced coding design (e.g., turbo code/decoder and low density parity check code/decoder), recon-figurable hardware architecture design, dynamic spectrumaccess, and multi-input and multi-output signal processingtechnologies. He received the IEEE VTS Japan Researcher’sEncouragement Award in 2001. He is a voting member ofIEEE Standards Coordinating Committee 41. He is also avoting member of the IEEE 1900.6 Working Group.

NAOTAKA SATO ([email protected]) received B.E.and M.E. degrees in electrical engineering from Tokyo Uni-versity of Science, Japan, in 1991 and 1993. From April1993 to April 1995 he was an engineer in Sony Corpora-tion, Tokyo, Japan. From April 1995 to May 1999 he wasan RF engineer in Sony Electronics, Inc., San Diego, Califor-nia. From May 1999 to September 2001 he was a senior RFengineer in Sony Corporation, Tokyo, Japan. From October2001 to March 2005 he was senior RF engineer in SonyEricsson Mobile Communications Japan, Inc., Tokyo, Japan.In April 2005 he joined the Communication TechnologyLaboratory, Sony Corporation, Tokyo, Japan, as a seniorresearcher working on cognitive radio and 4G cellular tech-nology. His research interests include dynamic spectrumaccess and RF system architecture. He is a voting memberof IEEE Standards Coordinating Committee 41. He is also avoting member of the IEEE 1900.6 Working Group.

The interfacesdefined in this

standard will ensurethe interoperability

of sensors and othersystem components

of the CRS. Thus,the standard willbring freedom to manufacturers to

develop their ownadvanced sensing

techniques.



Cooperative speccommunication be

PUt

SUt1

SUt2


INTRODUCTIONAs we know, almost all existing wireless commu-nication networks are allocated a fixed spectrum,resulting in a large portion of the assigned spec-trum being used sporadically [1–3]. According toa report by the Federal Communications Com-

mission (FCC), the percentage of the assignedspectrum that is occupied ranges only from 15 to85 percent, varying widely in time and geograph-ical position [2]. Hence, the limitation andunderutilization of available spectrum resourceaccelerates new research. Ultra-wideband(UWB) and cognitive radio (CR) are candidatetechniques to improve utilization of the assignedspectrum. Under the power spectral densityemission limit of Part 15, which is –41.3dBm/MHz or significantly lower (as low as –75dBm/MHz), UWB can share wideband spectrumwith other existing wireless systems. However,the application of UWB is limited because of itsultra-wideband frequency range, poor agility,and high complexity. The term cognitive radiowas first introduced by Joseph Mitola. As apromising candidate, CR has the ability to shareor reuse spectrum in an opportunistic manner byemploying spectrum overlay and/or spectrumunderlay approaches, which results in an increaseof spectrum utilization [1–3]. In [1], the authorsdefined CR as an intelligent wireless communi-cation technology that is aware of its surround-ing environment, uses the methodology ofunderstanding-by-building to learn from theenvironment, and then adapts its internal statesto statistical variations in the incoming radio fre-quency stimuli by making corresponding changesin certain operating parameters (e.g., transmitpower, carrier frequency, and modulation strate-gy) in real time.

Moreover, high-data-rate wireless communi-cation systems are limited not only by the limit-ed spectrum, but often more significantly by thefading effects due to multipath propagation, theDoppler effect, and the angle spread of thewireless channel. Diversity is one of the mosteffectual methods to resist the fading effect. Aswe know, multiple-input multiple-output(MIMO), multicarrier modulation (MCM) andcode-division multiple access (CDMA) are

TAO LUO, BEIJING UNIVERSITY OF POSTS AND TELECOMMUNICATIONSFEI LIN, SHANDONG INSTITUTE OF LIGHT INDUSTRY

TAO JIANG, HUAZHONG UNIVERSITY OF SCIENCE AND TECHNOLOGYMOHSEN GUIZANI, KUWAIT UNIVERSITY

WEN CHEN, SHANGHAI JIAOTONG UNIVERSITY

ABSTRACTFor high-data-rate wireless communication

systems, two major issues are the underutiliza-tion of limited available radio spectrum and theeffect of channel fading. Using dynamic spec-trum access, cognitive radio can improve spec-trum utilization. Almost all proposed CRsystems are based on multicarrier modulationsince multiple users can access the MCM sys-tems by allocating subcarriers. Generally, MCMmainly includes two different schemes, orthogo-nal frequency-division multiplexing and filteredmultitone modulation. Considering mutualinterference elimination, synchronization, andtransmission efficiency, we conclude that FMTis better than OFDM in MCM-based CR sys-tems. Additionally, cooperative diversity canreduce the fading effect since the space diversitygain can be obtained through the distributedantennas of each user. Hence, in this article, wecombine CR with the cooperative diversity tech-nique, and then construct three cooperativediversity cognitive models: the collaborativespectrum sensing model, the cooperative com-munication cognitive model, and the hybridmodel. Additionally, radio resources can beextended from time-frequency dimensions tospace-time-frequency dimensions in the pro-posed models, which effectively improves bothspectrum utilization and MCM-CR system per-formance. Finally, extensive simulations areconducted to show the validity and effectivenessof the proposed models.

MULTICARRIER MODULATION ANDCOOPERATIVE COMMUNICATION IN

MULTIHOP COGNITIVE RADIO NETWORKS

The authors combineCR with the cooperative diversitytechnique, and construct three cooperative diversitycognitive models:the collaborativespectrum sensingmodel, the coopera-tive communicationcognitive model, andthe hybrid model.

JIANG LAYOUT 2/7/11 10:43 AM


commonly considered as candidates. The MCMschemes, such as orthogonal frequency-divisionmultiplexing (OFDM) and filtered multitone(FMT) modulation [4], are approaches to over-come the intersymbol interference (ISI) causedby multipath propagation. CDMA can suppressnarrowband noise and interference by spread-ing the signal bandwidth to a wideband spec-trum. Hence, MCM such as OFDM andmulticarrier CDMA (MC-CDMA) are hailed aspromising candidates for realizing spectrumoverlay and spectrum underlay CR applications[2, 5], respectively. In MC-CDMA-based CR(MC-CDMA-CR) systems, it is possible toabandon distributed sensing in a way that thetransmitting secondary user (SU) can spread itssignal across the entire band, including thatoccupied by the primary user (PU), whichresults in a base station and/or signaling chan-nel not being needed. Furthermore, MC-CDMA allows narrowband PU interferers to beexcluded locally at the SU receiver, henceimproving its performance. In other words,MC-CDMA technology is more suitable forspectrum underlay CR applications. However,we mainly focus on spectrum overlay CR sys-tems in this article. Therefore, MC-CDMA-CRis not discussed below.

Additionally, MIMO can improve the channelcapacity and performance of wireless communi-cation systems by using space and time resources.However, it is not sufficient that only one anten-na be built in the mobile station due to the limi-tations of its cost, size, and complexity.Moreover, the MIMO technique does not workwell when the fading is large-scale. Consequent-ly, cooperative diversity [6], also called virtualMIMO, has been proposed, in which users cantransmit data by sharing antennas of other userssurrounding them. Similarly, a cooperative com-munication system can obtain the space diversitygain and improve reliability. The basic ideasbehind cooperative communication can be tracedback to the groundbreaking work of Cover andEl Gamal on the information theoretic proper-ties of the relay channel with additive whiteGaussian noise (AWGN). However, the aim ofrelays is only to help the source transmit infor-mation, whereas the users in cooperative com-munication systems can act as both informationsources and relays. Nosratinia et al. have proventhat even though the interuser channel is noisy,cooperation can still lead not only to an increasein capacity for both users, but also to a morerobust system [6]. In this case, the achievablerates of users are less susceptible to channelvariations. There are three main cooperative sig-naling protocols: amplify-and-forward (AF),detect-and-forward (DF), and coded cooperation(CC) methods [6]. In this field, the key tech-niques mainly include power allocation, coopera-tive partner selection, performance evaluation,and so forth.

In fact, CR and cooperative communicationhave developed rapidly in their own fields.However, there is little advantage in improvingspectrum utilization when only cooperative com-munication is used, whereas CR is not good forimproving the symbol error rate performance ofeach user. Consequently, combining CR with

cooperative communication may be a good solu-tion to both problems. Ghasemi [7] and Gane-san [8] first proposed the collaborative(exchanging spectrum hole vectors with eachother) and cooperative communication schemeto improve the detection probability of spec-trum sensing in CR systems. Nevertheless, bothof them are only based on AF and do not takethe locations of SUs into consideration. Afterthat, Devroye introduced cooperative communi-cation to data transmission, and obtained funda-mental limits of achievable rates in CR systems.Obviously, this model is ideal and needs cooper-ative communications among SUs. Finally,Simeone et al. studied the cognitive relayingscheme between PUs and SUs [9], which is ofcourse a simple and elementary discussion. Con-sequently, in this article we expand the modelsof Ghasemi and Devroye, and give an overviewof the cooperative diversity cognitive models,which combine MCM, cooperative diversity, andCR techniques together.

This article is organized as follows. In thenext section we discuss MCM techniques in themultihop CR system. Then three cooperativediversity cognitive models are proposed. Simula-tion results show the effectiveness of the pro-posed models in the following section.Conclusions are drawn in the final section.

MULTICARRIER MODULATION TECHNIQUESIN MULTIHOP CR NETWORKS

Recently, some work has been reported on mul-tihop CR networks [10], in which each node hasa list of available frequency bands and mustwork adaptively among these frequency bandsbecause of dynamic spectrum access. It is wellknown that two nodes cannot communicate ifthey work on different frequency bands. Hence,routing in multihop CR networks becomes acritical and challenging issue. In general, solu-tions of this problem mainly focus on the meth-ods in the network layer, whose processing delayis on the order of milliseconds. However, thehigh-speed wireless channel in multihop CR net-works varies on the order of microseconds dueto multipath fading, the Doppler effect, anddynamic occupancy of the subchannel by PUs.Therefore, the solutions proposed in the net-work layer may cause heavy interference to PUs.To this end, adopting MCM to intersectionnodes (i.e., RUs) of multihop CR networks maybe a good solution in the physical layer. Becauseof the usage of MCM, the intersection node canallocate some unused subcarriers to differentinformation flows; thus, all flows can be trans-mitted simultaneously.

In fact, considering that the access of multi-ple users can be implemented by the allocationof subcarriers in an MCM system, almost all ofthe proposed spectrum overlay CR systems arebased on MCM technology, specifically theMCM-CR system. Moreover, almost all pro-posed MCM-CR systems are based on OFDM[2], such as IEEE 802.22, the spectrum poolingsystem proposed by Timo A. Weiss, and theNext Generation (xG) communication networksproposed by the Defense Advanced Research

The longer the delayspread of the

channel is, the longer the length

of the CP is, andthus the less the

efficiency is. Obviously, these

methods conflict withthe concept ofimproving the

spectrum utilizationof the CR technique.



Projects Agency (DARPA) [2]. Recently, cosinemodulated multitone (CMT)-based CR (CMT-CR) and FMT-based CR (FMT-CR) systemshave also been proposed [3].

In an MCM-CR system, a maximum likeli-hood detection (MLD) model is deduced underthe constraint of the interference temperature[3]. Moreover, the optimal detection region, andthe probability of detection and false alarm areobtained in [3]. Spectrum allocation and accessalgorithms are well studied in [2, 3]. Based onthe MLD scheme and Markovian chain predic-tion (MCP) model, an efficient SU access algo-rithm based on spectrum hole vector (SHV) isproposed to meet the quality of service (QoS)requirements of the SUs in a centralized MCM-CR system in [3]. However, one issue of theMCM-CR system is that it requires a large sizeinverse discrete Fourier transform (IDFT)/DFToperation due to the varying locations of eachsubcarrier in a wide spectrum, resulting in aheavy implementation complexity and seriousdelay. To this end, a scheme combining Cooley-Tukey’s recursive algorithm with pruning algo-rithm can be adopted. Now, the main issuebecomes: which technique is better in an MCM-CR system, OFDM or FMT? We discuss moredetail in the following.

Recently, some challenges in the physicallayer of OFDM-CR systems have been analyzed(e.g., mutual interference and synchronization)[2]. In OFDM-CR systems, spectrum partition-ing is realized in the form of overlapping sub-bands, in which adjacent subcarriers are at thenulls of the sinc(f) function. Therefore, spectrumefficiency is high. However, the requirement ofsynchronization is also very strict, especially forfrequency synchronization. If the subcarriers arenot orthogonal, the sidelobes of the sinc-shapedspectrum on each subcarrier may fully interferewith PUs even if the parameter of the OFDMsystem used by PUs and SUs is the same. More-over, the worse case is that OFDM is not adopt-

ed by PUs, or the parameters of OFDM for bothPUs and SUs are different even though OFDMis used by PUs. To decrease the interference, thereceived OFDM signal is windowed in the timedomain before it is fed into the operation ofFourier. Another method is to leave some virtu-al subcarriers (VCs) free. Furthermore, in anOFDM-CR system, the cyclic prefix (CP) or so-called guard interval is added to each transmit-ted symbol to avoid ISI, which occurs inmultipath channels and destroys orthogonality.Unfortunately, like VCs, CP leads to a loss oftransmission efficiency. The longer the delayspread of the channel, the longer the length ofthe CP, and thus the lesser the efficiency . Obvi-ously, these methods conflict with the concept ofimproving the spectrum utilization of the CRtechnique.

Compared with OFDM-CR systems, an FMT-CR system does not require CP between severalcontinuous transmitted symbols [4]. Instead, thebandwidth of each subcarrier is chosen to bequasi-orthogonal in the frequency domain, whichis also called subcarrier spectral containment,and it can be achieved by the use of steep rolloffbandpass filters (i.e., filter bank). As a result, thetime domain response of these filters may over-lap in several successive transmitted symbol peri-ods. Therefore, it is necessary for equalizationper subchannel to reduce any remaining ISI,even if the channel is in an ideal state. A highlevel of subcarrier spectral containment is goodfor CR systems because the leakage of signalenergy between adjacent subchannels may beneglected since it is as low as –70 dB where thesubcarriers are closely spaced, as shown in Fig.1. Figure 1 illustrates the frequency response ofthe first five subcarriers in an MCM system,where the solid and dashed lines denote FMTand OFDM, respectively. Due to the tight spec-tral containment achieved by the prototype filterin FMT, negligible power leaks into adjacentbanks. Hence, fewer VCs are needed to complywith the regulatory power spectral mask than inan OFDM system.

Therefore, it is obvious that only a few VCsare needed since the CP is not necessary forFMT systems; thus, the transmission efficiencyof FMT systems is better than that of OFDMsystems. Without loss of generality, we definethe transmission efficiency as

(1)

where Nc, Nvc, and LCP denote the number ofsubcarriers, the number of VCs, and the lengthof the CP, respectively. For example, when thebandwidth of occupied spectrum equals 20 MHz,Nc = 64, LCP = 16, and Nvc = 12 in an OFDMsystem based on the standard of HIPERLAN/2or IEEE 802.11a, while LCP = 0 and Nvc = 2 ~4 in an FMT system. Accordingly, the efficiencyof OFDM and FMT is ηOFDM = 65 percent andηFMT ≈ 94 percent, respectively.

Moreover, synchronization among differentusers is not serious in an FMT system because ofits tight spectral containment. The disadvantageof the FMT is its complexity due to the filterbank and equalization per subchannel. However,

η=+

−=

−+

×N

N L

N N

N

N N

N Lc

c CP

c vc

c

c vc

c CP100%,

Figure 1. Frequency response of the first five subcarriers in MCM system withNc = 64 (solid: FMT, dashed: OFDM).

fT/Nc0.020

-70

-80

Am

plit

ude

/dB

-60

-50

-40

-30

-20

-10

0

0.04 0.06 0.08 0.1



it may not be a serious issue in the future due tothe fast development of digital signal processingtechniques.

In summary, taking mutual interference sup-pression, synchronization, and transmission effi-ciency into consideration, it is much better toadopt FMT than OFDM in MCM-CR systems.

PROPOSED COOPERATIVE DIVERSITYCOGNITIVE MODELS IN AN

MCM-CR SYSTEMWhen the idea of CR is taken into account inthe cooperative communication system, thecooperative communication system with cogni-tive relay can be proposed to further improvethe spectrum efficiency. On the other hand,cooperative diversity CR systems can also beproposed when the idea of cooperative diversityis considered in a CR system. Thus, based onChasemi and Devroye’s model, combiningMCM, cooperative diversity, and CR tech-niques, we propose three cooperative diversitycognitive models, including the collaborativespectrum sensing model, the cooperative com-munication cognitive model, and the hybridmodel, as shown in Figs. 2, 3, and 4, respective-ly. Solid and dashed ellipses denote PUs in theprimary network and SUs in the secondary net-work, respectively. Furthermore, dash-dottedlines denote the link to transmit the controldata, including signaling and spectrum hole vec-tor information. The solid and dashed linesdenote the transmissions of the users’ self dataand the partners’ data for cooperative retrans-mission, respectively.

COLLABORATIVE SPECTRUM SENSING MODELIn the process of spectrum sensing, the fadingeffect, noise, and interference may cause someerrors (result in false alarm detection and falsedismissal detection), which are very harmfulbecause they may introduce serious interferenceto PUs, resulting in a decrease of the spectrumutilization ratio. The left model in Fig. 2, whichis one of the proposed collaborative spectrumsensing models between PUs and SUs, meansthat a PU leaves or broadcasts the spectrumhole information to SUs over the air or througha wired channel. Obviously, the PU is the

licensed user, who would not do this. However,if a PU would like to allow SUs access to his/herlicensed spectrum when he/she was not usingthem, he/she should offer the spectrum holeinformation to help SUs. Certainly, this pro-posed collaborative spectrum sensing model issimple and introduces no any interferences toPU because spectrum hole information is veryaccurate in this case. Nevertheless, the disadvan-tage of this model is that it needs an additionalspecialized channel to transmit the spectrumhole information from PU to SUs, as well as thepermission and authorization of the PU. Theproposed collaborative spectrum sensing modelamong SUs is shown in the right part of Fig. 2.Collaborative means that SUs can exchange orcooperatively retransmit their SHV informationwith each other. For some reasons, such as deepfading or distance from PUs, some SUs (e.g.,SUt2) may not detect the used spectrum success-fully, which will cause a false dismissal detection.However, their neighbors (e.g., SUt1) can dothem a favor. Therefore, the detection probabili-ty of SU t2 can be improved by cooperativelycommunicating with its adjacent SUs (e.g., SUt1)or exchanging SHV information with each other.SHV exchange and cooperative communicationwere first proposed by Chasemi [7] and Ganesan[8], respectively. However, only AF was consid-ered in [8]. Obviously, there are still many keytechniques of the proposed collaborative spec-trum sensing model that need to be further stud-ied, such as the spectrum sensing algorithm, theinfluence of the detection probability and loca-tions of SUs, the choice of cooperative partner,and so forth.

COOPERATIVE COMMUNICATION COGNITIVE MODEL

In this proposed model, we assume that theSHV has been detected successfully by SUs. Thetop part of Fig. 3 shows the cooperative commu-nication between PUs and SUs. PUs (e.g., PUtand PUr) are source/receiver, and SUs (e.g.,SUt1 or SUt2) act as relays to retransmit PUs’data. The rationale of this choice is that helpingthe PU to increase its throughput entails (for afixed demand of rate by the PU) a diminishedtransmission time for the PU, which leads tomore transmission opportunities for the SUs.That is to say, the PUs’ data can be transmitted

Figure 2. Proposed collaborative spectrum sensing model.

Between PUs and SUs

PUt

SUt1

SUt2

SUr

PUr

Among SUs

PUt

SUt1

SUt2

SUr

PUr

The disadvantage ofFMT is its

complexity due tothe filter bank and

equalization per subchannel.

However, it may not be a serious

issue in the futuredue to the fastdevelopment of

digital signal processing

techniques.



quickly, and some seconds are then saved forSUs to access the idle spectrum. The bottompart of Fig. 3 shows the cooperative cognitivemodel among SUs, where the SUs act as cooper-ative partners with each other. In such a case,SUs can work in either half-duplex (left bottom)or full-duplex (right bottom) mode. When it is inhalf-duplex mode, time-division multiple access(TDMA) is often adopted by SUs, and SUs asrelays (e.g., SUt2 in the left bottom of Fig. 3) donot transmit data themselves. For full-duplexmode, CDMA is always selected by SUs, andSUs act as both source and relay (e.g., SUt1 andSUt2 in the right bottom of Fig. 3). Similar to theproposed collaborative spectrum sensing, thereare also many key techniques that need to bestudied, such as the transmitter scheme of coop-erative communication, capacity and diversitygain analysis, power allocation algorithm, receiv-er designing, partner selection algorithm, and soon.

HYBRID MODELAs aforementioned, the proposed model ofcollaborative spectrum sensing is helpful forspectrum detection, while the proposed modelof cooperative cognitive communication ismainly for data communication. However,spectrum detection and data communicationwill be as a whole in future practical CR sys-tems. Therefore, the combinat ion of theaforementioned two models (i .e., a hybridmodel) is proposed in this subsection, whichis shown in Fig. 4. Obviously, in the proposedhybrid model, there are two steps needed:

collaborative spectrum sensing and coopera-tive communication. Certainly, they can alsobe implemented simultaneously. Both collab-orative spectrum sensing and cooperativecommunication are between PUs and SUs onthe left of Fig. 4, while they are both amongSUs on the right . Apparently, the same col-laborative partners are not needed during thedifferent steps. For example, collaborativespectrum sensing occurs between PUs andSUs during the first step, while cooperativecommunication is among SUs during the sec-ond step. Therefore, the probability of spec-trum detect ion is very important . In thefuture, further research is needed on thehybrid model, including spectrum sensing, theinfluence of probability of spectrum detec-tion, the selection of cooperative users, thesynchronization and channel estimation algo-rithm, and so forth.

Note that we need to point out that the MCMtechnology can be used in any of the three pro-posed models. Moreover, the space diversitygain can be obtained by adopting cooperativecommunication technology. Therefore, the intro-duction of cooperative diversity expands CRresearch from two dimensions (time-frequency)to three dimensions (space-time-frequency). Insummary, all resources can be fully and effec-tively utilized to improve the performance ofMCM-CR systems.

Finally, it is necessary to point out that thePU’s model to occupy the channel is no limita-tion in this article. This is because we mainlyfocus on the model description and performance

Figure 3. Proposed cooperative communication cognitive model.

Between PUs and SUs

PUt

SUt1

SUt2

SUr

PUr

Half-duplex

Among SUs

PUt

SUt1

SUt2

SUr

PUr

Full-duplex

PUt

SUt1 SUr

SUt2 SUr2

PUr

In the proposedhybrid model, whichmay be close to thesystems easily tocarry out, there aretwo steps needed:collaborative spectrum sensingand cooperative communication. Certainly, they canalso be implementedsimultaneously.



analysis, which results in the detection probabili-ty of the spectrum being the chief factor consid-ered in the three proposed models, regardless ofthe PU’s model to occupy the channel.

SIMULATION RESULTS

To get insight into the effectiveness of the pro-posed models and validate some related analyti-cal results, extensive computer simulations havebeen conducted in which M-phase shift keying(MPSK) modulation is used, and channel coeffi-cients hi,j(i ∈ {s, r}, j ∈ {r, d}) are independentsamples of zero mean complex Gaussian randomvariables with variance σ2

i,j, where s, r, and ddenote the source, relay, and destination, respec-tively. Moreover, we assume σ2

s,d = 1, σ2s,r = 1,

σ2r,d = 10, and the total transmitted power P =

Ps + Pr, where Ps and Pr represent the powerallocated to source and relay, respectively.

PERFORMANCE LOSS WITH PROBABILITY OFSPECTRUM DETECTION

Here, we adopt the half-duplex model seen inthe right part of Fig. 4, in which SHV informa-tion is exchanged for collaborative spectrumsensing and cooperative communication amongSUs. We assume that there is only one commu-nication link between PUt and PUr, and severallinks among SUs. Obviously, the collaborativeprobability of spectrum detection Pd = 1 – (1 –Pds)n, where Pds (we have assumed that it is thesame for all SUs) is the probability of spectrumdetection by each SU without any help, and nis the number of SUs selected to join in SHVexchange for collaborative spectrum sensing.Then we further assume that only one SU (e.g.,SUt2) is randomly selected to be a relay part-ner for cooperative communication after col-laborative spectrum sensing. Based on theseassumptions, the unconditional symbol errorrate (SER) performance in an AF cooperationcommunication system is illustrated in Fig. 5,in which Pds = 0.9, and equal power allocation(EPA) is used. In Fig. 5, for comparison, thecurve with perfect spectrum sensing (Pd = 1) isalso plotted. Obviously, it can be concludedfrom Fig. 5 that there is about 4.0 dB loss with-out cooperation (n = 1 and Pd = Pds = 0.9)

when SER is 10–3, and the SER performance isvery close to the results with perfect sensinginformation when the number of collaborativeSUs equals 3 (n = 3 and Pd = 0.99). That is tosay, there is about 4.0 dB gain when two moreSUs join for collaborative spectrum sensing.Obviously, the complexity of the systemincreases because of the cooperative partners’joining.

POWER CONTROL SCHEME OF SUS UNDER THECONSTRAINT OF INTERFERENCE TEMPERATURE

In this subsection, under the constraint of theinterference temperature, we discuss a CR sys-tem based on the cooperative cognitive half-duplex model on the bottom left of Fig. 3 inwhich SUt1, SUt2, and SUr are source, relay, anddestination, respectively, and PUt is the PU.Without loss of generality, Q denotes the inter-ference temperature level. Hence, the problembecomes how to obtain the optimal performance

Figure 4. Proposed hybrid model.

Cooperative spectrum sensing andcommunication between PUs and SUs

PUt

SUt1

SUt2

SUr

PUr

Cooperative spectrum sensing andcommunication among SUs in half-duplex

PUt

SUt1

SUt2

SUr

PUr

Figure 5. SER performance using the half-duplex model in Fig. 4 with Pds =0.9.

P/N0/dB

5 0

10-3

10-6

SER

10-4

10-5

10-2

10-1

100

10 15 20 25 30

n = 1n = 2n = 3Perfect sensing


of the SUs’ network under constraint Q. It iswell known that power control is one simplesolutions. In order to obtain the maximal ergodiccapacity, the approximate optimal power alloca-tion scheme (OPAE) is proposed in the conduct-ed simulations, and an AF cooperativecommunication protocol is selected among SUs.Furthermore, cooperative communication isobviously not adopted when the source-to-desti-nation channel is better than the relay-to-desti-nation one. Figure 6 illustrates the ergodiccapacity with the changing of Q and the channelvariances of the source-to-PU (σ2

s,p), respectively,where σ2

s,p = 0.1 in the upper part and Q = –1dB in the lower part. Obviously, it can be seenfrom Fig. 6 that the capacity performance of theproposed OPAE scheme outperforms that of theexisting EPA, and the capacity performancemonotonically increases with increasing Q andmonotonically decreases with increasing σ2

s,p.

CONCLUSIONS

In this article, we have first studied the MCMtechniques in a multihop MCM-CR system.Considering mutual interference elimination,synchronization, and transmission efficiency, weconclude that FMT is better than OFDM in anMCM-CR system. Second, we propose threecooperative diversity cognitive models: the col-laborative spectrum sensing model, the coopera-tive communication cognitive model, and thehybrid model. The introduction of cooperativediversity expands CR research from time-fre-quency dimensions to space-time-frequencydimensions. Therefore, all resources can be fullyand effectively utilized to improve the perfor-mance of an MCM-CR system. Finally, somesimulations have been conducted to verify thatthe collaborative spectrum sensing can improvethe probability of spectrum detection, resultingin enhancement of the performance of theMCM-CR system.

ACKNOWLEDGMENTS

The work presented in this article was supportedin part by the National Science Foundation ofChina with Grants 60872049, 60971082,60972073, 60872008, 60702039, and 60972031;National Key Basic Research Program of Chinawith Grant 2009CB320407; the National HighTechnology Development 863 Program of Chinaunder Grant 2009AA011803; the Program forNew Century Excellent Talents in the Universityof China under Grant NCET-08-0217; theResearch Fund for the Doctoral Program ofHigher Education of the Ministry of Educationof China with Grant 200804871142; NationalGreat Science Specific Project with Grants2009ZX03003-001, 2009ZX03003-011; and theSEU SKL project with Grant W200907.

REFERENCES[1] S. Haykin, “Cognitive Radio: Brain-empowered Wireless

Communications,” IEEE JSAC, vol. 23, no. 2. Feb. 2003,pp. 201–20.

[2] I. F. Akyildiz et al. “Next Generation Dynamic SpectrumAccess Cognitive Radio Wireless Networks: A Survey,”Comp. Net., Sept. 2006, pp. 2127–59.

[3] T. Luo et al., “A Subcarriers Allocation Scheme forMulti-carrier Modulation Based Cognitive Radio Sys-tems,” IEEE Trans. Wireless Commun., vol. 7, no. 9,Sept. 2008, pp. 3335–40.

[4] I. Berenguer, Filtered Multitone (FMT) Modulation forBroadband Fixed Wireless Systems, Master’s thesis,Univ. of Cambridge, 2002.

[5] V. Chakravarthy et al., “Novel Overlay/Underlay Cogni-tive Radio Waveforms Using SD-SMSE Framework toEnhance Spectrum Efficiency — Part I: TheoreticalFramework and Analysis in AWGN Channel,” IEEETrans. Commun., vol. 57, no. 12, Dec. 2009, pp.3794–3804.

[6] A. Nosratinia et al., “Cooperative Communication inWireless Networks,” IEEE Commun. Mag., vol. 42, no.10, Oct. 2004, pp. 74–80.

[7] A. Ghasemi et al., “Collaborative Spectrum Sensing forOpportunistic Access in Fading Environments,” Proc.IEEE DySPAN ‘05, 2005.

[8] G. Ganesan and Y. Le, “Cooperative Spectrum Sensingin Cognitive Radio, Part I: Two User and Part II: Mul-tiuser Networks,” IEEE Trans. Wireless Commun., vol. 6,no. 6, June 2007, pp. 2204–22.

[9] O. Simeone and U. Spagnolini, “Cooperation and Cog-nitive Radio,” IEEE ICC ‘07, 2007, pp. 6511–15.

[10] Y. Shi and Y. T. Hou, “A Distributed OptimizationAlgorithm for Multi-Hop Cognitive Radio Networks,”IEEE INFOCOM, 2008, pp. 1292–1300.

BIOGRAPHIESTAO LUO [M‘09] ([email protected]) is now working as aprofessor at Beijing University of Posts and Telecommunica-tions, and Key Laboratory of Universal Wireless Communi-cations, Ministry of Education, P.R. China. His recentresearch interests are in the areas of wireless communica-tion theory and technologies including OFDM, MIMO,WAVE, cooperative communication, and cognitive radiotechnologies.

FEI LIN ([email protected]) is currently is currently an asso-ciate professor at Shandong Institute of Light Industry,Jinan, China. Her current research interests are in the areasof wireless communications, especially cooperative commu-nications, cognitive wireless access, and MIMO.

TAO JIANG [M‘06, SM’10] ([email protected]) is currently afull professor at the Wuhan National Laboratory for Opto-electronics, Department of Electronics and InformationEngineering, Huazhong University of Science and Technolo-gy, Wuhan, P.R. China. He has authored or co-authoredover 70 technical papers in major journals/conferences andfive books/chapters in the areas of communications. Hiscurrent research interests include the areas of wirelesscommunications and corresponding signal processing,especially for cognitive wireless access, vehicular technolo-gy, OFDM, UWB and MIMO, cooperative networks, nano


Figure 6. Ergodic capacity with different Q (top) and σ2s,p (bottom), respectively.

Q/dB-1.5 -2

3

2 Ergo

dic

capa

city

(b/

s/H

z)

4

5

-1 -0.5 0 0.5 1 1.5 2

σ2s,p

0

1

0

Ergo

dic

capa

city

2

3

4

0.2 0.4 0.6 0.8 1

EPA OPAE



networks, and wireless sensor networks. He has served oris serving as a member of the symposium Technical Pro-gram Committees of many major IEEE conferences, includ-ing INFOCOM and GLOBECOM. He was invited to serve asTPC Symposium Chair for the International Wireless Com-munications and Mobile Computing Conference 2010. Heis serving as Associate Editor of some technical journals incommunications, including EEE Communications Surveys &Tutorials and Wiley’s Wireless Communications and MobileComputing He was a recipient of Best Paper Awards atIEEE CHINACOM ’09 and WCSP ’09.

MOHSEN GUIZANI [S’83, M’90, SM’98, F‘09] ([email protected]) is currently a professor and the associate dean ofacademic affairs at Kuwait University. He is also an adjunctprofessor at Western Michigan University (WMU). He waschair of the CS Department at WMU from 2003 to2008and chair of the CS Department at the University of WestFlorida from 1999 to 2003. He received his B.S. and M.S.degrees in electrical engineering; M.S. and Ph.D. degrees incomputer engineering in 1984, 1986, 1987, and 1990,respectively, from Syracuse University, New York. Hisresearch interests include wireless communications and

mobile computing, and optical networking. He is Editor-in-Chief of the Wireless Communications and Mobile Comput-ing Journal (Wiley) and the Journal of Computer Systems,Networks, and Communications (Hindawi, Inc.). He is theauthor of seven books and more than 250 publications inrefereed journals and conferences. He has guest edited anumber of special issues in IEEE journals and magazines.He has also served as member, Chair, and General Chair ofa number of conferences. He received both the Best Teach-ing and Excellence in Research Awards from the Universityof Missouri-Columbia in 1999. He is the past Chair of TAOSand current Chair of WTC IEEE Communications SocietyTechnical Committees. He is a member of ASEE and seniormember of the ACM.

WEN CHEN ([email protected]) received his Ph.D. fromthe University of Electro-Communications, Tokyo, Japan, in1999. In 2001 he joined the University of Alberta, Canada.Since 2006 he has been a full professor in the Departmentof Electronic Engineering, Shanghai Jiaotong University,China. His interests cover network coding, cooperativecommunications, cognitive radio, and MIMO-OFDM sys-tems.



Service/contentprovider

2


INTRODUCTIONWireless communications continues to pervadeall aspects of our lives — wireless distribution ofaudio and video around the home, wireless solu-tions for logistics, wireless ticketing and accesscontrol, wireless sensors for agriculture, medicalapplications, and so on. While many peopleappreciate the profound impact that wirelesscommunications are having and will have on ourlives, it will be some time before the vision of wire-less everywhere will be realized — mainly becauseintroducing large-scale changes to the way manysystems work is complex, and requires significanttime, effort, and energy. While there have beenmany important advances in wireless technologyin recent years, there are economic challenges inproviding high-speed wireless access to less pop-ulated areas. This gap between those who bene-fit from digital technology and those who do notis known as digital divide.

A key technology that can help to bridge thedigital divide is satellite communications as it canbe used in areas where there is no terrestrialalternative. In the developed world, satellite net-

works can be interworked with existing terrestrialnetworks, be they wireless or fixed systems, coreor access network, and function as a high-speedbackbone network to support a wide variety ofservices for users in a variety of roles.

In this regard, there are many lessons tolearn from recent mobile satellite experience.Indeed, in urban/suburban areas, fixed andmobile technologies (e.g., asymmetrick digitalsubscriber line [ADSL], Global System forMobile Communications [GSM]) are welladvanced. Satellites, performing in isolation,cannot compete with terrestrial systems in theseurban areas. They can only provide niche ser-vices to areas inaccessible to terrestrial tech-nologies. While these markets are politicallyimportant, they are small and bring poor rev-enue for satellite operators. The future of next-generation satellite systems is clearly in anintegrated architecture with terrestrial systems.Their success also hinges on their ability to pro-vide, in full cooperation with terrestrial sys-tems, broadband data rate applications similarin spirit to today’s Internet. This is also benefi-cial for terrestrial system operators as it willenable them to increase the capacity of theirsystems, support large-scale deployment of dif-ferent emerging bandwidth-intensive services,and satisfy the ever growing community ofInternet users.

Two critical issues arise when consideringsatellite systems in this context. First, satellitesystems are very costly in general; second, thereare challenges in integrating satellite and terrestri-al networks, particularly when terminal mobilityis necessary. This article will give some insighttoward solving both of these problems.

In this article we focus on interworkingbetween the satellite part of the network and itsterrestrial counterpart. Interworking relatedoperations are performed at newly defined enti-ties called interworking gateways (IGWs). Thescope of this article is to define the modules ofthe technological solutions that will be incorpo-

TARIK TALEB, NEC EUROPE LTD.YASSINE HADJADJ-AOUL, UNIVERSITY OF RENNES 1

TOUFIK AHMED, UNIVERSITY OF BORDEAUX 1

ABSTRACTThe current trend in telecommunications ser-

vices provisioning is shifting toward global ubiq-uitous networking and unified servicearchitecture. Given the diversity of access tech-nologies, this global ubiquitous networking can-not be possible without efficient interworkingbetween the different access players. This leadsto the necessity of defining, implementing, anddeploying common services control architecture,able to support a wide variety of services forusers in a variety of roles (consumer, producer,or manager of communication and media). Thisarticle defines some issues related to the inter-working operation between the satellite and ter-restrial domains. It suggests some solutions anddiscusses their potential.

CHALLENGES, OPPORTUNITIES, ANDSOLUTIONS FOR CONVERGED SATELLITE AND

TERRESTRIAL NETWORKS

The authors definesome issues relatedto the interworkingoperation betweenthe satellite and terrestrial domains,and suggest somesolutions and discusstheir potential.

TALEB LAYOUT 2/7/11 10:44 AM


rated in IGWs and evaluate their performancesvia computer simulations.

The remainder of this article is organized asfollows. The next section portrays the key com-ponents of the envisioned architecture. We thendescribe our proposed context-aware completeend-to-end QoS approach devised for interwork-ing between the satellite and terrestrial domains.The article concludes in the final section.

ENVISIONED NETWORK TOPOLOGY

Although geostationary Earth orbit (GEO) sys-tems are widely in use, and low/medium Earthorbit (LEO/MEO) will come onto the scene inthe longer term, this article does not target anyparticular satellite constellation type. The devel-oped solutions will be designed to be applicableto all constellation types (GEO, MEO, andLEO). The satellites are only assumed to bebidirectional interactive, acquiring onboard pro-cessing (OBP) capabilities and intersatellitelinks (ISLs) [1]. Terminals are interactive. Ter-minals outside the reach of the terrestrial net-work have direct access to the satellites.Terminals within reach of the ground Internetinfrastructure have the ability to connect eitherdirectly to satellites or via the IGWs. The over-all objective of this article is to define the neces-sary intelligence that should be added to IGWsto guarantee context-aware complete end-to-end quality of service (QoS) for users. Thus, dif-ferent levels of convergence are considered. Thefirst level of convergence concerns efficient datatransmission based on IP. The second levelrefers to the control and signaling for providingresource allocation and management. The thirdlevel of convergence deals with the provisioningof a generic service delivery platform based inIP Multimedia Subsystem (IMS). Finally, mobil-ity management and seamless connectivity areconsidered for both network-link handover andlink-layer handover.

CONTEXT-AWARE COMPLETEEND-TO-END QOS APPROACH

EFFICIENT DATA TRANSMISSIONFirst, in light of the rapid globalization of theInternet and the resultant universality of IP, thedata traffic load to be generated from the inter-worked satellite/terrestrial networks is expectedto be all-IP as well. Investigating the interactionsof IP protocols with the network is of vitalimportance.

Satellite systems are well known for theirunique characteristics — long propagationdelays, large delay-bandwidth product, errorsdue to propagation corruption and handovers,and variable round-trip time (RTT) and linkhandovers. These features put limitation on theworking of most transmission. With this regard,the authors have recently developed the Recur-sive, Explicit, and Fair Window Adjustment(REFWA) method to enhance the efficiency andfairness of TCP in satellite systems [2]. The useof the REFWA scheme has been extended fur-ther to the case of hybrid wired/wireless net-works as well. While the REFWA schemeexhibits good performance, its performanceremains limited in large bandwidth environments(such as satellite systems) due to its window-based nature. Development of a new rate-basedcongestion control protocol that is specifically tai-lored for satellites and can exploit well the largedelay bandwidth product feature of the satellitesystems is required. REFWA can be a good can-didate for that by changing its window-based fea-ture to a rate-based one. Indeed, this is possibleby having IGWs send data at rates exactly equalto the feedback value computed by REFWA.This is similar in spirit to the concept of theExplicit Control Protocol (XCP). Using suchrate-based congestion control mechanism, andsimilar in spirit to the connection splitting in [3],the full end-to-end path will be decoupled into

Figure 1. RTT-based connection setup + connection decoupling.

Sender Receiver A

Receiver B

IGW1

IGW2

Connection B

Connection A

In light of the rapidglobalization of the

Internet and theresultant universalityof IP, the data trafficload to be generatedfrom the interworked

satellite/terrestrialnetworks is expected

to be all-IP as well.Investigating the

interactions of IP pro-tocols with the

network is of vitalimportance.



separate segments — end terminal to IGW, asegment traversing the satellite network, and thefinal segment between the IGW and the remoteend terminal (Fig. 1). The use of the protocolpertains to the segment traversing the satellites.The other two segments will employ controlmechanisms optimized for their characteristics(wireless or wired). Furthermore, the necessaryintelligence required for coordination betweenthe used data transmission mechanisms will beadded to IGWs to ensure reliable delivery ofdata while meeting the end-to-end QoS require-ments.

To illustrate the idea with more clarity, wehave conducted simulations using NS-2. Thewindow-based nature of REFWA is replaced bya rate-based one as explained earlier. The per-formance of the modified REFWA is comparedagainst that of XCP and TCP as shown in Fig. 2.We tested the system under homogeneous trafficconditions using 100 heavy FTP sources during300 s, a duration long enough to capture andstudy the behavior of our proposed transportprotocol. At the beginning of the simulation, thehosts behind each satellite terminal are activatedrandomly following a uniform distribution rang-ing from 10 to 100 ms.

Figure 2 shows the average window size forthe three protocols. From the figure, it is clearthat the average window size, when REFWA isin use, converges immediately to the optimalwindow value. In contrast, in case of XCP, ittakes 15 s before the system reaches its optimalwindow. With TCP/Reno, the system oscillatesaround the optimal value without reaching asteady state. The simulation results also demon-strate the good performance of the modifiedREFWA as it achieves the highest goodput.Indeed, REFWA outperforms XCP andTCP/Reno by 6.93 and 20.90 percent, respective-ly, in terms of goodput.

For the sake of further transmission efficien-cy and better QoS, short RTT connectionsshould be established via the terrestrial wirelessnetwork. Indeed, terminals communicating withnearby users do not have to drain their energy toconnect directly to satellites. Long RTT connec-tions can be set via satellites. Different tech-

niques can be used for periodic monitoring ofthe network conditions. IGWs will be constantlyupdated with feedback on network dynamics.Based on this feedback and the RTT of connec-tions, IGWs decide the path for communication— either via the satellites or via only the terres-trial network. For this purpose and in order toblur the separation between the satellite and ter-restrial domains, there is need for exchange ofstate information (e.g., instant link loads)between the two domains. A hierarchical archi-tecture of gateways can be considered. With thisregard, the number of levels of this hierarchy,the size of each level, the amount of control traf-fic that should be exchanged and the length ofthe monitoring interval time should be decidedin a way that enhances the accuracy in the assess-ment of network dynamics while minimizing theoverhead in terms of signaling messages. Such acontext-aware routing scheme will yield a betterload balancing over the entire network and willenhance the end-to-end (E2E) QoS [4].

In the considered interworked satellite/terres-trial network, the IGW also provides the inter-face for any service/content provider who desiresto provide service/content over both the terres-trial and satellite networks. If the provided datais bursty in nature (e.g., video data) and the tar-geted population of users is potential, it will behighly useful to send data from the provider tousers using satellite channels. In such a commu-nication scenario (Fig. 3), a service subscriberissues a request for a particular data/video titleto the IGW via the terrestrial network (wirelessor wired). The IGW informs the service providerof the request, and the latter allocates the neces-sary resources to satisfy the user’s request. Toensure reliable transmission of data (dependingon the underlying transport protocol) the clientkeeps acknowledging successful receptions ofdata to the IGW via acknowledgment (ACK)packets sent over the terrestrial network. IF theACK packets are delayed or lost, the overall net-work performance may be impacted. Addingintelligence in terms of an adequate delayedacknowledgment mechanism along with a robusterror recovery mechanism to IGWs can help copewith these issues.

RESOURCE ALLOCATION AND MANAGEMENTIn light of the limited resources of any powerfulnetwork, QoS can be maintained only via effi-cient resource management/allocation mecha-nisms. In the case of converged satellite/terrestrial networks, devising an efficientresource allocation method is a highly challeng-ing task due to the fluctuating nature of thewireless links. Indeed, in DVB-S2 networks, forinstance, a novel satellite-tailored adaptive cod-ing and modulation (ACM) technique is intro-duced to cope with the wireless channelfluctuations. ACM renders the resource reserva-tion process even more difficult since the chan-nel capacity changes frequently as the channelexperiences noisy periods.

While there are many approaches to solvingresource management issues in networks, onethat is suitable when there are limited and costlyresources is connection admission control(CAC).

Figure 2. The average window size for REFWA, TCP, and XCP.

Time (s)

50 0

15 20

0

Ave

rage

win

dow

siz

e (1

00 T

CP

flow

s)

10 5

100 150 200 250 300

50 0

15 20

0

10 5

100 150 200 250 300

50 0

15 20

0

10 5

100 150 200 250 300

TCP/Reno

XCP

REFWA



A large library of CAC schemes has beenproposed in the literature. These techniques canbe classified as either resource-reservation-basedor statistical-multiplexing-based schemes.Resource reservation CAC systems have someknown scalability issues and may often lead toself-induced congestion due to the heavyresource reservation process. Besides, staticreservation falls short of satisfying the flexibilityrequirements of typical network operators. Fur-thermore, statistical multiplexing CAC approach-es cannot completely eliminate congestion duringsome peak noisy periods. However, they enableresource sharing between users and yieldreduced waste of resources.

This feature renders statistical multiplexingschemes more suitable for converged satellite/terrestrial networks. Several research works havedevised different CAC schemes to guarantee areasonable QoS level under different networkconditions [5]. A common shortcoming of theseschemes resides in their inefficiency in dealingwith the varying nature of the physical layercapacity of the satellite network.

The authors’ research work presented in [6]takes into account the satellite channel fluctua-tions and presents an interesting CAC mecha-nism that also ensures fairness among terminalscompeting for the capacity of the same satellitechannel. A shortcoming of the proposedapproach lies in its lack of a bandwidth alloca-tion mechanism and multiservice support.Indeed, CAC should be exerted in conjunctionwith a bandwidth allocation mechanism, espe-cially in converged satellite/terrestrial networkswhere the link capacity may vary as a result ofthe ACM mechanism. In this case, combinedaction among various layers, a cross-layerapproach, of the networks is likely to improvethe performance of the overall system by pro-tecting, for example, prioritized flows from pack-et drops during congestion events. In this area ofresearch, there is a particular interest in thedevelopment of a cross-layer bandwidth allocationmechanism that can assist CAC and furtherenhance its functionality.

For this purpose, we suggest using a channelprediction mechanism based on the least meansquare (LMS) algorithm to tackle the excessivedelay incurred by the feedback in satellite net-works. Based on the proposed model, a self-con-figuring mechanism to cope with variablenetwork conditions is derived. From this cross-layer approach, both optimized bandwidth allo-cation and guaranteed per-class QoS areexpected.

To illustrate the idea, we have conductedsimulations using Opnet. We test the systemunder homogeneous traffic conditions using FTPsources during 1500 s. In the first phase of thesimulation, the system is maintained free fromnoise. At the beginning of the second phase ofthe simulation (i.e., t = 245 s), a source of noiseis introduced, which directly impacts the signal-to-noise ratio (SNR), which decreases by approx-imately 2 dB. At the third phase of thesimulation (i.e., t = 790 s), another source ofnoise is introduced. Finally, at the last phase,which starts at t = 1000 s, the sources of noiseare eliminated successively at t = 1000 s and t =

1210 s. This performance evaluation scenarioallows us to see how quickly and accurately ourscheme adapts to both degrading and improvingconditions.

Figure 4 shows the instantaneous SNR expe-rienced by a satellite terminal vs. the predictedSNR. In the first stage of the simulation, weobserve that the estimation error is relativelyimportant. This is principally due to the randominitialization used in LMS. After 105 s, we clear-ly see that the estimation becomes more precise,which demonstrates the effectiveness of our pro-posed prediction mechanism. At this point, theaccuracy of the proposed algorithm is approxi-mately 98.5 percent.

Figure 5 indicates the performance of theproposed mechanism in terms of throughput.The figure clearly shows that the bandwidthmanager using predicted SNR values, whichallow selecting the appropriate modulation andcode rate, outperforms the conventionalapproach. As a consequence, the experiencedthroughput is increased by approximately 4.6percent.

The proposed cross-layer CAC mechanism,which relies on predicted SNR values, protectsthe network from congestion while maintaining agood trade-off between bandwidth utilizationand end-to-end delay. These performances areparticularly interesting for the provision ofdelay-sensitive and bandwidth-intensive applica-

Figure 3. Service delivery pattern.

Service/contentprovider

IGW

Client

1

3

3

2



tions over converged satellite/terrestrial net-works.

MOBILITY MANAGEMENTThe success of any communication system hingeson its ability to provide acceptable QoS. In thecontext of mobile environments, QoS provision-ing depends in turn on an efficient managementstrategy for mobility. In satellite networks,mobility management is a challenging task.Indeed, supporting continuous communicationover satellite systems may require changing spotbeams (and links in LEO/MEO systems) as wellas the IP address of the communication end-points. Thus, both link-layer and network-layerhandovers are required for satellite networking.In case of non-geostationary (NGEO) systems,mobility management becomes more complex asboth the satellite network and mobile users areon the move.

In satellite networks, handovers can be broad-ly classified into two categories: network-linkhandover and link-layer handover. The formeroccurs when one of the communication end-points changes its IP address due to motion ofsatellites or mobility of the user terminal. Thelatter occurs when one or more links betweenthe end terminals change. It consists of satellitehandover, ISL handover, and spot beam hand-over. Spot beam handover is the most commontype of handover. They occur frequently due tothe small area covered by spot beams and themobility of users (or high speed of NGEO satel-lites). In this article, we initially focus on spotbeam handovers and then extend the study toother handover types.

Spot Beam Handover Management — For better fre-quency utilization, the footprint of an individualsatellite is divided into smaller cells, called spotbeams. To ensure uninterrupted ongoing com-munications, a current communication linkshould be handed off to the next spot beamwhen needed. A spot beam handover involvesthe release of the communication link betweenthe user and the current spot beam and acquir-ing a new link from the next spot beam to con-tinue the communication. Due to the small areacovered by spot beams, users’ mobility, and high

satellite speed in the case of NGEO systems,spot beam handovers are the most common typeof handovers experienced in satellite systems.

Efficient management of handovers is partic-ularly linked to the resource allocation problemdiscussed above. Indeed, the selection of a suit-able policy for channel allocation can ensurechannel availability during handover. Thus,channel allocation strategies and handover guar-antee are the prime issues in managing handoverrequests. It should be noted that it is more desir-able to ensure and guarantee smooth ongoingcalls rather than block a newly arriving call. Tosolve the spot beam handover problem, severalhandover schemes have been proposed in therecent literature. A thorough survey on thesetechniques is available in [7].

Sophisticated network planning is required toassign more capacity to spot beams when a hightraffic rate is expected. Statistical methods, cou-pled with a user behavior model and precise pre-dictions of satellite tracks relative to the Earthsurface, allow general characterization of thetraffic load for a particular satellite or spotbeam. Via a cross-layer design, this can help inanticipating imminent handover events andlocating the new point of attachment to thesatellite network [8]. While a cross-layer opti-mization can be implemented at either enddevices or intermediate nodes in the network,such as IGWs, it is relatively easier and morefeasible to implement changes at mobile nodes.Indeed, at the communication endpoint, thephysical layer of a mobile host instantly mea-sures the radio strength or link quality. Whenthe mobile node moves into the overlappingarea of two or more spot beams, and differentsignals are consequently detected by the physicaland data link layers, a warning message notifyingof an imminent handoff event along with a list ofthe new possible spot beams are sent to theapplication layer. In case of multiple spot beams,the application layer refers to a set of tools tosort out the spot beam to which the mobile nodeis most likely going to be connected. Indeed, theapplication layer may use history on the user’smobility pattern to predict the new spot beam.Referring to a spatial conceptual map, alongwith the user’s personal information, its currentposition, and its velocity heading, the applicationlayer can make an accurate prediction of themost probable future spot beam. Prior contextu-al knowledge of the coverage area of the satel-lite network and the type of application canfurther increase the accuracy of the prediction.Once the next spot beam is determined, themobile host informs the IGW of the next spotbeam. Based on the current conditions (e.g.,maximum number of free channels) of each spotbeam, the IGW decides whether the call shouldbe accepted or denied. If the handover cannotbe made without degrading QoS of already exist-ing users or causing network congestion, theIGW denies the handover request and sends animmediate negative acknowledgment to informthe mobile host that the request has been turneddown. Simultaneously, a list of available spotbeams can be sent along with the negativeacknowledgment to induce the mobile host tohand over with another spot beam. The mecha-

Figure 4. The instantaneous SNR vs. the predicted SNR.

Time (s)

250

2

SIN

R (d

B)

0

4

6

8

10

12

14

16

18

20

22

0 500 750

QPSK (1/4, 1/3, 2/5, 1/2, 3/5, 2/3, 3/4, 4/5, 5/6, 8/9, 9/10)

16 APSK (2/3, 3/4, 4/5, 5/6)

32 APSK (3/4, 4/5, 5/6, 8/9, 9/10)

8 PSK (2/3, 3/4)

1000 1250 1500

Predicted SINRInstantaneous SINR



nisms by which IGWs admit or turn down hand-over requests (from a user) should be consistentwith the underlying resource allocation strategy.An actual design and implementation of thiscontext-aware cross-layer architecture at themobile terminals and the intelligence requiredby IGWs to manage handovers define an inter-esting topic of research in this particular field ofresearch.

Satellite and ISL Handover Management in NGEO Sys-tems — The solution suggested above may alsobe highly interesting when it is put in the contextof mobile satellite communication systems (e.g.,LEO, MEO). In these systems, satellite hand-overs are more important as users need to firstchoose among different satellites and will thenbe served by the spot beam covering the user. Inaddition to the above solution, there is a need todevelop complementary solutions that select themost suitable satellite for communications thatcan reduce bandwidth waste and call blockingprobability, and also fulfill the QoS require-ments. In regard to QoS, the application typeshould be taken into account in the satellitehandover management strategy.

Another important issue in the context ofconverged terrestrial/satellite networks is how tomanage network layer handovers. While thereare IP-based mobility solutions which havealready reached the marketplace, they areunsuitable for converged satellite/terrestrial net-works. This is mainly because they result in avery large amount of signaling traffic whenemployed in a satellite context, due to the con-stant and rapid motion of the satellites. Conse-quently, alternative network-layer handovermechanisms are required.

SEAMLESS CONNECTIVITYWhile in the above subsections the focus was ondefining issues pertaining to congestion control,resource allocation, and mobility management inthe converged satellite/terrestrial network, anddevising possible solutions, in the remainder ofthis article we discuss possible scenarios forseamless use of the interworked satellite/terres-trial network and provide guidelines for theirrealization.

IMS-Based Service Delivery Architecture — A possiblesolution for service integration between satelliteand terrestrial networks can be based on IMS,which represents a key element in the satellitearchitecture, supporting seamless and universalaccess to personalized services. Indeed, theadoption of IMS will favor the rapid emergenceof new secure services and will enable seamlessprovisioning of multimedia services. IMS pro-vides a service delivery platform (SDP) on top ofconvergence network technologies. This will helpin generating new revenues, reducing the com-plexity of the IWG while significantly decreasingthe cost of the satellite network management,which impacts directly the satellite services cost.Additionally, IMS allows more efficient handlingof the multicast and broadcast traffic initiatedfrom the satellite terminals. The rest of this sec-tion describes how to handle multicast servicesusing IMS-based architecture.

Current mechanisms for delivering multicastservices over satellite links use snooping (layer2) or proxying (layer 3) to allow the delivery ofIGMP/MLD membership messages to a satellitegateway over the air interface. A proxy andsnooper do not change anything in the IGMP/MLD messages but only forward the request fur-ther toward theIGW. In fact, over the satellitenetwork, the broadcast property exists only onthe forwarding link (i.e., the satellite return linkprovides only directional links). The host cannotlisten to the signaling reports transmitted byother hosts on the return link. This leads themto individually send out their report, which gen-erates excessive traffic. This would result inflooding and high latency problems.

Flooding occurs when many hosts (i.e.,IGMP/MLD clients) reply to a broadcast requestfrom the IGMP/MLD Querier sent out by therouter to sense the presence of clients in eachmulticast group. As highlighted earlier, unlikeLAN, the satellite return link does not provide abroadcast property but only a unidirectional con-nection. Typically, hosts cannot listen directly toreplies from other hosts. Thus, all the hosts haveto respond to the IGMP/MLD Querier after theexpiration of their timer. Moreover, satellitemulticast groups can be very large and verydynamic. This leads to a waste of bandwidth andCPU over the satellite link and the Gatewayrespectively.

Another important issue of IP multicastbehavior over satellite networks is the latency instopping transmission after the last host leaves amulticast group. The latency is the delay neededfor the Querier to become aware that the multi-cast group is empty in order to stop multicastforwarding on it. The authors explain in [9] howto tackle the flooding and latency issues in pro-viding multicast services. In this context, the IMSservice delivery platform is used, which allowsthe IGMP messages to be aggregated and trans-mitted as a Session Initiation Protocol (SIP)-based message for managing multicast groups.The argument for using SIP/IMS protocols is toallow hosts to join and leave a multicast group asIGMP does, verify the authentication and autho-rization of the user, signal any cryptographiccontext (e.g., using MIKEY), and easily supportany future extension/augmentation that can beimplemented in SIP.

Figure 5. Throughput using instantaneous vs. predicted SNR.

Time (s) 250

Thro

ughp

ut (

b/s)

0.0E+00

5.0E+05

1.0E+06

1.5E+06

2.0E+06

2.5E+06

3.0E+06

3.5E+06

4.0E+06

0 500 750 1000 1250 1500

Bandwidth manager using instant SNRBandwidth manager using predicted SNR



CONCLUDING REMARKSThis article highlights some of the opportunitiesbehind integrated satellite and terrestrial net-works. For the realization of such convergednetworks, it addresses a number of issues per-taining to transmission efficiency, resource allo-cation and management, mobility management,and seamless connectivity. While the mainobjective of this article is to highlight the relat-ed issues and define new directions for the com-munity of satellite researchers, it also suggests anumber of solutions, as seen from the network-ing perspectives, that, once they are put togeth-er, form a complete context-aware end-to-endQoS approach that solves the aforementionedissues.

REFERENCES[1] J. Farserotu and R. Prasad, “A Survey of Future Broad-

band Multimedia Satellite Systems, Issues and Trends,”IEEE Commun. Mag., vol. 38, no. 6, 2000.

[2] T. Taleb, N. Kato, and Y. Nemoto, “REFWA: An Efficientand Fair Congestion Control Scheme for LEO SatelliteNetworks,” IEEE/ACM Trans. Net. J., vol. 14, no. 5, Oct.2006. pp. 1031–44.

[3] M. Marchese, M. Rossi, and G. Morabito, “PETRA: Per-formance Enhancing Transport Architecture for SatelliteCommunications,” IEEE JSAC, vol. 22, no. 2, Feb. 2004,pp. 320–22.

[4] T. Taleb et al., “Explicit Load Balancing Technique forNGEO Satellite IP Networks with On-Board ProcessingCapability,” to appear, IEEE/ACM Trans. Net. J.

[5] R. Abi Fadel and S. Tomhe, “Connection AdmissionControl and Comparison of Two DifferentiatedResources Allocations Schemes in a Low Earth OrbitSatellite Constellation” ACM Wireless Net. J., vol. 10,no. 10, May 2004.

[6] Y. H. Aoul and T. Taleb, “An Adaptive Fuzzy-Based CACScheme for Uplink and Downlink Congestion Control inConverged IP and DVB-S2 Networks,” IEEE Trans. Wire-less Commun., to appear.

[7] P. K. Chowdhury, M. Atiquzzaman, and W. Ivancic,“Handover Schemes in Satellite Networks: State-of-the-Art and Future Research Directions,” IEEE Commun.Surveys & Tutorials, vol. 8, no. 4, 4th qtr. 2006.

[8] T. Taleb et al., “A Cross-Layer Approach for an EfficientDelivery of TCP/RTP-Based Multimedia Applications inHeterogeneous Wireless Networks,” IEEE Trans. Vehic.Tech., vol. 57, no. 6, Nov. 2008, pp. 3801–14.

[9] T. Ahmed et al. “IMS-based IP Multicast Service Deliveryover Satellite Network” 8th IEEE PIMRC ‘07, Sept. 2007.

BIOGRAPHYTARIK TALEB [S‘04, M‘05, SM‘10] ([email protected])is currently working as a senior researcher at NEC EuropeLtd, Heidelberg, Germany. Prior to his current position untilMarch 2009, he worked as am assistant professor at theGraduate School of Information Sciences, Tohoku Universi-ty, Japan. From October 2005 to March 2006 he worked asa research fellow with the Intelligent Cosmos ResearchInstitute, Sendai, Japan. He received his B.E. degree ininformation engineering with distinction, and M.Sc. andPh.D. degrees in information sciences from GSIS, TohokuUniversity, in 2001, 2003, and 2005, respectively.

YASSINE HADJADJ ([email protected]) is an associate profes-sor at the University of Rennes 1, France, where he is alsoa member of the IRISA Laboratory. He received a B.Sc. incomputer engineering with high honors from SENIA Uni-versity, Oran, Algeria, in 1999. He received his Master’sand Ph.D. degrees in computer science from the Universityof Versailles, France, in 2002 and 2007, respectively. Hewas an assistant professor at the University of Versaillesfrom 2005 to 2007, where he was involved in severalnational and European projects such as NMS, IST-ATHENA,and IST-IMOSAN. He was also a post-doctoral fellow at theUniversity of Lille 1 and a research fellow, under the EUFP6EIF Marie Curie Action, at the National University of Dublin,where he was involved in the DOM’COM and IST-CARMENprojects, which aim at developing mixed Wi-Fi/WiMAXwireless mesh networks to support carrier grade services.His main research interests concern the fields of wirelessnetworking, multimedia streaming, congestion control andQoS provisioning, and satellite communications. His workon multimedia and wireless communications has led tomore than 25 technical papers in journals and internationalconference proceedings.

TOUFIK AHMED ([email protected]) is a professor at Institut Poly-technique de Bordeaux (IPB) in the ENSEIRB-MATMECASchool of Engineering. He is doing his research activities inCNRS LaBRI Lab, UMR 5800 at the University of Bordeaux 1.He received a B.Sc. in computer engineering with high hon-ors from the National Institute of Computer Science,Algiers, Algeria, in 1999, and M.Sc. and Ph.D degrees incomputer science from the University of Versailles, France,in 2000 and 2003, respectively. In November 2008 heobtained his Habilitation à Diriger des Recherches degreefrom the University of Bordeaux 1 on adaptive streamingand control of video QoS over wired/wireless IP networksand P2P architectures. He was a visiting scientist at theSchool of Computer Science of the University of Waterloo in2002 and a research fellow at PRiSM laboratory of the Uni-versity of Versailles until 2004. His main research activitiesconcern QoS for multimedia wired and wireless networks,end-to-end signaling protocols, P2P networks, and wirelesssensor networks. His work on QoS and video delivery hasled to many publications in major journals and conferences.

A possible solutionfor service integra-tion between satel-lite and terrestrialnetwork can bebased on the IP Multimedia Subsystem that represents a key element in the satellite architecture,supporting seamlessand universal accessto personalised services.



Powervaluelevel

After RS groupin

RS9

R

RS4-75dBm-90dBm-90dBm-60dBm-97dBm-90dBm-93dBm-87dBm

ector 1

4 RS5

RS1RS2Sector 2

BS1

RS3 RS6

RS7

Sect

or 3

BS2


INTRODUCTIONThe specification of the third-generation (3G)Long-Term Evolution (LTE) radio interface wasrecently finished by the Third Generation Part-nership Project (3GPP) [1]. It aims to providedownlink/uplink peak rates of at least 100Mb/s/50 Mb/s and round-trip times of less than10 ms. The first commercial LTE deploymenttook place in Stockholm, Sweden, and Oslo,Norway, in December 2009. However, the LTEsystem cannot meet the requirements of futurebroadband wireless networks, which is officiallycalled International Mobile Telecommunications(IMT)-Advanced by the International Telecom-munications Union — RadiocommunicationStandardization Sector (ITU-R). The IMT-Advanced system is expected to supportenhanced peak data rates on the order of 100Mb/s for high-mobility and 1 Gb/s for low-mobil-ity environments, respectively [2]. Also, it is ableto provide a high degree of commonality offunctionality worldwide while retaining the flexi-bility to support a wide range of services andapplications in a cost-efficient manner. In order

to meet these new challenges, 3GPP has startedto develop further advancements for 3G LTEsystems, referred to as LTE-Advanced, as a can-didate for IMT-Advanced.

In order to meet the requirements of IMT-Advanced, more spectrum bands are needed.Besides the existing spectrum for 3G mobilecommunication systems, spectrum bands locatedat 450–470 MHz, 698–790 MHz, 2.3–2.4 GHz,and 3.4–3.6 GHz have also been identified for3G and IMT-Advanced systems by the ITU dur-ing World Radio Conference 2007 (WRC ’07)[3]. Most of them are above the 2 GHz band,where the radio propagation is more vulnerableto non-favorable channel conditions. With tradi-tional cellular architectures, the density of basestations (BSs) has to be significantly increased tomeet service coverage requirements, offeringhigh data rates at these high-frequency bands.Obviously, this is not a favorable method since itwould greatly increase deployment costs. Instead,a cost-effective solution would be the multihopcellular architecture with relaying, which short-ens the transmission distance and increases theamount of users under good channel conditions,thus allowing for higher throughput. Recently,standardization efforts of integrating cooperativerelaying technologies into LTE-Advanced net-works have commenced [4].

In LTE and LTE-Advanced networks, orthog-onal frequency-division multiplexing (OFDM)has been chosen as the multiple access methodsince it can provide high data rates and spectrumefficiency. In OFDM-based systems, users aremultiplexed in time and frequency by means of ascheduler that dynamically assigns subcarriers todifferent users at different time instances accord-ing to predefined scheduling metrics. Therefore,the OFDM-based multihop transmission bymeans of relay stations (RSs) has been recog-nized as an efficient technique to meet therequirements of future broadband wireless net-works. The RSs have the capability of forward-ing the traffic between the base station (BS) andthe mobile stations (MSs). The main objective ofintroducing multihop relaying technology into

KAN ZHENG, BIN FAN, JIANHUA LIU, YICHENG LIN, AND WENBO WANG,BEIJING UNIVERSITY OF POSTS & TELECOMMUNICATIONS

ABSTRACT

Recently there has been an upsurge of inter-ests in the multihop infrastructures for orthogo-nal frequency division multiplexing-based cellularnetworks in both academia and industry. In thisarticle, we first present an overview of the inter-ference coordination strategies in OFDM-basedmultihop cellular networks. Then, based on theframework of third-generation LTE-Advancednetworks with multihop relaying, several typicalstatic or semi-static interference coordinationschemes are proposed to improve the coverageand increase the cell edge data rate. By applyingthese schemes, the radio resources can be reusedwith certain limitations on either the frequencyor time domain, or even both of them. Dynamicsystem-level simulations are also carried out todemonstrate the effectiveness of the proposedinterference coordination schemes.

INTERFERENCE COORDINATION FOR OFDM-BASEDMULTIHOP LTE-ADVANCED NETWORKS

The authors presentan overview of theinterference coordina-tion strategies inOFDM-based multihop cellular networks. They pro-pose several typicalstatic or semi-staticinterference coordina-tion schemes toimprove coverageand increase the celledge data rate.

ZHENG LAYOUT 2/7/11 10:46 AM Page 54


OFDM-based cellular networks is to achieveboth higher throughput and better service cover-age with the assistance of cost-effective relayarchitecture.

Preliminary study on cooperative relayingtechnology indicates that multihop relayingoffers certain performance advantages such ascoverage extension and capacity improvement[5]. However, numerous challenges at the physi-cal (PHY) and medium access control (MAC)layers in OFDM-based multihop cellular net-works still remain [6]. For example, time andfrequency resources are typically reused in multi-ple cells, thus leading to co-channel interferenceimpairments among the coverage of neighboringBSs and/or RSs. Such co-channel interferencebetween cells plays an important role in affectingthe performance of OFDM-based multihop cel-lular networks. The impact of interference ismore obvious for cell edge users, who are moresusceptible due to poor channel gains with theirserving BS or RS. Limited reception caused byinterference at the cell edge is an issue of greatimportance for wireless operators who want toprovide full coverage within their service areasand guarantee a prior agreed quality of service(QoS) to their subscribers. To the best of theauthors’ knowledge, there have been few worksin the literature on specific interference coordi-nation schemes in OFDM-based multihop cellu-lar networks.

The scope of this article is hence to examinehow users can share the available radio resourcesefficiently, in terms of bandwidth and time allo-cation, in order to mitigate intercell co-channelinterference and thus enhance user throughput,especially for cell edge users. We first brieflyintroduce the state-of-the-art interference miti-gation schemes in OFDM-based cellular net-works. Then, with the introduction of the systemframework based on 3G LTE specifications, sev-eral static or semi-static interference coordina-tion schemes are proposed. We also analyze anddiscuss their performance extensively throughsimulations.

STATE-OF-THE-ART

In OFDM-based cellular networks, interferencecoordination strategies have been studied inorder to increase achievable reuse of the scarcespectrum with reasonable complexity and over-head. An exhaustive exposure of the state of theart is outside the scope of this article. Interfer-ence coordination aims at applying restrictions toradio resource management in a coordinated wayamong cells. These restrictions can be either onthe available radio resources or in the form ofrestrictions on the transmit power that can beapplied to certain radio resources. Such restric-tions provide the possibility for improvement insignal-to-interference-plus-noise ratio (SINR),and consequently to the cell edge throughput andcoverage. Interference coordination also requirescommunication between different nodes in thenetwork to (re)configure resource restrictions.Based on the requirement of the inter-BS com-munication interval, most of the existing and cur-rently in development interference coordinationstrategies can be categorized into three types [5]:

Static coordination: Internode communica-tion is very limited since it corresponds to thesetup of restrictions only. Reconfiguration of theresource allocation restrictions among nodes isdone on a timescale of days.

Semi-static coordination: The correspondingsignaling rate of internode communication isgenerally on the order of tens of seconds to min-utes. Reconfiguration of the restrictions is doneon a timescale of seconds or longer.

Dynamic interference coordination: Itrequires much internode communication toexploit multiuser diversity among neighboringcells with high computational complexity. In thiscase, internode signaling or data transferringmay be needed at each scheduling instant.

In the LTE standardization process, manyappealing and feasible interference coordinationalgorithms were extensively studied for OFDM-based networks. Typical interference coordina-tion strategies, such as soft frequency reuse(SFR) and fractional frequency reuse (FFR),utilize the resources of frequency and radiatedpower to coordinate BS transmissions with pre-defined resource constraints for different typesof users as follows:

Soft frequency reuse: The whole availablebandwidth is divided into multiple non-overlap-ping subbands. Each cell selects one subband asits major band and the others as its minor bands.Major bands can be used in the whole cell areawith full transmit power while minor bands areonly in the inner zone of the cell with reducedtransmit power. The performance of the celledge user (CEU) can be improved by using themajor bands to mitigate intercell interference.Also, the high data rate of the cell center user(CCU) can be achieved since both major andminor bands are available for its transmission [7].

Fractional frequency reuse: It splits the givenbandwidth into inner and outer subbands. Theinner subbands are completely reused by allcells, while the outer subbands are dividedamong neighboring cells with a frequency reusefactor greater than one. Intercell interference isreduced at the cell edge by assigning resourceson outer subbands to CEUs [8].

With the development of LTE-Advanced,multihop relaying techniques have been intro-duced into the cellular network. In a multihopcellular network, an RS usually transmits thesame or a different format of the information asthat received from the BS or MS, which is likelyto be regarded as a certain kind of repetition. Sothe capacity of this relaying network is decreasedfrom the system point of view. Therefore, it isquite necessary to design an efficient resourceallocation scheme in a multihop cellular net-work, which turns to high radio resource reuseamong RSs and BSs. However, compared withtraditional cellular networks, more complicatedand serious interference exists in an OFDM-based multihop cellular network. The downlinkco-channel interference in such a network can beclassified as:

Intercell interference: The co-channel inter-ference introduced by the frequency reusebetween multiple cells; that is, interference fromBS → RS links, BS → MS links, and RS → MSlinks in the neighboring cells, respectively.

Interference coordi-nation aims at apply-ing restrictions to the

radio resource man-agement in a coordi-

nated way amongcells. These restric-tions can be either

on the availableradio resources or in

the form of restric-tions on the transmit

power that can beapplied to certain

radio resources.



Intracell interference: The co-channel inter-ference induced by resource reuse within thesame cell; that is, interference from BS → MSand the RS → MS links in the same cell, respec-tively.

Similar to a traditional OFDM cellular net-work, there are also three types of interferencecoordination strategies to deal with co-channelinterference in OFDM-based multihop cellularnetworks: static, semi-static, and dynamicschemes. With static/semi-static interferencecoordination schemes, both inter- and intracellinterference can be diminished by allocatingresources under the constraint of the predefinedresource coordination pattern. Also, it is criticalto exploit the characteristics of multihop trans-mission to obtain a well-designed coordinationpattern that can achieve throughput gain for allusers in the cell. For dynamic coordinationschemes, the power and resource allocation isdynamically coordinated among neighboringcells at each transmission time [5]. Optimalintercell resource coordination and large mul-tiuser diversity gain can be obtained by central-ized control or non-cooperative competitiongaming among the cells. However, due to itshuge signaling overhead and high complexity,such a dynamic scheme is not practical in cur-rent OFDM-based multihop cellular networks.

As of today, most available interference coor-dination schemes for OFDM-based multihopcellular networks are only extensions of existingcoordination strategies used in traditional cellu-lar networks. The particularities and opportuni-ties of multihop relaying transmissions in dealingwith co-channel interference problems, however,have not been fully exploited yet.

FRAMEWORK OF AMULTIHOP RELAYING NETWORK

In this section we briefly introduce the underly-ing multihop relaying framework for downlinktransmission. As depicted in Fig. 1a, there areusually three types of links involved in end-to-end communication in multihop relaying cellularnetworks: the link from BS to RS (BS → RS),the link from RS to MS (RS → MS), and thelink from BS to MS (BS → MS). Furthermore,for the sake of clarity, we refer to the BS → RSlink as the relay link, while both the RS → MSand BS → MS links are called access links.Relaying should happen only when it canimprove the end-to-end throughput.

For the sake of description, the time-division-duplex (TDD) frame structure defined in LTE isused as an example to enable relaying technolo-gy in cellular networks. As shown in Fig. 1b,each radio frame of length Tf = 10 ms consistsof two half-frames with length Thalf = 5 ms each.Usually, each half-frame includes four commonsubframes of length Tsub = 1 ms and three spe-cial fields: downlink pilot time slot (DwPTS),uplink pilot time slot (UpPTS), and guard peri-od (GP). Each subframe comprises two slotswith length Tslot = 0.5 ms. For more detailedinformation on this frame structure, please referto the 3GPP specifications [1].

In practical implementation, the half-duplex

RS is usually assumed to be deployed in LTE-Advanced networks, where the transmission andreception take place in different subframes.When multihop relaying happens, the completetwo-way transmission over the air has four com-munication phases (i.e., BS → RS, RS → MS onthe downlink [DL], and MS → RS, RS → BS onthe uplink [UL]). In each phase one subframe isused for transmission; the basic transmissiongranularity in the time domain is one subframe,consisting of two successive time slots. Based onthis framework, we present several advancedinterference coordination schemes for multihopcellular networks in the next section, which canbe also applied in frequency-division duplex(FDD) systems.

ADVANCED INTERFERENCECOORDINATION SCHEMES

Different from traditional cellular networks, co-channel interference occurs not only betweenneighboring BSs or RSs but also between nearbyBSs and RSs involved in multihop transmission.In order to mitigate these interferences, suitableradio resource allocation and scheduling mecha-nisms in such a multihop cellular networkbecome vital. In this section we propose severalstatic or semi-static interference coordinationschemes for downlink transmission.

In general, the interference between neigh-boring RSs or BSs is measured when the net-work is initialized, or the user distribution andservices slowly vary. Then, according to mea-surement results, the RSs that do not causesevere interference to each other are groupedtogether and reuse the same radio resources.The restrictions on radio resource usage for BSsand RSs can be carried on along the frequencyor time domain (i.e., one-dimensional), or bothof them (i.e., two-dimensional) in an OFDM-based multihop cellular network. Then, withthese restrictions, each BS or RS can scheduleits serving users without explicit internode com-munication.

ONE-DIMENSIONALINTERFERENCE COORDINATION SCHEMES

In order to illustrate 1D interference coordina-tion schemes clearly, we show a typical deploy-ment scenario in Fig. 2a, where each cell ispartitioned into three sectors and a fixed RS isput on the bore sight line of the directionalantenna in each sector. First, every neighboringBS or RS takes turns transmitting the referencesignal while others keep silent and measure thereceived signals during the measurement period.The received signal power from RSi at RSj (i ≠ j)is denoted Pji. Then each RS reports the mea-sured Pji to its anchor BS. Next, an interferencemeasurement table is formed at the anchor BS,which can be used as the guide for RS grouping.Define the interference power threshold Pth asthe maximum value of interference power an RScan tolerate with acceptable communication.The simple grouping criterion is that RSi andRSj can be selected into the same resource reuseset only when Pji < Pth. For example, the inter-

Most interferencecoordination schemesfor OFDM-based mul-tihop cellular net-works are only theextensions of exist-ing coordinationstrategies used intraditional cellularnetworks. The oppor-tunities of multihoprelaying transmis-sions in dealing withco-channel interfer-ence problems havenot been fullyexploited.



ference power threshold Pth = –80 dBm is pre-defined in this section. According to measure-ment results in Fig. 2a, the nine RSs concernedare grouped into three RS reuse sets (i.e., G1 ={RS1, RS7, RS8}, G2 = {RS2, RS5, RS6}, and G3= {RS3, RS4, RS9}). Note that every two setshave a null intersection (i.e., Gm ∩ Gn = ∅, if m≠ n). In the resource partitioning stage, all RSsin the same reuse set can be allocated to theradio resources with the same pattern. There-fore, in the remaining parts of this article, weonly concentrate on resource coordinationamong the RSs in the different groups (e.g., RS1∈ G1, RS2 ∈ G2, and RS3 ∈ G3).

Frequency Domain Interference Coordination Scheme —In this scheme different priorities of the radioresources for different users are defined in thefrequency domain before resource allocation. Ateach BS or RS, the radio resources with highpriority are first assigned to its remote users thatexperience strong interference with high proba-bility. On these high-priority radio resources, thesignals can be transmitted with full radiatedpower. If the data amount of the servicedemands exceeds the throughput the BS or RScan provide only with high-priority radioresources, other radio resources with low priorityhave to be allocated under the constraint oflower radiated power.

In Fig. 2b we give an example of high-priorityradio resource allocation for three RS groups,which is based on the measurement results in

Fig. 2a. In this frequency domain (FD) interfer-ence coordination scheme, every two successivesubframes are allocated for BS and RS transmis-sion, respectively. In the first subframe a part ofradio resources along the frequency domain isorthogonally allocated to the relay links (BS →RS) of the neighboring sectors. Meanwhile, sincemost single-hop users are usually close to theBS, they are insensitive to intercell interference.The other radio resources are shared by the sin-gle-hop transmissions between BS and MS in theneighboring sectors. The radio resources are notevenly divided among the BS → RS link and theBS → MS link, which is dependent on the ratioof the single-hop user number and two-hop usernumber. In the second subframe the RSs trans-mit to their serving MSs, the remote MS are firstallocated high-priority radio resources, which areorthogonal to each other in the frequencydomain in order to eliminate strong co-channelinterference. Then, for those users near theirserving RSs, low-priority radio resources can beused for transmission with lower radiated powerwithout much performance loss. Usually, a line-of-sight (LOS) channel is assumed in the BS →RS link, and a non-line-of-sight (NLOS) in theRS → MS link. The quality of radio channels onthe BS → RS link is much better than that onthe RS → MS link. Hence, the higher modula-tion and coding scheme (MCS) for data trans-mission can be applied in BS → RS links.Consequently, the throughput balance betweenthe BS → RS link and the RS → MS link can be

Figure 1. Illustration of the framework in a multihop cellular network: a) illustration of the different links in a multihop cellular network;b) frame structure.

Onesubframe

DwPTS GP UpPTS

One slot

Subframe #0 Subframe #2 Subframe #3 Subframe #4 Subframe #5 Subframe #7 Subframe #8 Subframe #9

One radio frame

One half-frame

Access linkMS BS RS MS

Relay link Access link

Single hop

End-to-end connection End-to-end connection

First hop Second hop

DwPTS

(b)

(a)

GP UpPTS



Figure 2. Illustration of 1D interference coordination schemes in OFDM-based multihop cellular networks: a) deployment of one RS persector; b) resource allocation by the frequency domain IC scheme; c) resource allocation by the time domain IC scheme; d) flowchart ofthe frequency domain IC scheme; e) flowchart of the time domain IC scheme.

Subframe#3

Subframe#2

• Let same type of links in the neighboring sectors be in different subframes

• Reduce the radiated power of the BS in the subframes for BS → MS link.• Use beamforming in the BS → RS link (optional)

(e)

Resource allocation

• Group each three successive subframes together as a basic coordination unit.

• In each sector, assign each subframe to one type of link, (i.e. BS → RS, BS → MS, or RS → MS link.

Power coordination

Start

End

(d)

Frequency

Time

BS → RS1

RS1-MS BS-RS1

UplinkRadio frame i Radio frame i+1

Subframe#0

Subframe#9

Receivedsignal

from RS

Powervaluelevel

(a) (b)

(c)

After RS grouping

RS5

RS6

RS7

RS8

RS9

RS1

RS2

RS3

RS4RS1 -75dBmRS2 -90dBmRS3 -90dBmRS5 -60dBmRS6 -97dBmRS7 -90dBmRS8 -93dBmRS9 -87dBm

Receivedsignal

from RS

Powervaluelevel

RS1 -80dBmRS2 -60dBmRS3 -90dBmRS4 -70dBmRS5 -82dBmRS6 -90dBmRS7 -90dBmRS9 -76dBm

BS → MS1

subframe

RS1 → MS

1subframe

1subframe

1subframe

1subframe

BS-RS2 RS2-MS BS-MS

1subframe

1subframe

1subframe

BS-MS BS-RS1 RS1-MS

1subframe

1subframe

1subframe

Frequency

RS3RS1

RS2

Time

BS → RS2

BS → MS1

subframe

RS2 → MS

1subframe

Frequency

Time

BS → RS3BS → MS

1subframe

RS3 → MS1

subframe

RS2

RS3RS1

Sector 1

RS4 RS5

RS1

RS9

RS2

RS9

Sector 2

BS1

BS3

RS3 RS6

RS7

Sect

or 3

BS2

Downlink Downlink

Subframe #3Subframe #2


Subframe #0Subframe #9Subframe #8

Downlink Downlink

BS-MS

Start

Coordination in the first frame

Coordination in the second frame

Power allocation

End

Yes

No

Allocate orthogonal frequency resources to BSs inneighboring sectors for BS → RS link.Allocate other resources to BSs for BS-MS link.

•

•

Resource allocation

Full radiated power for the BS → RS linkResources are used by its neighboring BSs for BS → RS link?

••

Half radiated powerfor BS → MS link.

Full radiated powerfor BS → MS link.

Full radiated power on high-priority resourcesHalf radiated power on low-priority resources

••

Power allocation

RSs in different groups allocate orthogonal frequency resources to their CEUs as high-priority resourcesEach RS uses the remained frequency resources as its low-priority resources

•

•

Resource allocation



achieved with the suitable design. Figure 2dgives the flowchart of the FD interference coor-dination scheme.

Time-Domain Interference Coordination Scheme — Mul-tihop relaying transmissions in the downlink usu-ally consist of two phases in sequence (i.e., firstfrom BS to RS and then from RS to MS). Con-sidering this characteristic, the subframes allo-cated to different transmission phases can becoordinated among cells in the time domain.According to the channel path loss modeldefined in [9], the strongest interference to theremote users served by an RS mostly comesfrom their neighboring RSs. Thus, it is advisableto ensure that the resources for RS → MS linksof the neighboring cells are kept orthogonal inthe time domain to mitigate the interference.The time domain interference coordinationscheme is proposed and described as follows.

In the time domain, different subframes canbe used by one of three types of links (i.e., BS →MS, BS → RS, and RS → MS). Figure 2c showsan example of resource allocation pattern forthree neighboring sectors, where BS → MS rep-resents the single-hop link from BS to MS; BS→ RSi, i ∈ {1, 2, 3}, is the relay link of the two-hop transmission; and the access link of the two-hop transmission is denoted RSi → MS, i ∈ {1,2, 3}. Different subframes are allocated to dif-ferent types of links within one sector, whilethose of the same type of links among the neigh-boring sector are not overlapped in the timedomain. Thus, every three successive subframesare used together as a unit for the coordinationamong these three links by the time domaininterference coordination scheme. The flowchartof the time domain interference coordinationscheme is also shown in Fig. 2e.

By this allocation method, the interferencefrom neighboring BSs becomes the dominantfactor, affecting the performance of RS → MSlinks in the anchor cell. To overcome this prob-lem, beamforming can be applied at the BSs fortransmission of the relay link, which concen-trates the transmit power in the direction of aparticular user and minimizes the interferencefrom BS → RS links in the neighboring cell toRS → MS links in the anchor cell. Furthermore,a neighboring BS can reduce its transmit powerin BS → MS links for single-hop users located inthe cell center, thus also mitigating the interfer-ence to RS → MS links in the anchor cell. Mean-while, the interference on other links such as BS→ MS and BS → RS can also be decreased. Inaddition, the ratio of radio resources betweensingle- and two-hop links can be varied with thedistribution of different types of users.

Performance Comparison — The simulations are car-ried out to demonstrate the performances of thepriority-based interference coordination (IC)schemes. Most simulation assumptions followthe evaluation methodology as in [9, 10] and aresummarized in Table 1. The scenario with inter-site distance (ISD) of 1500 m is considered,where the operating bandwidth of 10 MHz islocated at the central frequency of 2 GHz fordownlink transmission. A penetration loss of 10dB is assumed for the access links (i.e., BS →

MS and RS → MS). For simplicity, only a single-antenna configuration is assumed for each node.Statistics are collected from a total of 20 dropsin the simulations. In each drop, users are uni-formly distributed around each BS with a densityof 90 users/cell. The simulation time span is 50 s(1000 radio frames) in each drop. The idealhexagonal cell is assumed for each BS, and twotiers of cells are considered with respect to onereference cell in the center (i.e., a total of 19hexagonal cells). Moreover, each cell is parti-tioned into three 120˚ sectors, where one 120˚directional antenna for each sector is applied atthe BS. For each sector, only one RS is deployedon the bore sight line of the directional antennaof the BS. The distance between the RS and itsanchor BS is two thirds of the cell radius.

According to the received SINR, all users canbe classified into two types: CEU and CCU. Theratio between the number of CEUs and CCUs ineach cell is set to be 1:2. If the frequency domain(FD) IC scheme is applied, the signal is trans-mitted on the high-priority resources with fullradiated power, while only half of the full radiat-ed power is allowed on the low-priority ones.With the time domain (TD) IC scheme, the radi-ated power ratio between each BS and RS in thenetworks is set to be 2:1 for fair comparison. Wealso provide the performance of the networkwithout the IC scheme for comparison.

The full-buffer traffic model is first assumedfor simulations, in which there are infinite datapackets in the queues. Figure 3a shows thecumulative distribution function (CDF) of thereceived SINR per cell in the network with/with-out 1D IC schemes. Since the interference powerto the high-priority resources used by the CEUsis decreased by using the FD IC scheme, theSINR performance of CEUs can be improved.However, for the CCUs in the network with theFD IC scheme, the corresponding SINR valuesare decreased because the radiated power on thelow-priority resources is reduced. Moreover, itcan be seen that the SINR values of all users inthe network with the FD IC scheme are limitedto be no more than 16 dB. This is because theperformance of the two-hop users is restrictedby the quality of the relay link (BS → RS), whoseSINR is no larger than 16 dB due to the inter-sector interference. Different from the FD ICscheme, the SINR performance is apparentlyimproved for all users including CEUs andCCUs by using the TD IC scheme. Meanwhile,parts of interference to the relay link (BS → RS)come from the neighboring RSs with lower radi-ated power instead of the BSs by using the TDIC scheme. Therefore, the highest SINR valueof the two-hop users, restricted by the relay link,is increased to around 18 dB. The comparison ofper user throughput performance in the net-works with/without 1D IC schemes is also givenin Fig. 3b. Similar to the SINR performance, theFD IC scheme can achieve significant through-put improvement for the CEUs with little per-formance degradation of the CCUs. In thenetwork with the TD IC scheme, the throughputis increased for both CEUs and CCUs. Further-more, the aggregated cell throughput of the net-works without or with 1D interferencecoordination schemes is given in Fig. 3c. Com-

Multihop relayingtransmissions in the

downlink usuallyconsist of two

phases in sequence (first from BS to RSand then from RS to

MS). Consideringthis characteristic,

the subframes allocated to differenttransmission phasescan be coordinatedamong cells in the

time domain.



pared to the network without the IC scheme, alarge throughput gain for the CEUs (i.e., 15.9percent) can be obtained with the FD IC scheme.However, it decreases the aggregated cellthroughput of the CCUs. On the other hand, theTD IC scheme can improve the throughput per-formance of both CEUs and CCUs, with morethan 18.3 and 15.6 percent throughput gain com-pared to the case without the IC scheme, respec-tively.

Thus, we can conclude that the FD IC schemecan only improve the performance of the CEUswhen imposing little restriction on the radio allo-cation. On the other hand, the performance ofboth CEUs and CCUs can be enhanced by theTD IC scheme. However, the TD IC scheme hasless flexibility because three subframes have tobe grouped together for resource allocation.

Next, we study the effects of 1D IC schemeson real-time voice over IP (VoIP) service inFig. 3c. Since the SINR performance of CEUscan be improved by applying the IC schemes,the VoIP capacity of the CEUs is increased,with 14.3 and 28.6 percent gain correspondingto the FD IC and TD IC schemes compared tothe case without the IC scheme, respectively.Usually, at least one or two resource blocks(RBs) are needed to support the transmissionof each VoIP packet for the CCUs. The smallSINR degradation due to the FD IC schemedoes not affect the number of allocated RBs,so the VoIP capacity of the CCUs does notdecrease. With the TD IC scheme, the SINRperformance of the CCUs is also enhanced andleads to VoIP capacity improvement. Thus, theVoIP capacity of all users in the network withthe IC schemes outperforms that without the

IC scheme, with 7.5 and 15.1 percent relativegain with the FD IC and TD IC scheme,respectively.

TWO-DIMENSIONALINTERFERENCE COORDINATION SCHEMES

For network deployment with more than one RSper sector, the coordination scheme operatingsolely in the FD degrades the potential frequen-cy diversity gain, while either too small timegranularity or a group with too many subframesis needed for the scheme operating only in theTD. It is therefore necessary to develop a 2Dtime-frequency interference coordination schemewith more flexibility for an OFDM-based multi-hop cellular network.

Assume that there are two RSs deployed persector as shown in Fig. 4a. According to theinterference measurement results, all RSs aregrouped into the following six sets (i.e., G

~

1 ={RS1, RS15, RS17}, G

~

2 = {RS2, RS16, RS18}, G~

3= {RS3, RS12, RS14}, G

~

4 = {RS4, RS11, RS13},G

~

5 = {RS5, RS8, RS9}, G~

6 = {RS6, RS7, RS10}).Recall that the RSs in the same set can reusethe same radio resources. The intersector co-channel interference can be mitigated by meansof the TD coordination scheme, while the intra-sector interference is dealt with by the FD coor-dination scheme. In Fig. 4b, for TD resourceallocation, all RSs in six sets can also be com-bined into three pairs: P1 = {RS1, RS2}, P2 ={RS3, RS4}, and P3 = {RS5, RS6}. Then differ-ent kinds of links including RS → MS, BS →RS, and BS → MS in each paired RS set are dis-tinguished in the TD within one sector. More-over, the same kind of links between theneighboring sectors are allocated to differentsubframes without overlap in the TD.

Then further radio resource allocation withinone sector among different RS sets is performedin the FD. The total frequency bandwidth foreach RS set is partitioned into two orthogonalbands,high-priority and low-priority. The high-priority bands of two different RS sets in thesame sector are orthogonal. For each RS, it isnecessary to first allocate the radio resources inthe high-priority band to the remote MSs withfull radiated power, while those in the low-prior-ity band with lower radiated power can be usedby the MSs that are close to the RS. In this waythe intrasector co-channel interference, whichcomes from the different reuse sets, can bedecreased efficiently. Also, we show theflowchart of the proposed 2D interference coor-dination scheme in Fig. 4c. Note that this schemecan be extended to cases with larger numbers ofRSs.

Performance Comparison — The performances ofthe multihop cellular networks with the 2D ICscheme are also evaluated under the scenariowith the parameters given in Table 1. In thesesimulations two RSs are deployed per sectorwhile one 60˚ directional antenna is deployedfor each BS → RS link. The number ratio ofCEU to CCU is set as 1:2, and the transmitpower ratio between each BS and RS is 3:1.

Figure 5a gives the cumulative distributionfunction (CDF) of the received SINR in the net-

Table 1. Parameter assumptions in a multihop cellular network.

Parameters Values

Intersite distance (ISD) 1500 m

Bandwidth/carrier frequency 10 MHz/2 GHz

Antenna configuration (BS/RS/MS) 1/1/1

PathlossBS-MS 128.1 + 37.6 log10(R)

BS-RS LOS/NLOSType D LOS in [9]/128.1 + 37.6 log10(R), R in km

RS-MS LOS/NLOS Type C/Type A in [9]

Transmit power/height BS 46 dBm/32 m

RS 37 dBm/12.5 m

Penetration loss 10 dB

Thermal noise spectral density –174 dBm/Hz

Noise figure 5 dB

Traffic model Full buffer



work with or without the 2D IC scheme. Withthe resource coordination in both the TD andFD, the SINR performance of all users isimproved by the 2D IC scheme, especially forthe CEUs. Since the 60˚ directional antenna isapplied to the BS → RS link in the case withtwo RSs per sector, less power leaking occurscompared to the case of only one RS per sector.Hence, the intersector interference on the BS →MS link is greatly decreased. In addition, as themain interference of the BS → RS links is notcaused by leaking power of the directional anten-na from neighboring sectors, the performanceimprovement of two-hop users with high SINR isnot obvious when the 2D IC scheme is applied.In Fig. 5b the per user throughput performanceof the network with/without the 2D IC scheme isalso compared. Similarly, it can be seen that allusers in the network can achieve significantthroughput improvements, especially the CEUs,by using the 2D IC scheme. Moreover, we givethe cell throughput of the network with or with-out the 2D IC scheme in Fig. 5c. Compared tothe case without the IC scheme, aggregated cellthroughput is increased by more than 25.2 per-

cent by using the 2D IC scheme, while about16.4 and 61.9 percent throughput gains areobtained for the CEUs and CCUs in the net-works, respectively.

In Fig. 5d we compare the VoIP capacity ofthe network with or without the 2D IC scheme.Both CEUs and CCUs can obtain SINR perfor-mance gain when the 2D IC scheme is applied inthe networks, resulting in improvement of theVoIP capacity of all users. For example, about46.6 percent VoIP capacity gain is achieved bythe 2D IC scheme when all users are considered.

CONCLUSIONS

Interference coordination is essential forOFDM-based cellular multihop networks inorder to solve the problem of co-channel inter-ference and achieve high spectral efficiency. Inthis article we present several semi-statictime/frequency interference coordinationschemes, demonstrating their applicability andefficiency in 3G LTE networks toward furtheradvancements. In our analysis we have shownthat the frequency domain interference coordi-

Figure 3. Performance comparison between OFDM-based multihop cellular networks with/without 1D IC scheme: a) SINR distribution;b) per user throughput distribution; c) aggregated cell throughput; d) VoIP capacity.

CEU

1.46 1.491.26

CCU

(c)

6.21

7.25

6.27

All users

7.67

8.73

7.53

0

5

10

0

Agg

rega

ted

cell

thro

ughp

ut (

Mb/

s)

(d)

All users

5761

7.5% 15.1%

53

CEU

1618

14.3% 28.6%

14

0

5

65

0

VoIP

Cap

acit

y (u

sers

/MH

z/se

ctor

)

10

15

20

25

30

35

40

45

50

55

60w/o ICwith FD ICwith TD IC

SINR (dB)

(a)

-2-4

0.1

0

CD

F

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

0 2 4 6 8 10

All usersAll users

CEU

CEU

12 14 16 18 20Per user throughput (kb/s)

(b)

0

0.1

0

CD

F

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

600500400300200100

w/o ICwith TD ICwith FD IC

w/o ICwith TD ICwith FD IC

w/o ICwith FD ICwith TD IC



nation scheme can only improve the perfor-mance of CEUs, while the performance of bothCEUs and CCUs can be enhanced by the timedomain interference coordination scheme with aslight strict limitation. Furthermore, the frequen-cy-time domain interference coordinationscheme is proposed to achieve throughput gainwith high flexibility.

ACKNOWLEDGMENT

This work was supported in part by the ChinaNSFC under Grant 60802082, National KeyTechnology R&D Program of China underGrant 2009ZX03003-008-01, and Research Fundfor the Doctoral Program of Higher Educationunder Grant 200800131023.

Figure 4. Illustration of 2D interference coordination schemes in OFDM-based multihop cellular networks: a) deployment of two RSs persector; b) resource allocation by the 2D IC scheme; c) flowchart of the 2D IC scheme.

(a)(b)

BS-MSBS-RS6 RS6-MS

BS-RS5 RS5-MS

1subframe

1subframe

1subframe

RS2RS1

RS3

RS4

RS5RS6

RS8 RS13

BS1RS7

RS1 RS2

RS14 Subframe#3

Subframe#2

Frequency

Time

BS-MSBS-RS4 RS4-MS

BS-RS3 RS3-MS

1subframe

1subframe

1subframe

Frequency

Time

BS-MSRS2-MS BS-RS2

RS-MS BS-RS1

1subframe

1subframe

1subframe

Frequency

Time


Subframe#0

Subframe#9

Subframe#8

Downlink Downlink

RS9 RS4

BS3RS10

RS17 RS18

RS3

RS5 RS11

BS2RS6

RS15 RS16

RS12

Time-domain coordination(mitigate the inter-sector interference)

(c)

Start

End

Frequency-domain coordination(mitigate the intra-sector interference)

Power coordination• Half radiated power in the BS → MS subframes.• Use beamforming in the BS → RS subframes (optional).

Resource allocation• Conjuct each three successive subframes together, and group the same kinds of links in each sector.• Assign the subframes to different kinds of links, each for certain kind of link.• Keep the same kind of links in the neighboring sectors in different subframes.

Power allocation• Full radiated power on high-priority resources.• Half radiated power on low-priority resources.

Resource allocation• In the subframes for BS → RS and RS → MS links, orthogonally allocate the frequency resource to BSs and RSs in different groups as high-priority resources.• Assign the remained frequency resources to BSs and RSs as low-priority resources.



REFERENCES[1] 3GPP TR 36.211 v. 8.6.0, “Evolved Universal Terrestrial

Radio Access (E-UTRA), Physical Channels and Modula-tion (Release 8),” Mar. 2009.

[2] ITU-R M.1645 Rec., “Framework and Overall Objectivesof the Future Development of IMT-2000 and SystemsBeyond IMT-2000,” June 2003.

[3] ITU, “Final Acts — WRC-07,” WRC ‘07, Dec. 2007;http://www.itu.int/publ/R-ACT-WRC.8-2007/en.

[4] Y. Yang et al., “Relay Technologies for WiMAX and LTE-Advanced Mobile Systems,” IEEE Commun. Mag., vol.47, no. 10, Oct. 2009, pp. 100–5.

[5] IST WINNER D3.5.1, “Relaying Concepts and SupportingActions in the Context of CGs”; http: / /www.ist-winner.org/WINNER2Deliverables/D3.5.1v1.0.pdf.

[6] B. Can et al., “Implementation Issues for OFDM-BasedMultihop Cellular Networks,” IEEE Commun. Mag., vol.45, no. 9, Sept. 2007, pp. 74–81.

[7] 3GPP R1-050507, “Soft Frequency Reuse Scheme forUTRAN LTE,” Huawei, 3GPP RAN WG1 Meeting #41,May 2005.

[8] 3GPP R1-050738, “Interference mitigation-Considera-tions and Results on Frequency Reuse,” Siemens, 3GPPRAN WG1 Meeting #42, Aug. 2005.

[9] IEEE 802.16j-06/013r3, “Multihop Relay System Evalua-tion Methodology (Channel Model and PerformanceMetric),” Feb. 2007.

[10] 3GPP TR 36.814 v. 1.4.1, “Further Advancements forE-UTRA Physical Layer Aspects (Release 9),” Sept. 2009.

BIOGRAPHIESKAN ZHENG [M‘03, SM‘09] ([email protected]) received hisB.S., M.S., and Ph.D. degree from Beijing University ofPosts & Telecommunications, China, in 1996, 2000, and2005, respectively, where he is currently an associate pro-fessor. He worked as a researcher in companies includingSiemens and Orange Labs R&D, Beijing, China. His currentresearch interests lie in the field of signal processing fordigital communications, with an emphasis on cooperativecommunication and heterogeneous networks.

BIN FAN received his B.S. and Ph.D. from Beijing Universityof Posts & Telecommunications, China, in 2005 and 2010,respectively. Now he works in Orange Labs R&D, Paris,France. His current research interests lie in the field ofcooperative communication and radio resource manage-ment.

JINHUA LIU received a B.S. from Beijing University of Posts &Telecommunications in 2008, where he is currently work-ing toward an M.S. degree. His current research interestlies in the field of signal processing for cooperative com-munication.

YICHENG LIN received his B.S. and M.S. from Beijing Universi-ty of Posts & Telecommunications, China, in 2007 and2010, respectively. His current research interest lies in thefield of signal processing for radio resource management.

Figure 5. Performance comparison between OFDM-based multihop cellular networks with/without 2D IC schemes: a) SINR distribution;b) per user throughput distribution; c) aggregated cell throughput; d) VoIP capacity.

(c)

CCU

8.36

7.18

CEU

2.77

1.71

All users

11.13

8.89

0

5

0

Agg

rega

ted

cell

thro

ughp

ut (

Mb/

s)

VoIP

cap

acit

y (u

sers

/MH

z/se

ctor

)

10

(d)

CEU

58.8%

46.6%

27

17

All users

85

58

0

10

0

20

30

40

50

60

70

80

90 w/o ICwith 2D IC

SINR (dB) Per user throughput (kb/s)

(a)

-2

0.2

0

CD

F

0.1

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

0 2 4 6 8 10 12 14 16 18

CEU

CEU

All users

All users

20 22

(b)

0

CD

F

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

700 800 600 500 400 300 200 100 0

w/o ICwith 2D IC

w/o ICwith 2D IC

w/o ICwith 2D IC


INTRODUCTIONEvery year freeway accidents and obstructionsresult in traffic congestion [1], environmentalpollution, and damages in property, personalinjury, and/or fatalities. The overall cost to soci-ety is in the billions of dollars and increasing [1].With the availability of advanced sensor tech-nologies, preventing and/or minimizing theimpact of these occurrences has been a majorfocus in highway incident detection research inthe past decade. An important goal in theseefforts has been to reduce the response times indetecting the incident. Timely and accurate inci-

dent detection is critical to notify approachingdrivers, dispatch emergency response teams todeal with the injured, clear the road of theimpeding obstruction, to return traffic patternsto the normal flow, and prevent subsequent sec-ondary incidences.

Automated incident detection (AID) meth-ods have been developed for detecting potentialincidents since the 1970s. However, early deploy-ments of these systems were relatively ineffectivesince they had either high false alarm rates(which rendered them ineffective) or low detec-tion rates (which rendered them unreliable). Therecent advances in sensor technologies and wire-less devices have created a new paradigm fordesign and development of AID systems. Recentsystems have incorporated complex algorithmsto predict and detect incidents and their loca-tions [2]. Unfortunately, even these elaboratemethods have unacceptable levels of false posi-tive and false negative rates.

In this article, we propose an AID systemthat leverages vehicular ad hoc networks(VANETs) and is implemented as an applicationon the VGrid framework [3]. Within a geograph-ic area, an ad hoc network of vehicles equippedwith sensors and in-vehicle processing capabilitycan form an ad hoc cluster of sensors and gridcomputers, which we refer to as VGrid. VGrid isa control and management infrastructure frame-work, where VGrid collects and processes real-time traffic data using in-vehicles sensors andcomputers to make distributed control decisions.VGrid can complement and extend the existingfixed infrastructure by distributing previouslycentralized control functions, such as trafficstatistics collection and dissemination of trafficadvisory messages, to local intelligent agents,including in-vehicle sensors and computers.

As part of its design, VGrid vehicles broad-cast beacon messages that include informationabout the vehicle’s speed, position, and lane.These messages are then collected by nodes1

that are within the broadcast range of transmit-ting vehicles. Using the aggregated information,each node independently estimates and main-tains information on the occupancy, lanechange, and speed for different sections of the


ec. 5 Sec. 6 Sec. 7

3

ec. 5 Sec. 6 Sec. 7

1 2

ec. 5 Sec. 6 Sec. 7

1 Nodes refer to VGrid vehicles that can store and processinformation and, using wireless communication, commu-nicate with other VGrid vehicles. Throughout this articlewe use the terms VGrid vehicle and node interchangeably.


BEHROOZ KHORASHADI, FRED LIU, DIPAK GHOSAL, MICHAEL ZHANG, AND CHEN-NEE CHUAH,UNIVERSITY OF CALIFORNIA, DAVIS

ABSTRACTIn this article, we study an ad hoc distributed

automated incident detection algorithm for high-way traffic using vehicles that are equipped withwireless communications, processing, and stor-age capabilities (referred to as VGrid vehicles).Each VGrid vehicle periodically broadcasts bea-con messages with its speed, location, and laneinformation. Using these beacons, each VGridvehicle builds and maintains information aboutdifferent sections of the road. Using such infor-mation, each VGrid vehicle independently per-forms an anomaly detection algorithm based onthe traffic density, speed, and the number oflane changes to identify incidents. The robust-ness of the detection is improved by a votingscheme in which a consensus, among participat-ing VGrid vehicles, is achieved when a fixednumber of votes are accumulated. We use a sim-ulation tool called VGSim to study the perfor-mance of our detection algorithm in a highwayscenario. The results show that our distributedincident detection algorithm has low false posi-tive rate, zero false negative rate, and can stillachieve incident detection with as little as 10percent penetration of VGrid vehicles.

DISTRIBUTED AUTOMATEDINCIDENT DETECTION WITH VGRID

The authors study anad hoc distributedautomated incidentdetection algorithmfor highway trafficusing vehicles thatare equipped with wireless communications, processing, and storage capabilities(referred to as VGridvehicles).

LIU LAYOUT 2/7/11 10:47 AM Page 64


road. The information is processed to detectanomalous traffic patterns to identify staticobstructions in the road. The initial individualdetection phase is used to detect all potentialincident points, and then collaboration and veri-fication among the nodes is used to filter outerroneous predictions. The proposed methoddecreases both the false positive and false nega-tive rates to acceptable levels. In addition,because each vehicle acts as a low-cost sensorthat collects and processes the informationreceived from other sensors, the detection timeis, in most cases, very low (on average about 20s). The low cost of each device also allows moreefficient deployment compared to systems thatrequire devices that must be installed at road-sides or in the road itself.

The rest of this article is organized as follows.In the next section we discuss related work onAID. We then give a detailed description of theproposed incident detection approach. Webriefly describe the simulation tool and the sce-nario for which we evaluate the performance ofour proposed incident detection algorithm. Wethen discuss the results and finally, we concludewith a summary of our results and outline futurework.

RELATED WORK

In general there are four main types of AIDdetection algorithms. They are based on patternrecognition, catastrophe theory, statistical analy-sis, and artificial intelligence (AI) techniques [4].Of these four types, statistical and pattern recog-nition based methods were the first to be devel-oped in the 1970s. Later, with the emergence ofmore high-powered computers, AI techniqueswere applied.

The earlier work in AID used techniquesbased on decision trees for traffic pattern recog-nition [5], time series analysis [6], and Kalmanfilters [7]. Based on these techniques, threemajor AID algorithms were developed: the Cali-fornia Algorithm, the McMaster Algorithm, andthe Minnesota Algorithm. The California Algo-rithm utilizes decision trees to identify anoma-lous lane densities between detection points.The California Algorithm, also referred to asTraffic Services Corporation (TSC) Algorithm-2,uses a five-minute roll-wave suppression logicthat helps reduce false alarms due to shockwaves from downstream. The key advantages ofthe California Algorithm is that it is comprehen-sive in nature and has low false positive rates.Unfortunately, the California Algorithm makesdecisions based solely on the occupancy rate ofthe road; other parameters such as vehicle speedand volume are not used.

The McMaster Algorithm is based on catas-trophe theory [8]. It classifies traffic conditionsinto subcategories and determines if traffic pat-terns fall within these categories. They are robustin the sense that they are not affected byupstream input failures. However, the McMasterAlgorithm only tracks data from the fast lane.Finally, the Minnesota Algorithm, which is simi-lar to the California Algorithm, uses the occu-pancy rates. This, however, leads to the samedisadvantage as the California Algorithm but

also has the added limitation of creating a largenumber of false alarms in low traffic density sce-narios.

Examples of AID systems based on statisticalmethods include the DES algorithm summarizedin [4], and the Standard Normal Deviation(SND) [4] and Single-Station Incident Detection(SSID) algorithms [9]. SSID uses the standardstatistical T-test to analyze the differences inoccupancy.

The introduction of AI techniques has result-ed in new AID approaches. The study reportedin [2] uses a combination of fuzzy logic andgenetic algorithms (GAs). While this approach ishighly adaptive, both fuzzy logic and GAs arehighly complex algorithmic approaches that haveyet to be proven to be efficient. The AI-basedapproaches for AID systems have also focusedon artificial neural networks which contain mul-tiple layers, multiple inputs, and a complexstructure. This approach has been shown to per-form better than fuzzy logic and GA-basedapproaches; however, complexity and accuracyare still major drawbacks. The use of AI in con-junction with modern camera technology wasinvestigated in [10]. Unfortunately, this approachis very expensive since cameras must be installedand maintained. An example of this is the videoor 3D image processing adopted in [11].

In recent years, VANET has provided aunique framework to develop new techniques forAID. In the research reported in [12, 13], a cen-tralized detection approach is proposed for AID.These methods rely on information collectedand reported periodically at roadside nodes byVANET vehicles. These roadside nodes thenaggregate the information collected from indi-vidual vehicles to detect obstructions upstreamfrom the roadside node. Although this workaddresses a similar problem (AID usingVANET), there are fundamental differencesthat distinguish it from the work presented inthis article. A major difference is that themethod proposed in this article is a fully dis-tributed approach that leverages the fully dis-tributed VGrid computing framework in whichvehicles form an ad hoc network with no infra-structure coordination to detect anomalies invehicular traffic flows. This distributed design ofthe AID scheme has significant advantages, asmentioned later in this article. In addition, theproposed detection method is evaluated usingVGSim [14], a simulation platform designedspecifically to simulate a realistic mobility modelfor vehicles and an integrated wireless network-ing protocol stack.

To date, the most effective and rapid detec-tion system seems to be based on cell phonecallers and/or roadside call boxes [4]. It has beenargued that using cell phones to call in roadsideincidents can be dangerous, and could result insecondary accidents and/or other traffic flowproblems. Furthermore, studies have shown thatas cell phone conversations increase, the likeli-hood that some highway traffic situation will gounreported and/or unnoticed also increases [15].The use of call boxes can be very effective butrequires the overhead of deployment; the morefrequent the call box stations, the greater thecost.

The key advantagesof the California

Algorithm is that it iscomprehensive innature and has alow false positive

rates. Unfortunately,the California Algo-rithm makes deci-

sions based solely onthe occupancy rateof the road; other

parameters such asvehicle speed and

volume are not used.



All the previously mentioned AID techniquessort through data collected from roadside sen-sors or other similar data collection devices. Assuch, these algorithms are all constrained by thesame limitation; they can only achieve a certainlevel of accuracy and efficiency. VANET tech-nologies and the proposed VGrid framework[14] enable new approaches to AID. They pro-vide both spatially and temporally fine-graineddata regarding the speed, position, and lanechanges of vehicles. Furthermore, the distributedgrid computing (VGrid) framework implement-ed over VANET allows complex coordinatedcomputation over the data to accurately androbustly detect various patterns in the trafficflow and thereby detect incidents.

OBSTRUCTION DETECTION ALGORITHM

In order for vehicles to detect potential roadobstruction, vehicles must exchange traffic infor-mation. The VGrid framework [14] provides thisin the form of traffic beacon messages. Thesemessage are broadcasted periodically by eachVGrid vehicle and, in-turn, collected by all VGridvehicles that are within the transmission range ofthe broadcast. Over time each VGrid vehiclebuilds an independent and local view of the roadand its traffic pattern. Each vehicle periodicallysearches this aggregated information for specificmarkers that indicate road obstructions. Once avehicle detects an obstruction, it initiates a votingphase, in which nodes within the transmissionrange vote to reach a distributed consensus onthe detection. In the following subsections wedescribe the details of the algorithm.

BEACON MESSAGES AND ROAD PROFILINGThe core of this algorithm relies on the dissemi-nation of vehicle traffic information by all VGridvehicles. The beacon messages are broadcast at4 Hz (4 times/s) and contain the fields shown inTable 1. The Manual on Uniform Traffic Con-trol Devices (MUTCD) standardizes the parti-tioning of roads into logical sections bydemarcating highway sections into 100 ft incre-ments. Although this is primarily used for thedesign of highway systems, it additionally pro-

vides coordination between vehicles to acquireroad ID and position information which can beused to organize obstruction location informa-tion. To determine lane information multiplemethods exist, and we believe the simplestmethod is to use slightly higher accuracy GPSsystems than are available in today’s vehicles.Any GPS unit that can obtain ±1 ft accuracy issufficient to determine the lane informationneeded for this detection method. More accu-rate GPS units are also available that have sub-centimeter accuracy; however, that level of detailfar exceeds the needs of lane determination.

Through detailed experiments we found thata beaconing rate between 0.5 Hz and 20 Hz wasrequired to ensure the required performance ofour proposed AID scheme. For beaconing rateless than 0.5 Hz there was noticeable degrada-tion of the detection rate of the algorithm. Forbeaconing rate greater than 20 Hz, there werepacket drops due to collisions, which in turnresulted in higher detection time. Indeed, theimpact was higher for higher traffic densities,which were caused by local congestion near theobstruction point. The results shown in this arti-cle have been obtained for a beaconing rate of 4Hz. Note that the size of the beacon messages issmall (24 bytes), resulting in minimal impact onnetwork congestion when beacon frequency iswithin a given threshold. A more detailed analy-sis of the impact of beacon frequency on perfor-mance is reported in [14].

The beacon messages are collected by eachVGrid vehicle, and then used to form a dynamicview of the road and the traffic patterns. EachVGrid vehicle gathers traffic information anduses it to track the movement and position ofother VGrid vehicles. Tracking is done by simplystoring the beacon messages of each distinctvehicle. Since beacon messages are periodicwhen a new beacon is received, the receivingvehicle can infer the trajectory of the sourcevehicle. For example, assume that a vehicle xreceives a message from vehicle y at time 2 swith a beacon message that indicates a positionof 150 in lane 3. Later, vehicle x receives anotherbeacon from vehicle y at time 5 s with position200 and lane 3. Vehicle x infers that vehicle ytraveled from position 150 to 200 in 3 s. In theevent of a lane change, vehicle x makes theassumption that y changed lanes halfway betweenpoints 150 and 200.

Each VGrid vehicle maintains three informa-tion arrays for each lane, where each cell in thearray corresponds to a particular section of road.We choose road sections to be the length of avehicle in the simulation (7.5 m or 75 cells in thesimulation). In order to compensate for the reduc-tion in penetration rate (the percentage of VGridenabled vehicles) we then partition the roadinversely proportional to the penetration rate:

road_section_size = vehicle_length/penetration_rate(1)

These road section divisions are not relative tothe vehicle itself but to the location. The coordi-nation between vehicles can be achieved rela-tively simply as highways have already beennaturally partitioned using mile markers. Sec-

Table 1. Beacon message fields.

Field Type Size

Source IP address Integer 4 bytes

Position Integer 4 bytes

Lane number Byte 1 bytes

Time-to-live (TTL) Byte 1 bytes

Message timestamp Long 8 bytes

Vehicle speed Short 2 bytes

Road ID Integer 4 bytes

Total bytes 24 bytes

Over time each VGridvehicle builds anindependent andlocal view of theroad and its trafficpattern. Each vehicleperiodically searchesthis aggregated information for specific markers thatindicate road obstructions.



tions zero can start at the beginning of eachunique mile marker on the highway.

We use the following three arrays to collectinformation used for the detection algorithm.

The road section count (RSC) array indicatesthe number of vehicles that have been trackedover a particular section. In order to create theRSC array, vehicle x will increment each cell inthe RSC when a vehicle is either tracked over asection or is located in a particular section. Con-sider the previous example. If vehicle x receivesa beacon from vehicle y located in section 150 ofthe road, x will increment its count or RSC arrayby one at index 150 for lane 3. When the secondbeacon is received x will increment indexes151–200 in lane 3. If y were to change lanes theincrement would be for 151–175 in lane 3 and175–200 in lane 2 (assuming lane 2 was the lanevehicle y moved into).

The road section speed (RSS) array containsthe weighted average of speeds tracked over aparticular section of the road. The indices (whichcorrespond to road sections) in the RSS arrayare initialized to the maximum, which for simu-lation study reported in this article is set to 30m/s (≈60 mph). Vehicles are tracked over roadsections in the same fashion as in the RSC array.Speed is either calculated by tracking a vehicleover time or extracted directly from the beaconmessage itself. The RSS for an index i in thearray is updated or calculated using an exponen-tially weighted moving average given by

RSS[lane][i] = (1 – α)RSS[lane][i] + α * vtracked_speed, (2)

where 0 < α ≤ 1, lane is the lane which the vehi-cle is being tracked, and vtracked_speed is theobserved speed of the vehicle being tracked.Again, as a vehicle y is tracked over multiple sec-

tions, Eq. 2 is applied to each section vehicle ypasses over. A weighted moving average is usedto quickly reflect the speed changes in trafficover time. We chose an α = .25 to quickly exhib-it changes in traffic speeds in a particular sectionbut not overly emphasize the information dis-seminated by any one particular vehicle. Theend goal was to have newly disseminated trafficinformation about a recently created obstructionquickly dampen stale data in a road sectionwhere an obstruction did not previously exist.

The vehicle lane change (VLC) array con-tains a counter for each section of the road. If avehicle is observed changing out of a lane in aparticular section the index is decremented. If avehicle is observed entering a lane at a particularsection, the index in VLC is incremented. Thelane change point is an estimation of a vehicle’sactual location of change. However, becausebeacon messages are so frequent (4 Hz, which is4 beacon messages/s), and the sections of theroad are large, it was found that this estimationwas adequate for the detection process.

Figure 1 illustrates the obstruction detectionscenario at three different time instances, andTable 2 shows the RSC, RSS, and VLC arrayvalues for each time frame. In this example, weassume that all vehicles are within each other’stransmission range and there are no packet loss-es (due to interference and collision).

It is important to note that the proposedmethod relies on accurate positioning informa-tion. We assume that the VGrid vehicles have aGPS device that is accurate to within a few feetto pinpoint lane location. While currently mid-range GPS devices are accurate to within 10–15ft, the higher end units are accurate within 1–2 ftand would satisfy the requirements to accuratelydetermine the vehicles, lane, and position. Wealso found that there are production quality

Figure 1. Illustration of how vehicles move through in time. Table 2 represents the corresponding array val-ues as seen by vehicle 1 or 2 as both vehicles are present in all three time instances.

Lane 1

Lane 2 T=0

Sec. 1 Sec. 2 Sec. 3 Sec. 4 Sec. 5 Sec. 6 Sec. 7 Sec. 8 Sec. 9 Sec. 10

1 3

2 4

Lane 1

Lane 2 T=1

Sec. 1 Sec. 2 Sec. 3 Sec. 4 Sec. 5 Sec. 6 Sec. 7 Sec. 8 Sec. 9 Sec. 10

1

5

2 3

Lane 1

Lane 2 T=2

Sec. 1 Sec. 2

VGrid Vehicle X Road obstruction

Sec. 3 Sec. 4 Sec. 5 Sec. 6 Sec. 7 Sec. 8 Sec. 9 Sec. 10

1

5

X

2

The coordinationbetween vehicles can

be achieved relatively simply as

highways alreadyhave been naturally

partitioned usingmile markers.

Sections zero canstart at the

beginning of eachunique mile marker

on the highway.



units that can pinpoint position accuracy to with-in 2 cm.2 These devices are designed specificallyto measure vehicle positions and are currentlyused to collect acceleration and braking distancemeasurements. Accuracy within 2 cm is, ofcourse, not a requirement for functional perfor-mance of our proposed AID system. However, itis important to note that such technology isavailable and can be applied in the future. Addi-tional technologies that are able to track lanechange include devices that track a vehicles tiremovement, accelerometers coupled with GPSsystems, or other types of sensors such as thoseembedded within the road itself designating eachlane. Although lane determination is a requiredcomponent of this proposed AID system, wehave described several technologies that areavailable for lane detection. Consequently, webelieve that this requirement would not be acause for barrier to entry for our proposed AIDsystem.

DETECTION METHODDetection is carried out by each vehicle by peri-odically scanning the three arrays in search ofanomalous traffic patterns. In order to quantifythe search object, we first formulated the look of

a road obstruction from a purely human per-spective. Figure 2 depicts what we would expectfrom tracking vehicles around an obstructionpoint. The most prominent characteristic of anobstruction point will be the lack of vehicles inthe lane downstream from the obstruction. Inaddition, vehicles upstream from the obstruction,in the same lane, would attempt to merge out ofthat lane while no vehicles should enter thatlane. Conversely, vehicles after the obstructionpoint should attempt to merge back into thelane due to the available space. Finally, vehiclesstuck behind the obstruction point will travel atmuch slower rates than the normal traffic, andvehicles past the obstruction point should beable to travel at average traffic speeds (whereaverage speed is calculated by averaging allspeeds collected from beacon messages withinthe area).

The detection method consists of two phases.In the first phase individual VGrid vehicles inde-pendently search through the accumulated andaggregated traffic information collected frombeacon messages to identify potential obstruc-tion points. Once a vehicle identifies a potentialobstruction point, it moves to the second phaseof the algorithm, in which it broadcasts an

Table 2. Representation of RSC, RSS, and VLC arrays for vehicle 2 in Fig. 1.

Time Array Lane # Sec: 1 Sec: 2 Sec: 3 Sec: 4 Sec: 5 Sec: 6 Sec: 7 Sec: 8 Sec: 9 Sec: 10

T = 0

RSCLane 1 0 0 0 1 0 0 0 0 1 0

Lane 2 0 1 0 0 0 1 0 0 0 0

RSSLane 1 30 30 30 29 30 30 30 30 28.5 30

Lane 2 30 27.5 30 30 30 29 30 30 30 30

VLCLane 1 0 0 0 0 0 0 0 0 0 0

Lane 2 0 0 0 0 0 0 0 0 0 0

T = 1

RSCLane 1 1 0 0 1 1 0 0 0 1 1

Lane 2 0 1 1 1 2 2 1 1 1 0

RSSLane 1 29 30 30 29 7 30 30 30 28.5 29

Lane 2 30 27.5 20 20 14 20 29 29 29 30

VLCLane 1 0 0 0 0 –1 0 0 0 0 0

Lane 2 0 0 0 0 1 0 0 0 0 0

T = 2

RSCLane 1 1 1 1 2 1 0 1 1 2 1

Lane 2 0 1 1 1 2 3 3 1 1 0

RSSLane 1 29 28.5 28.5 28 7 30 27 27 26.5 29

Lane 2 30 27.5 20 20 14 21.5 25.2 25 29 30

VLCLane 1 0 0 0 0 –1 0 +1 0 0 0

Lane 2 0 0 0 0 1 0 –1 0 0 0

2 An example of such adevice can be found athttp://www.racelogic.co.uk/?show=vbox


obstruction detection vote. The vote contains thevehicles unique id and also the details concern-ing the obstruction location. The first vehicle tocollect enough votes exceeding some pre-definedvote_threshold, broadcasts an alert message veri-fying the obstruction point. In this article, weconsider positive detection only once an alertmessage is generated. Once a consensus isreached, alert messages can be generated to allVGrid vehicles by the method described in [14].We define the time to detection as the time atwhich the obstruction occurs until the first alarmmessage is generated. Vehicles that have notdetected the obstruction point themselves arerestricted from broadcasting an alert message.Only vehicles that have both locally detected theobstruction point and acquired enough votes arepermitted to generate an alert message.

Phase 1 (Independent Search) — Phase one of thedetection algorithm begins by periodically (every1 second) scanning only the RSC array. As pre-viously mentioned, the point of, and following,an obstruction can be classified as a low densityarea, where few or no vehicles entered. In thecase of the RSC, a search is conducted for largedrops in observed count between adjacent sec-tions in the road. For example, if road section icontains an obstruction, each vehicle scans for asignificant contrast between RSC[lane][i – 1] andRSC[lane][i]. This is done by building an orthog-onal array RSCdif, which is defined as follows:

RSCdif[lane][i] = abs(RSC[lane][i] – RSC[lane][i – 1]),

where i is the section of road, lane is the lane inthe road, and abs(x) is the absolute value of x.

Since RSCdif contains the absolute valuechange between road sections, vehicles thensearch for a large fluctuation in RSCdif. Thescope of the investigation is made more efficientby implementing a sliding search window, whichreduces the range of the search to 200 m beforethe current vehicle’s position and 100 m after.This is done to increase the efficiency of search-es by avoiding an entire scan every cycle. We canassume that few messages will be received fromoutside of the search window range and that thesearch should not be conducted in road sectionswhere the vehicle has not passed or reached (interms of transmission range). The asymmetry ofthe search range is used to limit the scope of thesearch upstream from the vehicle (200 m before)based on the relative range of the vehicles trans-mission. The downstream (100 m ahead) limit isdue to the fact that sufficient information maynot have been collected for an accurate searchsince the vehicle has not yet reached the area inquestion. In order to classify a significant changein RSCdif, we track the mean and standard devi-ation of the values in RSCdif. Changes that arethree standard deviations from the mean areflagged as significant drops in road count. Whensuch a point is detected, vehicles examine thesection in question and use the VLC and RSS toverify those anomalous points. This is donethrough analysis of the VLC to verify that lanechanges 100 m before the flagged section areoutbound, and all lane changes 100 m after are

inbound. If this is verified the RSS is then usedto check that the average speed of section i – 1is less than 20 percent of the observed averagespeed of surrounding vehicles and that averagespeed of section i + 1 is at least 80 percent ofthe observed average speed. In essence, vehiclespeeds behind the obstruction should be slowand speeds in front of the obstruction faster.Phase 1 can be summarized as follows:

RSCdif[lane][i] > mean + 3 * stdev (3)

VLC[lane][i – k] < 0 ∀ k < 100 m (4)

VLC[lane][i + k] > 0: ∀ k < 100 m (5)

RSS[lane][i – 1] < avgSpeed * .2 (6)

RSS[lane][i + 1] > avgSpeed * .8 (7)

We use Eq. 3 to locate a potential obstruc-tion point and then use Eqs. 4, 5, 6, and 7 as afiltering mechanism to filter out false positives.Figure 3 represents a snapshot of the RSC arraycollected from a particular vehicle, in a simula-


Figure 2. Figure depicts what is expected from a road obstruction. Directly afterthe obstruction point it is expected that average density would be lower thanthe rest of the lanes. Before the obstruction vehicles should be attempting toexit the lane and return after the obstruction point.

VGrid enabled vehicle

VGrid enabled vehicleNon-VGrid enabled vehicle

Non-VGrid enabled vehicle

Area of low traffic density

Road obstruction

Expected vehicle trajectory

Figure 3. Road Section Count (RSC) array for a particular vehicle collectedduring actual simulation. Penetration rate = 100 percent, traffic density = 15percent, and transmission range = 500 m (19.2 dBm, which is a transmissionpower that yields a transmission range of approximately 500 m).

Position (cells)

Road section count (RSC)

2,500 0

10

0

Cou

nt

20

30

40

50

60

70

80

5,000 7,500 10,000 12,500 15,000

Lane 1Lane 2Lane 3Lane 4


tion in which an obstruction was present. TheRSCdif generated from Fig. 3 is represented inFig. 4.

Phase 2 (Coordinated Voting) — In the secondphase, vehicles cast votes to verify an obstruc-tion point. We found that in phase 1, approxi-mately 98 percent of the votes cast wereaccurate. However, we argue that not only doesthe second phase guarantee that votes do notgenerate inaccurate alerts, but it avoids even a 1percent false positive rate, which can render anAID system unreliable [4]. It can also protectagainst misbehaving nodes, which could broad-cast bogus votes. In this article we do notexplore the effect of misbehaving nodes. Fur-ther study of the second phase of this algorithmshowed that for particular traffic situations thevoting mechanism was critical in reducing thefalse alarm rate while still ensuring reasonabledetection times. In particular, we found that avoting threshold of three was sufficient to weedout the false positives generated in the firstphase of the algorithm. Figure 5 gives a high-level overview of the detection method as exe-cuted by an individual vehicle.

SIMULATION STUDY

SIMULATION TOOL

In order to carry out s imulations we useVGSim,3 which is a simulator based on Java inSimulation Time (JiST).4 In VGSim, the net-work simulation module is based on SWANS,5a Java based network simulator. In VGSim,vehicular movements and applications aretransformed into events that are processed bythe JiST event driven platform. The networksimulator and the vehicular traffic model runon a feedback loop that enables the interac-tion between the networking simulation andthe vehicular mobility. Information obtainedfrom the SWANS network simulator is fedinto the mobility model and then based on themobility model, updated antenna positions aredetermined for the SWANS network simula-tor. The work done in [3] proposes and vali-dates a mobility model specifically for theVGSim platform. The vehicular mobility mod-ule of VGSim is a Cellular Automata (CA)model, which implements a modified versionof Nagel and Schreckenberg (N-S) model. Adetailed description of this integration of themodified N-S model and the VGSim simula-tion platform is described in [14]. The simula-tion tool also has a graphical user interface(GUI) which provide a visual interface to thesimulations. The SWANS network simulatorand vehicle mobility simulator both update agraphical interface that allows network andvehicle mobility parameters to be changeddynamically.

SIMULATION PARAMETERSVGrid vehicles are equipped with low cost pro-cessor and memory to store data and carry outcomputation on the data. In addition each VGridvehicle carries an onboard GPS unit and wirelessnetwork enabled following the IEEE 802.11bstandard. As part of our study, we vary the per-centage of VGrid vehicles and refer to this per-centage as the penetration rate.

In order to quantify our algorithm, we usethe following three common AID performancemeasurements: false alarm rate (FAR), detec-tion rate (DR), and mean time to detection(MTTD),

(8)

(9)

(10)

where tdetection is the time of the first alert mes-sage and tstart is the start time of the incident.We chose an interval of 20 s to calculate FARbecause in the majority of the studies presentedin [4], the AID algorithms used the same FARmetric.

We consider a 1.5 km stretch of highway.Incidents are injected at cell position 9000(900 m) in lane 2 at time 0. The simulation

MTTDn

t tdetection starti

n= −( )

=∑

1

1

DR =#

#

of detected incidents

total of incidents

FAR = ×#

% of false alerm cases

seconds20100

Figure 5. A flow diagram of the detection method.


Figure 4. The RSCdif generated from RSC array illustrated in Fig. 3. Penetra-tion rate = 100 percent, traffic density = 15 percent and transmission range =500 m (19.2 dBm).

Position (cells)

RSCdif Array

25000

5

0

RSC

(i) –

RSC

(i –

1)

10

15

20

25

30

35

40

45

5000 7500 10,000 12,500 15,000

Lane 1Lane 2Lane 3Lane 4

Sleep for 1 s

Phase 1 Phase 2

Insufficientvotes

No anomalydetected

No detection

Anomaly verifed by vehicle

Search for anomalies inRSCdif array (Equation 3)

Voting thresholdis reached and

alert is generated

Vehicle totals allvotes received

thus far for thislocation

Vehiclebroadcasts a

“vote”

Verify anomaly with VLCand RSS arrays

(equations 4,5,6 & 7)

Anomaly detected

3 Details of VGSim canbe found at http://wwwc-sif.cs.ucdavis.edu/VGrid

4 This tool can be foundathttp://jist.ece.cornell.edu/

5 Documentation forSWANS can be found athttp://www.aqualab.cs.northwestern.edu/projects/swans++/



ends when the first alarm is generated. Trans-mission range is based on an estimation of thecorresponding decibel level designated in thesimulation. A transmission range of 500 mrefers to the maximum transmission distanceachievable given the transmission decibel level(19.2 dBm). As a result, the average transmis-sion range is far less than the designated trans-mission range. When the transmission range isset to 500 m, the average transmission is closerto 250 m.

The traffic injection is based on the trafficdensity parameter. The road is initially populat-ed in probabilistic fashion. If the traffic densityparameter is 20 percent, enough vehicles areplaced on the road to fill 20 percent of the cellson the road. Since each vehicle occupies 75 cells(7.5 m), calculating the number of vehicles toplace on the road for a given density is a simplelinear equation. When placing the vehicles theinitial positions of each vehicle is chosen at ran-dom. The injection pattern is based on exit thefrequency of each vehicle. When a vehicle exitsthe road, a new vehicle is injected at the begin-ning of the road. If the new vehicle cannot beinjected, the vehicle is queued and injected whensufficient room is available. Vehicle are random-ly marked as VGrid vehicles upon creation (bothin the initial population of vehicles or uponinjection) with a probability based on the chosenpenetration rate.

RESULTS

EFFECTS OF TRANSMISSION RANGE

Figure 6 shows the impact of transmission rangeon the mean detection time (MTTD) for a pen-etration rate of 100 percent. Figures represent-ing the other penetrations are omitted due tospace constraints. Our study found that a trans-mission range of 250 meters was not sufficientfor a 100 percent DR in all cases. In fact, forpenetration rates less than 25 percent, DR wasalmost zero except when traffic density was veryhigh. Conversely, for higher transmissionranges, such as 1000 meters, it was possible toincrease the detection rate even with low pene-tration rates and low traffic densities. We alsoobserved some anomalous behavior in the 20 to30 percent traffic density ranges seen by theerratic values for MTTD. This behavior is aresult of traffic mobility transitioning betweenfree flow (vehicles can travel at maximum veloc-ity) and congested traffic (vehicle velocity isconstrained by downstream traffic). In thesetransition ranges the variance in velocity ofindividual vehicles is much greater across theentire road, resulting in a higher variance inMTTD. The result in Fig. 6 shows that in orderto enable detection, VGrid vehicles must forma connected grid or mesh to coordinate detec-tion. If the transmission range is too small andthe VGrid vehicles are sparse, inter-vehiclecommunication becomes sporadic if not entirelydisabled. As a result, detection of obstructions(through coordination) becomes unlikely orimpossible. Results clearly show that there is aminimum required transmission range to enabledetection.

EFFECTS OF PENETRATION RATEAND TRAFFIC DENSITY

To test the limits of our detection algorithm, wefocused on penetration rates and traffic densityparameters in ranges that were less than idealfor detection. The detection rate is a function ofpenetration rate and traffic density. In general, ifenough VGrid vehicles were present, detectionwas possible. With a penetration rate of 5 per-cent, detection was not possible for any trafficdensities. Detection became possible with a 10percent penetration rate. However, at the lower

Figure 6. The MTTD for transmissions ranges of 250, 500, 750, and 1000 mand penetration rate of 100 percent.

Traffic density

Average detection time, all TX ranges

5 0

MTT

D (

s)

10 15 20 25 30 35 40 45 50 55 60 65 70

0.600.550.500.450.400.350.300.250.200.150.100.05

Penetration 1.0 Tx tange (meters) 250Penetration 1.0 Tx tange (meters) 250Penetration 1.0 Tx tange (meters) 750Penetration 1.0 Tx tange (meters) 1000

Table 3. DRs less than 100 percent.

Penetration rate Traffic density DR

5% All 0%

10% 5% 7%

10% 10% 20%

10% 15% 77%

10% 20% 90%

20% 5% 96%

Table 4. FAR greater than 0 percent.

Penetrationrate

Trafficdensity DR Obstruction

present

50% 50% and60% 1.1% No

50% 40% 0.56% Yes

50% 50% and60% 1.1% Yes

25% 60% 0.56% Yes

75% 60% 0.13% Yes



traffic densities (5, 10, 15, 20 percent), DR wasstill less than 100 percent. Table 3 represents allpoints were DR was less than 100 percent. Forother penetration rate and traffic density valuesthe DR was 100 percent.

An important metric of any AID system is itsfalse alarm rate (FAR). If FAR is too large, thesystem become unreliable. Even a FAR of 1 per-cent, can make the system unreliable [4]. For theproposed approach, we categorize the FAR fortwo different scenarios. The first is false alarmsgenerated in a traffic scenario where an accidentexists (an alarm generated for an obstructionother than the actual obstruction). The secondcase is false alarms generated in the situationwhere no obstructions exist (control simula-tions). Table 4 shows the data for which a FARgreater than 0 percent was observed. When con-sidering the FAR for the first scenario (anobstruction is present) it is important to pointout that all false alarms were generated for loca-tions upstream from the actual obstruction. Thiswas due to the traffic congestion generated bythe downstream road blockage. In highly con-gested traffic vehicles move very slowly, which isfalsely detected as obstructions.

Figure 7 shows a 3D plot of the effects ofpenetration rate and traffic density on MTTD.Excluding the low penetration rate and trafficdensity scenarios, we found that MTTD was, ingeneral, less than 1 min. It is interesting to notethe effect of traffic density on MTTD. Wefound that the more vehicles present, the fasterthe detection time. This is clearly shown in thefigure as detection time is much slower for the5 and 10 percent traffic densities. We alsofound, however, that for the higher traffic den-sities there was also a slight increase in MTTD.The increase in MTTD for the congested trafficscenarios is due to the fact that the detectiontime is correlated with the movement of thevehicles. Our algorithm is based on each vehi-

cle tracking traffic pattern over time and space.With very high traffic congestion, vehicles moveslowly, which has the same effect as slowingdown time. When the vehicles move slowly, it ismore difficult to quickly identify areas of lowtraffic density.

Table 5 is a comparison of the proposedmethod (referred to as VGrAID) and other AIDalgorithms. When presenting our results in Table5, we choose to aggregate our findings by pene-tration rate. The average FAR for a given pene-tration rate across all traffic densities is less thanthe previously stated FAR considering both thetraffic density and the penetration rate. Theresults and values presented for other AID algo-rithms are referenced from the “Summary ofAlgorithms” table in [4] with results of VGrAIDappended at the bottom. Note that the valuesfor DR, TTD, and FAR may not have been mea-sured in the same way. The table only gives aqualitative comparison of the different algo-rithms that exist.

CONCLUSION AND FUTURE WORK

In this article we have presented a distributedAID method that leverages ad hoc networkingand computing in vehicles with storage, comput-ing, and wireless networking capabilities. Weshow that information collected by the vehicles,analyzed, and shared in a distributed mannercan improve AID detection times compared tothose of traditional systems. In addition, sincenodes are mobile and the information they gen-erate is shared, the data is more accurate andtimely than other AID systems. With the lowcost of each device, we believe that implementa-tion of such a system could potentially be morecost effective than other traditional infra-structure-based systems. Finally, the collabora-tive nature of the system also greatly reduces thefalse alarm rate and can, in the future, protectagainst malicious behavior such as a malfunc-tioning sensor or intentional disruptive behavior.

As part of the future work we will explore theconcept of categorizing incidents based on addi-tional information collected by the vehicular gridnetwork. Furthermore, in this work we categorizean obstruction as a fixed blockage in the roadway.We can expand this definition to include slowmoving or partially disabled vehicles. As an exten-sion to this work, we also plan to modify the pro-posed algorithm to adapt to a centralizedinfrastructure where roadside sensors can beleveraged in order to accumulate traffic data col-lected and dumped by moving vehicles. Thiscould greatly reduce the number of VGrid vehi-cles needed to facilitate detection as the roadsidenodes could collect and aggregate informationover a period of time. In addition to studyingvariations in which VGrid AID can be applied,we will also study the communication impact ofthis AID system and how it could impact otherconcurrent safety and infotainment applications.

ACKNOWLEDGMENTThe authors would like to thank the reviewers fortheir comments. This research was funded byNSF grant CMMI-0700383. Dr. Behrooz Kho-rashadi is now employed with Qualcomm.

Figure 7. The MTTD as a function of both penetration rate and traffic density(transmission range is set at 500 m).

Traffic density Penetration rate

3D Comparison of penetration rate and traffic density vs MTTD

0.2

100

MTT

D (

s)

150

200

250

300

350

400

450

500

50

0 0

0.4 0.6

0.8 0.8

1 0.6

0.4 0.2

0



REFERENCES[1] D. Schrank and T. Lomax, “The 2007 Urban Mobility

Report,” Texas Transportation Institute, Texas A&MUniv., Sept. 2007.

[2] D. Srinivasan, R. L. Cheu, and Y. P. Poh, “Hybrid FuzzyLogic-Genetic Algorithm Technique for AutomatedDetection of Traffic Incidents On Freeways,” Proc. IEEEIntelligent Transportation Sys. Conf., Aug. 2001.

[3] C.-N. Chuah et al., “Distributed Vehicular Traffic Controland Safety Applications with VGrid,” IEEE Wireless HiveNet. Conf., Aug. 2008.

[4] T. Martin et al., “Incident Detection Algorithm Evaluation,”prepared for Utah Dept. Transportation, Mar. 2001.

[5] H. J. Payne, E. D. Helfenbein, and H. C. Knobel, “Devel-opment and Testing of Incident Detection,” FederalHighway Administration, Research Methodology andResults, vol. 2, 1976, report no: FHWA-RD-76-20.

[6] S. R. Ahmed and A. R. Cook, “Application of Time-SeriesAnalysis Techniques to Freeway Incident Detection,” Trans-portation Research Record, 1982, pp. 19–21.

[7] S. Willsky et al., “Dynamic Model-Based Techniques forthe Detection of Incidents on Freeways,” IEEE Trans.Automatic Control, vol. 25, no. 3, 1980, pp. 347–60.

[8] N. Persaud, F. L. Hall, and L. M. Hall, “Congestion Iden-tification Aspects of the McMaster Incident DetectionAlgorithm,” Transportation Research Record, tech. rep.1287, 1990, pp. 167–75.

[9] C. Antoniades and Y. Stephanedes, “Single-Station Inci-dent Detection Algorithm (SSID) for Sparsely Instru-mented Freeway Sites,” Transportation Eng., 1996.

[10] S. Kamijo et al., “Traffic Monitoring and AccidentDetection at Intersections,” Proc. IEEE ITSC, Oct. 1999.

[11] T. Martin et al., “Video-Based Automatic IncidentDetection for Smart Roads: The Outdoor EnvironmentalChallenges Regarding False Alarms,” IEEE Trans. Intelli-gent Transportation Sys., vol. 9, no. 2, June 2008.

[12] M. Abuelela, S. Olariu, and G. Yan, “Enhancing auto-matic incident detection techniques through vehicle toinfrastructure communication,” 11th IEEE ITSC ‘08, Oct.2008, pp. 447–52.

[13] M. Abuelela and S. Olariu, “Automatic Incident Detec-tion in VANETs: A Bayesian Approach,” Proc. IEEE VTC-Spring, Barcelona, Spain, Apr. 2009.

[14] B. Khorashadi, Enabling Traffic Control and Data Dis-semination Applications with VGrid — A Vehicular AdHoc Distributed Computing Framework, Ph.D. thesis,UC Davis, 2009; http://wwwcsif.cs.ucdavis.edu/~VGrid/VGrid/Publications.html.

[15] J. McKnight and A. McKnight, “The Effect of CellularPhone Use upon Driver Attention,” Nat’.l Public ServicesResearch Inst., 1991.

BIOGRAPHIESDIPAK GHOSAL received his B.Tech. degree in electrical engi-neering from the Indian Institute of Technology, Kanpur, in1983, his M.S. degree in computer science and automationfrom the Indian Institute of Science, Bangalore, in 1985,and his Ph.D. in computer science from the University ofLouisiana in 1988. He is currently a professor in theDepartment of Computer Science at the University of Cali-fornia, Davis. His primary research interests are in the areasof high-speed networks, wireless networks, vehicular adhoc networks, next-generation transport protocols, andparallel and distributed computing.

CHEN-NEE CHUAH is currently a professor in the Electrical andComputer Engineering Department at the University of Cali-fornia, Davis. She received her B.S. in electrical engineeringfrom Rutgers University, and her M. S. and Ph.D. in electricalengineering and computer sciences from the University ofCalifornia, Berkeley. Her research interests lie in the area ofcomputer networks and wireless/mobile computing, withemphasis on Internet measurements, network anomaly detec-tion, network management, multimedia, online social net-works, and vehicular ad hoc networks. She received the NSFCAREER Award in 2003 and the Outstanding Junior FacultyAward from the UC Davis College of Engineering in 2004. In2008 she was selected as a Chancellor’s Fellow of UC Davis.She has served on the executive/technical program commit-tee of several ACM and IEEE conferences, and is currently anAssociate Editor for IEEE/ACM Transactions on Networking.

MICHAEL ZHANG is currently a professor in the Civil and Envi-ronmental Engineering Department at the University of Cal-ifornia, Davis. His research is in traffic operations andcontrol, transportation network analysis, and intelligent

transportation systems. He received his B.S. degree in civilengineering from Tongji University, and M.S. and Ph.D.degrees in engineering from the University of California,Irvine. He is an Area Editor of the journal Network andSpatial Economics and an Associate Editor of Transporta-tion Research — Part B: Methodological.

BOJIN LIU ([email protected]) is a Ph.D. student in theComputer Science Department, University of California,Davis. He received his Bachelor’s degree in computing fromHong Kong Polytechnic University. His research interestsinclude vehicular ad hoc networks, wireless networks, andparallel and distributed systems.

BEHROOZ KHORASHADI received his Bachelor’s degree from theUniversity of California, Berkeley in 2004. He is a Ph.D.graduate from the Department of Computer Science at theUniversity of California, Davis in 2009 and is currentlyworking at Qualcomm’s Bay Area Research and Develop-ment (BARD) facility, Santa Clara, California. His researchinterests include vehicular ad hoc networks, wireless net-works, parallel and distributed systems, and network pro-tocol optimization. His current work at Qualcomm includesprojects dealing with indoor location services on mobiledevices and multicore systems on mobile devices.

Table 5. Comparison of VGrAID with other AID algorithms.

Name DR (%) TTD (min) FAR (%)

APID 86 2.50 0.05%

DES 92 0.70 1.87%

ARIMA 100 0.40 1.50%

Bayesian 100 3.90 0%

California 82 0.85 1.73%

Low-pass filter 80 4.00 0.30%

McMaster 68 2.20 0.0018%

Neural networks 89 0.96 0.012%

SND 92 1.10 1.30%

SSID 100 not reported 0.20%

TSC 7 67 2.91 0.134%

TSC 8 68 3.04 0.177%

Video image processing 90 0.37 3.00%

Cell phones 100 — 5.00%

VGrAID 10% penetration 74% 3.58 0.0%

VGrAID 20% penetration 99.6% 1.29 0.0%


VGrAID 50% penetration 100% .45 0.16%



Note: This table provides a high level qualitative comparison of the DR, TTD,and FAR metrics. These variations are the result of the various measurementmethodologies used in each study.



A

ISP3


INTRODUCTIONWireless mesh networks (WMNs) have been thesubject of much discussion in recent years for thepromising improvements they bring to manyproblems of ad hoc networks, as well as theirpracticality and effectiveness in areas of difficultterrain. A vast amount of research has been per-formed in different areas of WMNs, such as net-work performance, routing, and applications.However, most research work does not identifythe type of architecture design in which theywork and fail to address whether their solutionscould generate the same results when implement-ed for different types of WMN architectures.

WMN architecture could take different topolo-gies based on structural design and orientation of

its components with respect to each other and tothe network environment. Such variations in thearchitecture of WMNs could pose fundamentaldifferences in the physical characteristics and per-formance of the network, which in turn could yielddifferent results in the performed studies. Suchdifferences necessitate architecture-dependentsolutions for each problem. Thus, an in-depthstudy of different types of architectures is impor-tant and critical for WMNs, and could potentiallyelucidate some of the challenges facing WMNs inresearch areas such as applications, routing, net-work management, and network performance.

In this article we identify three mainstreamarchitectures for WMNs: campus mesh (CM),downtown mesh (DTM), and long-haul mesh(LHM), and explain their fundamental differ-ences as well as some important factors thatcould be affected by the type of architecture. Wethen discuss how the dynamics of routing, net-work management, and performance of WMNscould be affected by the characteristics of thearchitecture, without getting into detailed analy-sis or evaluation of those factors.

The remainder of this article is organized asfollows. In the next section we give a brief litera-ture review. We then identify three distinct typesof topological architectures. We then outlinemajor network characteristics that could beaffected by the type of the deployed architec-ture, and show performance variations using sys-tem throughput for different architectures. Weconclude our findings and make recommenda-tions to be considered in WMN research studiesin the final section.

RELATED WORKS

The IEEE 802.11s is associated with WMNs, anddefines the interoperability between WMNs andwireless ad hoc networks. The IEEE 802.11s medi-um access control (MAC) layer defines architec-ture for WMNs including protocols to supportbroadcast, multicast, and unicast connections for aself-configuring multihop network combining abackbone of fixed wireless mesh routers (WMRs)and clusters of mobile nodes (MNs) [1].

To the best of our knowledge, few researchpapers exist in the literature clearly addressing

AMIR ESMAILPOUR AND NIDAL NASSER, UNIVERSITY OF GUELPHTARIK TALEB, NEC EUROPE LTD.

ABSTRACTThe wireless mesh network and the associat-

ed IEEE 802.11s standard have attracted anenormous amount of research in the wirelessresearch community in the past few years. Nev-ertheless, WMN architecture has not receivedmuch needed attention compared to other topicsin this area of research. Based on topologicaldifferences, various network architectures arepossible for WMNs, and we believe such archi-tectures could affect wireless characteristics dif-ferently. In this article we provide an overview ofarchitectural design approaches for WMNs, thensummarize the state-of-the-art research findingsand suggest further topics that need to beaddressed. Additionally, we identify three differ-ent types of architectures for WMNs: campusmesh, downtown mesh, and long-haul mesh.Furthermore, we discuss and investigate differ-ent WMN characteristics that could be affectedby commonly deployed architectures. Among theconsidered characteristics we select routing, net-work management, and network performancefor further analysis, and look at the challengesthese architectures face with respect to thosecharacteristics. To illustrate these challenges, weperform a simple experiment to show that LHMand DTM under identical network environmentsshow significant differences in performanceparameters such as throughput and delay.

TOPOLOGICAL-BASED ARCHITECTURES FORWIRELESS MESH NETWORKS

The authors providean overview of architectural designapproaches forWMN, then summa-rize the state-of-the-art research findingsand suggest furthertopics that need tobe addressed. Additionally, we identify three differ-ent types of architec-tures for WMN:campus mesh, down-town mesh, andlong haul mesh.

ESMAILPOUR LAYOUT 2/7/11 10:34 AM Page 74


the importance of the architecture of WMNs.Akyldiz et al. [2] identify, for the first time, threedistinct architectures for WMN. They classifythem as infrastructure/backbone, client, andhybrid WMNs. Their classification is based on thefunctionality of WMRs vs. MNs. Nodes in thebackbone architecture have different functionalitythan those in the client architecture. In infra-structure WMNs, the authors introduce WMRs inthe backbone, connecting to MNs and collectingtraffic from them. Then WMRs forward the traf-fic to the Internet backbone via access points. Inclient WMNs the authors introduce peer-to-peerconnectivity of MNs and totally remove the needfor WMRs. The hybrid WMN is basically a com-bination of the other two architectures.

Waharte and Boutaba proposed a tree-basedarchitecture for WMNs [3]. They identify twotypes of WMRs, access points (APs) and networkgateways (NGs). In their tree-based architecture,APs collect traffic from MNs and pass it to NGsin a two-layer hierarchical fashion, where MNsconnect to the Internet through NGs rather thanexchanging peer-to-peer traffic. Therefore, trafficstreams will mainly be directed toward/from theNGs, causing a bottleneck in the access network.Based on these findings, the authors concludethat a tree-based or hierarchical architecturewould suit WMNs most perfectly.

Seamless mesh (sMesh) architecture is intro-duced in [4]. The authors propose an architec-ture entirely based on the backbone and WMRs.It is also totally transparent to MNs. This meansthe entire WMN is seen as a single AP from thepoint of view of MNs. They add a connectivitymonitoring system to the functionality of WMRsthat constantly monitors connectivity power foreach MN, and finds the strongest connected APto which the MN will communicate. In this archi-tecture a higher density of WMRs in the back-bone ensures that all MNs always have goodconnections to the WMN.

Recently, BelAir Networks has taken a newapproach to categorizing different architecturesof WMNs [5]. The authors distinguish differentstructures based on the number of radios used onWMN nodes. They identify three types of archi-tectures: single-radio, dual-radio, and multiradiomesh networks. Each structure includes a stringof APs connecting several clusters of MNs. Inthis type of WMN categorization, a single-radiowireless mesh has low capacity and does noteffectively scale to implement a complete net-work solution, whereas a dual-radio mesh archi-tecture could scale to a metro dimension, sinceusing different radios, it separates traffic andreduces interference to improve capacity. On theother hand, a multiradio mesh system separateswireless access and backbone by using a dedicat-ed point-to-point link to form a wireless back-bone. This provides for a high-capacity systemthat can support large networks with wirelessbroadband service for the end user.

WMNs can be implemented using differenttypes of wireless technologies such as wirelesslocal area network (WLAN) or (WiMAX, or cel-lular technologies such as Universal MobileTelecommunications System (UMTS) or LongTerm Evolution (LTE). In recent years, WiMAXnetworks and a new generation of cellular net-

works such as LTE have been proposed as thebackhaul for the WLAN to build a new generationof WMNs that integrate various wireless technolo-gies to fulfill one of the promises of fourth-genera-tion (4G) wireless technologies [6, 7].

Wireless technology vendors have recentlyaddressed a new architecture for WMN calledMetroMesh, which covers areas on a metropolitanscale [8]. However, they have not addressed otherpossible architectures to include smaller environ-ments such as campuses or larger environmentssuch as long-haul, and stopped short of character-izing WMNs based on different architectures.

DESCRIPTION OF THEWMN ARCHITECTURES

WMN is a self-configured, self-organizing, multi-path, and multi-hop network, which consists of,fixed WMRs in the backbone and MNs in theaccess network. In this section, we first introducethe general or baseline architecture for WMNincluding all equipment, interfaces, links, and pro-tocols. Then we describe how this architecture isimplemented in three completely different oper-ating environments, namely campus, downtown,and long haul mesh. All architectures follow thegeneral structure outlined in the baseline. Howev-er, they differ on how their backbone and accessnetworks are topologically designed and connect-ed to each other. These architectures could becombined in order to build more complex hybridstructures, and could be customized to fulfill spe-cific requirements set by clients.

WIRELESS MESH NETWORKBASELINE ARCHITECTURE

The WMN baseline architecture is based on theconnectivity and physical orientation of differenttypes of WMN nodes. There are essentially twotypes of nodes, WMRs and MNs, as illustratedin Fig. 1. WMRs can be classified into threetypes: backbone mesh router (BMR), accessmesh router (AMR), and Internet access point(IAP) depending on in which part of the meshnetwork they are located. Each MN is connectedthrough an access link to an AMR, which servesas a gateway to the backbone network. TheBMR is in the core of the mesh network anddoes not have access functionality, nor does ithave any MNs connecting to it. IAPs serve asgateways to the Internet for the entire WMN.All WMRs have gateway/bridging functionality,which is not required for MNs.

Although other technologies such as WiMAXand UMTS have been proposed for the back-bone of WMNs, in this study all assumptions areprimarily of WMNs based on the IEEE 802.11WLAN with a/b/g/s amendments. The backboneinterfaces are equipped with 802.11a (in theinfrastructure mode of operation), access inter-faces with 802.11b/g (in the ad hoc mode ofoperation), and the mesh definitions are basedon 802.11s. A general overview of WMNs is pre-sented in Fig. 1.

Generally, a WMN has two distinct parts: abackbone network and an access network. Thebackbone network is a collection of various types

All architectures fol-low the general

structure outlined inthe baseline. Howev-

er, they differ onhow their backbone

and access networksare topologically

designed and con-nected to each other.

These architecturescould be combined

in order to buildmore complex hybrid

structures.



of WMRs connected to each other. The accessnetwork, on the other hand, is a collection ofseveral clusters of MNs. Each AMR acts as aclusterhead for its corresponding cluster of MNs,collects traffic from MNs and forwards it toother AMRs or BMRs for uplink delivery to theInternet via IAPs. The backbone is a group ofdifferent types of WMRs organized in circular,ad hoc, or longitudinal fashion depending onwhether the architecture is CM, DTM, or LHM,respectively. WMRs are often fixed and stable,and have access to unlimited power supply. Theyuse proactive routing protocols such as OpenShortest Path First (OSPF). The access networkis a group of clusters, each containing severalMNs. These clusters are highly mobile and unsta-ble, use temporary sources of power, and use on-demand and ad hoc routing protocols such asAd Hoc On-Demand Distance Vector (AODV).

All the links in the backbone and access net-works are wireless, established using 802.11a and802.11b/g respectively. AMRs are equipped withtwo physical interfaces; one to connect otherWMRs in the backbone, and the other to con-nect to MNs in the access network. All WMRsin the backbone use multiple virtual interfaces toconnect to multiple peer WMRs to build a par-tial mesh as depicted in Fig. 1. MNs use802.11b/g contention MAC to access sharedchannels and connect to their clusterhead AMR.

The baseline details mentioned here areshared by the three identified architectures inthe next section. The difference is in the topolo-gy; the geographical location and physical orien-tation of the equipment with respect to oneanother and to the network environment. In the

next section we identify the fundamental differ-ences in each type of architecture.

CAMPUS MESH ARCHITECTUREIn the CM architecture (Fig. 2), a limited num-ber of buildings are located in a campus environ-ment, with generally good line of site (LOS),and a central management and administrationunit. WMRs could simply be installed on existinginfrastructure in campus. The number of MNs insuch environments is usually fixed, and MNshave little or no mobility, since the wirelessequipment in a campus environment has little orno movement once they are stationed in a loca-tion. The entire network is usually under a singleadministration and is controlled by a singleInternet service provider (ISP). Traffic can easilybe monitored, and the amount of exchangedtraffic can easily be predicted during differenttime periods, resulting in a more static and pre-dictable network requirement. Thus, the networkin a CM architecture is generally easy to deploy,monitor, manage, and upgrade.

These features provide a highly flexible envi-ronment for deployment of a WMN. Due to sin-gle administration provisioning, it is easy tomonitor and control different aspects of networkmanagement, such as routing, congestion, andinterference control. CM is the most flexibleenvironment of the three architectures. It is alsothe simplest architecture to deploy.

Figure 2 shows a typical CM architecture,where there are two rings of WMRs in the back-bone: inner and outer rings. The outer ring essen-tially represents AMRs connecting MNs to thebackbone. The BMRs are in the inner ring, with

Figure 1. Wireless mesh network baseline architecture.

L3: Internet gateway network infrastructure

L2: Backbone network partial mesh

L1: Access networkad hoc

Wired links

Wireless links 802.11a Routing: OSPF

Wireless links 802.11b Routing: AODV

WMR=Wireless mesh routher (BMR, AMR, IAP) with 2 physical and several virtual interfaces

: Wired link : Wireless link

S_MN: a Source MN D_MN: a Destination MN MN=Wireless mobile node with 1 interface

MN

MN

MN

MN MN

MN

S_MN

D_MN

AMR

Public internet

IAP IAP

BMR

BMR BMR

AMR AMR

Generally, a WMNhas two distinctparts: a backbonenetwork and anaccess network. Thebackbone network isa collection of vari-ous types of WMRsconnected to eachother. The accessnetwork, on theother hand, is a col-lection of severalclusters of MNs.



no direct connection to the access network. Dueto high concentration of MNs in a CM, some ofthe inner ring BMRs could also act as AMRs.The inner ring has several major functions; to actas a redundant array of routers or a backup pathin case of congestion or disconnection, to providemultipath routing options to the AMRs, and tocollect traffic and pass it to Internet access points(IAPs) for Internet connectivity.

DOWNTOWN MESH ARCHITECTUREIn the DTM architecture, many buildings rangingfrom small to large are scattered over severalblocks in a downtown environment, as shown inFig. 3. This type of architecture introduces manychallenges in terms of deployment, management,and control. Generally, LOS is not adequate, andtowers are not available or accessible in manylocations. The number of MNs varies with time,and MNs tend to change their locations frequent-ly around the downtown area. The network couldbe under different administration and manage-ment, or even different ISPs, which introducesmore technical and billing difficulties, such asroaming and network sharing among ISPs.

In terms of traffic load and prediction of traf-fic behavior, this type of architecture is quite dif-ferent from the CM. In a CM, the majority oftraffic is generated by the users on the campus,such as employees in an enterprise or students ina university campus. The number of employees,and the type of operation, application, and usageare well known to the administration over time.

DTM, on the other hand is usually the harsh-est environment, where one does not know whatto expect in terms of real-time traffic. The num-ber of users passing by, the type of traffic they areusing, the time and day they are passing by, andother factors could very well change the fluctua-tions of the amount of traffic. Exchanged traffic ishighly bursty depending on different client opera-tions, and different times of day, week, month, oreven year. Different ISPs provide different typesof services to their clients, which makes itextremely difficult for them to coordinate withone other. The complication in management andbilling coordination could increase significantly.

DTM thus requires a more advanced high-capacity network and costly equipment for deploy-ment. Coordination between ISPs is required, andconstant city involvement and licensing issuesshould be taken into consideration. Such techni-cal and management difficulties could make thesolution not as viable as originally thought.

LONG HAUL MESH ARCHITECTUREIn the LHM there are no buildings around, butrather a long set of WMRs along a stretch of high-way inside a city or in suburban areas, where thereis no infrastructure in place, or it is difficult andcostly to deploy one. The WMRs could be as farapart as their transmission range allows. Lack ofabstraction allows for long LOS using single pow-erful unidirectional antennas between each pair ofadjacent routers. Deployment could prove simple,where antennas are positioned at great heights,kilometers away from each other, depending ontheir transmission power. A second set of routers(BMRs) could be deployed on the other side ofthe roadway for redundancy, as depicted in Fig. 4.

LHM is a phenomenal solution for network-ing and communications, because it eliminatesthe need for extensive and costly infrastructure,as required by traditional wired and wirelesstechnologies. In [9] the authors propose LHMarchitecture for the first time, along with a rout-ing scheme that involves OSPF and BorderGateway Protocol (BGP) routing protocols inthe backbone and AODV with an alternativerouting path through the access network.

It is generally agreed that multipath routingis a viable solution for WMN routing. On thecontrary, with multipath routing, there is notmuch gain for LHM architecture, since there arenot many possible paths between a pair of sourceand destination nodes. The only backup solutionis an array of redundant BMRs that runs alongwith the main backbone of AMRs.

DIFFERENCES BETWEENCM, DTM, AND LHM ARCHITECTURES

The three identified architectures are fundamental-ly different in terms of technical details as well asmanagement. Each type of architecture possessesadvantages and pitfalls. There are numerous studiesthat evaluate the performance of various WMNs intheir respective environment and unique architec-ture [2]. Performance of each type of architecture isadjusted and optimized to its specific characteristicsand environment. For instance, physical and MAClayer characteristics could be adjusted or links withdifferent capacities could be used to improve per-formance of one type vs. another. In this article,however, the main objective is to highlight the dif-ferences and their consequences without gettinginto details of performance analysis. In the next sec-tion we create a simple case to highlight one of theareas that shows clear differences among architec-

Figure 2. Campus mesh architecture.

IAP

AMR

AMR

AMR

AMR

BMR

BMR

BMR

Cluster of MNs

Inner ring of BMRs in the backbone

used for redundancy Outer ring of AMRs in the access network

Public internet



tures. Differences between CM and DTM fall intotwo major categories:• Differences between the two in ISP manage-

ment problems• Differences in network issues such as routing,

congestion, and interference due to clusterdensity of MNsThe major difference between LHM and the

other two architectures consists of the structureitself, the type of equipment, and the lack ofmultipath routing in LHM. Table 1 summarizesthe basic characteristics and differences of thethree types of architecture.

PRETENTIOUS NETWORK CHARACTERISTICS

In this section we present some factors or net-work characteristics that could be seriouslyaffected and severely constrained by the imple-mented architecture in WMNs. We explain themost important factors that have direct effectson the WMN when a particular architecture isused. However, a long list of such factors can bepointed out based on applications and solutionsproposed for different areas of WMNs.

ROUTING IN WMNSSeveral routing protocols are proposed forWMNs such as AODV, Dynamic Source Routing(DSR), and OSPF. OSPF is proposed and

deployed in the backbone by many researchgroups and vendors, such as Nortel Networks.However, a difficulty with implementing OSPF inthe backbone lies in the fact that OSPF works ina hierarchical fashion. This structure includes abackbone area (i.e., area0) in the root and sever-al other areas all connected through area0.

Hierarchical OSPF areas would nicely fit intothe DTM architecture where the main buildingsare surrounded by smaller offices, and to a lesserextent for the CM architecture. However, in theLHM architecture it would be impractical todeploy OSPF in a hierarchical fashion. This isdue to the fact that traffic from all other areasneeds to go through area0 before reaching itsdestination. This causes an enormous amount oftraffic to pass through area0, which results in anarea0 bottleneck.

Multipath routing in the backbone is intro-duced to improve the quality and performanceof WMN routing. Several multipath routingprotocols have been proposed for WMNs alongwith their extensions as well as new metrics forperformance measurement [10]. Multipathrouting could be applied in DTM or CM archi-tectures. However, in an LHM architecturewhere the WMRs are stretched longitudinallyalong hundreds of kilometers of highways,LHM could not gain much by using multipathrouting.

Figure 3. Downtown mesh architecture.

BMR BMR

BMR

BMR

BMR

BMR

BMRBMRAMR

AMRAMR

AMR

AMR

AMR

IAP

IAP

ISP5

ISP4

ISP3

ISP1

IAP

ISP2

BMR

Public internet

Public internet

Public internet

With multipath rout-ing, there is notmuch gain for LHMarchitecture, sincethere are not manypossible pathsbetween a pair ofsource and destina-tion nodes. The onlybackup solution is anarray of redundantBMRs that runsalong with the mainbackbone of AMRs.



NETWORK MANAGEMENT

One of the issues facing WMNs is network man-agement and network ownership by multiple ISPs.As the size of a network increases, more than asingle ISP could get involved in managing theentire network [2, 11]. Traditionally, each WLANis managed by a single ISP. However, for largernetworks that could range over hundreds of kilo-meters or hundreds of buildings in a downtownarea, different parts of the network most likely fallinto territories of different ISPs, and could posecomplications in management such as roaming,billing and handoff between different networks.Although ISP management could introduce seri-ous issues in LHM and DTM architectures, itdoes not pose any problem in a CM architecture.In a CM architecture where most of the networkmanagement is handled by a single ISP or evenoccasionally handled locally by in-house networkadministration, there will be no need for ISP ormanagement coordination and consideration.Therefore, individual solutions could be developedfor CM problems that do not concern manage-ment issues similar to those of DTM or LHM.

Using multiple ISPs in large networks hasalso been proposed for load balancing as well asother network management issues. In wired net-works, BGP is the protocol of choice for net-work management and employment of multipleISP solutions. WMN architectures such as LHMand DTM could also use BGP in the backbone,and provide a viable solution and replacementfor the last-mile networks.

NETWORK PERFORMANCENetwork performance for WMNs has been thesubject of much debate in the wireless researchcommunity in the past few years. Many studies

have introduced new and improved solutionsthroughout various TCP/IP layers to optimizenetwork performance for WMNs [2]. Othersstudied performance issues that come frominterference. Usually, MNs in a cluster engage incommunication with AMRs and with other MNs,causing multiple levels of interference. Interfer-ence could be a major obstacle in a DTM net-work where backbone routers are closer to eachother and ad hoc MNs are moving. On the otherhand, when LHM architecture is deployed in asuburban area, there are only few WMRs, andthey are deployed far apart. In this kind of struc-ture, interference is minimal and does not affectthe functionality of other nodes and routers. In[9] the authors show that the performance degra-dation is caused by contention among MNs.

In this section we evaluate and compare theperformance of various architectures. We per-form throughput and delay measurements tohighlight differences among the different archi-tectures. One can find more detailed perfor-mance analysis for WMNs in the literature [2, 9].Our throughput hypothesis states that perfor-mance degradation is due to reduction inthroughput caused by contention among MNs formedium access on the link to the AMR, and sucha performance measure is highly affected by theorientation of the AMRs and MNs, and the typeof architecture in the WMNs. Therefore, it isexpected that throughput values for differentarchitectures show significant differences.

A simulation model and experiments wereimplemented and carried out in the OPNETmodeler 14.5 PL1. We implemented CM, LHM,and DTM architectures with the same networkenvironment such as amount of equipment,applications, and traffic. In each model there are12 routers and six clusters with two MNs in each

Figure 4. Long haul mesh architecture.

Long range point-to-point directional antenna

Redundant array of mesh routers

(BMRs)

IAP IAP

BMR BMR

AMR AMR

Clusters of MNs

Public internet

Long haul array of access routers

(AMRs)

Table 1. Basic characteristics and differences of the identified three architectures.

Architecture Area (km2) Management ISP Line of site Users

Campus 1 Enterprise Single Good 100s

Downtown 10 City Single/multiple Inadequate 1000s

Long-haul 1000 State Multiple ISPs Excellent Unlimited

Hierarchical OSPFareas would nicely

fit into the DTMarchitecture wherethe main buildingsare surrounded by

smaller offices, andto a lesser extent forthe CM architecture.

However, in the LHMarchitecture, it would

be impractical todeploy OSPF in a

hierarchical fashion.



cluster. OPNET has different options for gener-ating traffic at various layers with unique specifi-cations, such as mobile ad hoc network(MANET) traffic. MANET traffic is generatedbetween a pair of source and destination MNs.

In each simulation experiment we graduallyincrease the number of MNs in the source clusterfrom two to six and monitor the changes in thesystem throughput values. Figure 5a shows thechanges in throughput for LHM architecture. Asthe number of MNs increases from two to four,there is a twofold increase in the system through-put. However, by further increasing MNs to six,the contention problem causes the throughput todecrease to just above that of two MNs.

Figure 5b shows corresponding results forDTM architecture. These results are significantlydifferent from those of LHM. As we increasethe number of MNs from two to four, thethroughput increases by 25 percent. As weincrease the number of MNs to six, the through-put increases at a steady rate by another 25 per-cent. This shows that the contention is notaffecting the network performance as severely asin the LHM architecture. This linear perfor-mance change can be attributed to several rea-sons. Intuitively, in a dense DTM environmentwhen an MN has many contending neighbors, itseeks other paths to reach the destination, there-by increasing the system throughput at a morelinear way than the LHM Architecture. Similarresults were achieved when we tried the linkthroughput between the D_AMR and D_MN.

In terms of delay analysis, as illustrated inFig. 6, in both cases the delay is the least for thescenario with two MNs. As we move to fourMNs, the delay increases for DTM at a slow rateby less than 5 percent for four MNs and lessthan 8 percent for six MNs. However, in case of

LHM, the delay follows a slightly higher increaseto over 10 percent. However, as we move to 6MNs, the delay increases dramatically to over 70percent compared to the case of two MNs.

The results show that there is no proportionalrelation between the number of MNs and thesystem throughput or delay in LHM comparedto DTM. Clearly, contention exists in both cases.However, its effect becomes insignificant whenother architectural features as well as other con-straints are included in the equation.

In this article we are not trying to generalizethe results or declare which type of architecture isbest. We merely highlight the impact of thesearchitectures on network performance. Regardlessof the reasons, the results support the hypothesisthat different types of architectures generate sig-nificantly different results in performance analysisand measurements. Furthermore, they show thatdifferences are so significant that proof of a pointin one type of architecture could not support thesame point in another type of architecture.

CONCLUSION AND RECOMMENDATIONS

Recent research on WMNs has not adequatelyaddressed various WMN architectures. A WMNcan assume different types of network architectures,and the type of architecture can affect wirelesscharacteristics differently. In this article we identifythree types of architectures and various networkcharacteristics that can be differently affected byeach type. We point out three major areas in whichthe differences are highlighted among the threearchitectures, and show by simulations that perfor-mance measures such as throughput and delaycould vary significantly depending on the underly-ing architecture. We recommend that WMN archi-tecture should be considered as an integral part of

Figure 5. Overall system throughput for architectures: a) LHM; b) DTM.

Min

0.0

(a)

b/s(x106)

0.5

0.0

1.0

1.5

2.0

2.5

3.0

1.0 2.0 3.0

4 MNs

6 MNs

2 MNs

4.0Min

0.0

(b)

1.0 2.0 3.0

6 MNs

4 MNs

2 MNs

4.0



research activity, and that experiments in this areaclearly distinguish and identify the scope of the net-work as well as its type of architecture. Varioussolutions in different areas of WMN have beenproposed in recent years, and experiments havebeen developed to prove or disprove proposalsbased on single architecture types. We further rec-ommend investigating the validity of such proposalsunder various types of architectures.

In the future, we plan to propose standardstructural definitions for the identified architec-tures and other hybrid architectures that could bebuilt based on the three mainstream architecturesmentioned in this article. We will also identify acomplete list of factors at different layers (e.g.,application, network, and MAC layers) that couldbe affected by the architecture type. We wouldfurther like to investigate the accuracy of severalproposed solutions based on different architec-tures, and investigate their reliability to see if theyapply to all kinds of architectures or only one.

REFERENCES[1] IEEE 802.11s, “Amendment for Mesh Networking to the

IEEE802.11”; http://ieeexplore.ieee.org/xpl/standards.jsp.[2] I. F. Akyildiz, X. Wang, and W. Wang, “Wireless Mesh

Networks: A Survey,” Comp. Net., vol. 47, no. 4, 2005,pp. 445–87.

[3] S. Waharte and R. Boutaba, “Tree-Based Wireless MeshNetwork Architecture: Topology Analysis,” Proc. 1stInt’l. Wksp. Wireless Mesh Net., Budapest, Hungary,July 2005.

[4] Y. Amir et al., “Fast Handoff for Seamless WirelessMesh Networks,” Proc. 4th MobiSys, Uppsala, Sweden,2006, pp. 83–95.

[5] BelAir Networks, “Capacity of Wireless Mesh Networks,”White Papers, 2008; http : / /belairnetworks.com/resources/.

[6] M. Cesana, “The Evolution toward Heterogeneous HighSpeed Mesh Networking,” ICT Mobile SUMMIT-CAR-MEN Project Wksp. Carrier Grade Mesh Net., San-tander, Spain, June 2009.

[7] N. Bayer et al., “Towards Carrier Grade Wireless MeshNetworks for Broadband Access.” German Ministry of

Education and Research, 2009; http://www.deutsche-telekom-laboratories.de/~karrer/papers/opcomm06.pdf.

[8] StrixSystems, “WiFi Mesh Business Case TechnicalResearch Paper,” 2009; http://www.strixsystems.com/case-studies/WiFi-Mesh-business-case.asp.

[9] A. Esmailpour et al., “Ad-Hoc Path: An Alternative toBackbone for Wireless Mesh Networks,” Proc. IEEE ICC,Glasgow, U.K., June 2007, pp. 3752–57.

[10] N. S. Nandiraju, D. S. Nandiraju, and D. P. Agrawal,“Multipath Routing in Wireless Mesh Networks,” Proc.IEEE MASS, Oct. 2006, pp. 741–46.

[11] B. V. Daggett, “Localizing the Internet: Five Ways Pub-lic Ownership Solves the U.S. Broadband Problem,”Inst. Local Self-Reliance, tech. rep., Jan. 2007.

BIOGRAPHIESAMIR ESMAILPOUR ([email protected]) is currently a Ph.D.candidate at the University of Guelph, Canada. He receivedhis Bachelor of Science at the University of Ottawa and Mas-ter’s of Applied Science from Ryerson University, Toronto,Canada. He worked at Nortel Networks as a software engi-neer and Daimler Chrysler as a network engineer for sevenyears, and returned to academic studies and research to pur-sue his Ph.D. degree. His area of research is in wireless meshnetworks and quality of service for the IEEE 802.16 stan-dard. He is presently working on his Ph.D. thesis in radioresource management and QoS for mobile WiMAX.

NIDAL NASSER ([email protected]) completed his Ph.D.in the School of Computing at Queen’s University,Kingston, Ontario, Canada, in 2004. He is currently anassociate professor in the Department of Computing andInformation Science at the University of Guelph. He is anassociate editor of the Journal of Computer Systems, Net-works, and Communications, Wiley’s International Journalof Wireless Communications and Mobile Computing, andWiley’s Security and Communication Networks Journal.

TARIK TALEB [S‘04, M‘05, SM‘10] ([email protected])is currently working as a senior researcher at NEC EuropeLtd, Heidelberg, Germany. Prior to his current position untilMarch 2009, he worked as am assistant professor at theGraduate School of Information Sciences, Tohoku Universi-ty, Japan. From October 2005 to March 2006 he worked asa research fellow with the Intelligent Cosmos ResearchInstitute, Sendai, Japan. He received his B.E. degree ininformation engineering with distinction, and M.Sc. andPh.D. degrees in information sciences from GSIS, TohokuUniversity, in 2001, 2003, and 2005, respectively.

Figure 6. End-to-end delay results for a)LHM; b) DTM architectures.

Min0.0

(a)

Seconds

0.0

0.8

1.0

1.0 2.0 3.0

4 MNs

6 MNs

2 MNs

4.0Min

0.0

(b)

0.6

0.4

0.2

1.0 2.0 3.0

6 MNs 4 MNs 2 MNs

4.05.0 5.0



N4

N8

N11 N12

N16 N15

2

6

0

N3

N7


INTRODUCTIONTime synchronization entails an important func-tion in wireless sensor networks (WSNs), andthis function can be performed at different lay-ers depending on the objective of synchroniza-tion. For instance, sharp timing is fundamentalat low layers as it helps increase data rates (shortbit times), and enhance noise immunity (fre-quency hopping) and time-division multipleaccess (TDMA)-based scheduling. Furthermore,since radio is usually the most energy-consumingpart of a node, keeping the nodes awake theminimum required time to exchange data is acommon practice that requires synchronization.

Synchronization is also of great interest athigher layers. Akyildiz et al. identify in the appli-cation layer the sensor management protocol,which includes (among other administrative

tasks) time synchronization [1]. From a practicalpoint of view, WSNs are networks composed ofa large number of small devices that take mea-surements, process them, and communicate withother devices coordinating their operations. Thiscollaboration enables a complex sensing task,named data fusion [2]. Data fusion requires syn-chronization for two tasks: time scheduling andtimestamping. The first is needed when thenodes coordinate to perform cooperative com-munications. The second is commonly usedwhen data is fused taking into account the col-lecting instant; for example, to perform eventdetection, tracking, reconstruction of a system’sstate for control algorithms, and offline analysis.

In this article we describe in detail and evalu-ate the use in WSNs of the Multihop BroadcastSynchronization (MBS) protocol. An early ver-sion of this scheme was first introduced by us in[3]. The proposed protocol is very well suited forWSNs and fills an existing gap when synchroniz-ing WSNs with a global network time. As seen inthe following sections, MBS helps to achievehigh accuracy and energy efficiency when time-stamping at lower layers is not possible in multi-hop networks.

RELATED WORK

Creating a common temporal reference usingthe nodes communication capabilities has beenwidely studied [2]. According to the strategy, wecould distinguish between a posteriori and a pri-ori synchronization [4]. A posteriori methodskeep devices’ clocks running free, gatheringinformation between relative clocks and rear-ranging timestamps once the measurement pro-cesses are finished. These methods are usuallythe most energy-efficient because they optimizethe number of messages exchanged, but they donot offer real-time capabilities. On the otherhand, a priori methods overcome this by syn-chronizing all the nodes with a common timereference (global network time [GNT]) usingregular clock corrections. A common drawbackof these techniques is overload of the network

ÁLVARO MARCO AND ROBERTO CASAS, UNIVERSITY OF ZARAGOZAJOSÉ LUIS SEVILLANO RAMOS, UNIVERSITY OF SEVILLE

VICTORIA’N COARASA AND A’NGEL ASENSIO, UNIVERSITY OF ZARAGOZAMOHAMMAD S. OBAIDAT, MONMOUTH UNIVERSITY

ABSTRACTTime synchronization is a key issue in wire-

less sensor networks; timestamping collecteddata, tasks scheduling, and efficient communica-tions are just some applications. From all theexisting techniques to achieve synchronization,those based on precisely time-stamping syncmessages are the most accurate. However, work-ing with standard protocols such as Bluetooth orZigBee usually prevents the user from accessinglower layers and consequently reduces accuracy.A receiver-to-receiver schema improves time-stamping performance because it eliminates thelargest non-deterministic error at the sender’sside: the medium access time. Nevertheless, uti-lization of existing methods in multihop net-works is not feasible since the amount of extratraffic required is excessive. In this article, wepresent a method that allows accurate synchro-nization of large multihop networks, working atthe application layer while keeping the messageexchange to a minimum. Through an extensiveexperimental study, we evaluate the protocol’sperformance and discuss the factors that influ-ence synchronization accuracy the most.

SYNCHRONIZATION OF MULTIHOP WIRELESSSENSOR NETWORKS AT THE APPLICATION LAYER

The authors presenta method thatallows accurate synchronization oflarge multi-hop networks, working at the applicationlayer while keepingthe messageexchange to the minimum.

MARCO LAYOUT 2/7/11 10:48 AM Page 82


due to the messages required to estimate thecommunication delays.

Another key issue is whether there is a sendertransmitting the current clock values as time-stamps (sender-to-receiver) or not (receiver-to-receiver). According to the time analysisperformed by Maróti et al., the most significantdelays when transmitting messages over a wire-less link are those from the send, receive, andaccess processes [5]. The problem with sender-to-receiver methods is the uncertainty timeintroduced by the send and access processes. Asa receiver-to-receiver method, the ReferenceBroadcast Synchronization (RBS) protocol pro-poses the use of reference broadcast messages toestablish a common time reference and get ridof the transmitter-side non-deterministic errorsources (it is assumed that all devices listening tothe broadcast get the message at the same time).This eliminates the time uncertainty introducedby the send and access processes, and sets a tem-poral reference shared by all the nodes [4].

The biggest drawback of receiver-to-receiversynchronization methods is how to propagate thelocal timestamps of the broadcast-receivers toset a GNT. Elson and Estrin propose a postfacto synchronization method, that is, a methodthat performs synchronization only when it isneeded [6]. The scattering method they proposedoes not give a common time reference to thebroadcast sender; it only synchronizes receivers.However, various authors point out the need ofsetting a network time to propagate the synchro-nism over a multihop network using broadcasts[4, 7]. Thus, nodes in the broadcast domain needto share timing information among them todetermine the GNT.

On the other hand, Timing-sync Protocol forSensor Networks (TPSN) is a sender-to-receiverprotocol that achieves real time with high accu-racy optimizing message exchange. It avoids theindeterminism working at the medium accesscontrol (MAC) layer to precisely timestampmessages at the exact moment they are sent [8].The Flooding Time Synchronization Protocol(FTSP) also uses MAC layer timestamping atboth the sender and receiver sides. The protocolproposes a multihop propagation scheme thatdoes not need any initial configuration to propa-gate synchronization info. This ad hoc structurealso enables dynamically overcoming node andlink failures in the network [5].

However, these sender-to-receiver synchro-nization schemes require accessing the lower lay-ers, which is not always possible when usingstandard or complex protocols. Current WSNapplications are developed using a wide varietyof hardware, software, and communication pro-tocols. Some of them are proprietary; others usecommon development platforms that havebecome de facto standards (Motes, i-beans); andothers implement the two current standards suit-able to be used in WSNs: ZigBee and Bluetooth.In all these cases accessing the lower layers isnot possible, and this prevent us from usingsender-to-receiver synchronization.

The Multihop Broadcast Synchronizationprotocol is a receiver-to-receiver synchronizationscheme that nonetheless obtains a GNT workingat the application layer. MBS is suitable for

large multihop networks, keeping the number ofmessages in the same order of magnitude whencompared to the sender-to-receiver methods.

MULTIHOP BROADCAST SYNCHRONIZATIONPROTOCOL

Many of the functions performed by sensornodes require a precise time measurement (e.g.,bit time calculation in communications). Thedevices commonly used to provide an accuratetime base are oscillators. Although these clocksideally provide a heartbeat at a constant rate, allof them present a frequency tolerance (whoselimits are provided by its manufacturers) that isalso slightly affected by the operation tempera-ture and aging. Thus, two clocks, even manufac-tured in the same process, will not have a priorithe same frequency.

This way, nodes in a WSN measure time withoscillators at slightly different rates. This diver-gence, called clock skew, can go up to 150 partsper million (PPM) (i.e., each second the nodeswill commit an error than can go to 150 μs).Apart from the clock skew, there is an offsetamong clocks because each node starts at a dif-ferent instant. Given a node i, with clock skew siand offset ki, its clock reading ti can be written as

ti = sit + ki, i = 1…n. (1)

Thus, synchronizing the clock of node iimplies estimating and compensating clock skewand offset (si, ki). The most used procedure toperform this adjustments is broadly described inthe literature [2, 4, 5, 9]. There is a referenceclock tr to which all the nodes are synchronized.A sync-point k is defined as a pair of timestampscollected at the same time tk in the referencenode and in the nodes that want to be synchro-nized: {ti

k, trk}. Once each node stores several

sync-points at different instants, the offset andskew (ki*, si*) differences with the reference canbe calculated using linear regression:

(2)

where the bar indicates average. This way, everynode can estimate the global time (tr*) from itslocal clock:

(3)

Once a node is synchronized, it can propa-gate the estimated tr* to others, creating newsync-points that spread the GNT in multihopnetworks [5].

Strictly speaking, it is not possible to obtain async-point at the same instant in two nodes notphysically linked; there will be unknown delaysin the synchronization message exchange. Whiledeterministic uncertainties in wireless linksslightly affect the synchronization accuracy, non-deterministic ones drastically reduce it [1, 4, 5,

tt k

sr

i i

i

**

*.=

−

st t t t

t t

k t

irk

r ik

ik

rk

rk

i i

*

*

=−( ) −( )

−( )=

∑

∑2

−− s ti r* ,

Although theseclocks ideally provide

a heartbeat at a constant rate, all of

them present a frequency tolerancethat is also slightly

affected by the oper-ation temperature

and aging. Thus, twoclocks, even manu-

factured in the sameprocess, will nothave a priori thesame frequency.



8]. TPSN and FTSP eliminate the biggest uncer-tainties (send, receive, and access time) by time-stamping at the MAC layer the send and receiveinstant of a message. This way, to obtain therequired sync-points accurately, only one mes-sage is necessary.

When access to the low layers is not possible,reference broadcast messages eliminate theuncertainty at the sender side. Unfortunately,although the reception instant of a broadcastmessage is tight, this procedure does not providethe required sync-points directly because thesender is not synchronized [4, 10]. To the best ofour knowledge, existing methods require addi-tional message exchange between every receivernode to set the GNT. This makes multihop prop-agation very inefficient [4, 7].

Our MBS protocol obtains sync-points usingreference broadcasts with considerably less mes-sage overhead than other methods that also usereference broadcasts, and manages to synchro-nize all the nodes (even the broadcast sender) aswell, making multihop propagation easy.

PROTOCOL DESCRIPTIONTo illustrate how MBS works, we consider theexample network in Fig. 1. Nodes in MBS per-form two different tasks. Propagators are thosethat spread the GNT broadcasting synchroniza-tion messages. Timestampers are nodes that cannotify the propagators about the timestampwhen the previous broadcast message arrives. InFig. 1, N3, N6, N9, and N11 are propagators.N1, N7, and N10 are timestampers; the providedtimestamps are referred to the GNT, so theyhave to be synchronized when notifying the cor-responding propagator node. To overcome thisrequirement when initiating the process, N1 willbe the node whose local clock will be the GNT(i.e., N1 is the global time provider [GTP]).

The synchronization will be made in two hops,the same number of hops FTSP would need. Inthe first hop, N6 will synchronize the nodes in thedashed ring to N1’s clock (using a technique simi-lar to that of RBS). Then, N3 using N7 as times-tamper, and N9 and N11 using N10, willsynchronize the nodes in the dotted rings andpropagate the GNT one hop away. N6 must alsobe synchronized, and, as it is the only propagatorconnected to the GNT, it will be synchronized inthe second hop by any of the other propagators.

This process will be done using two differentmessages: SyncBC and TimeUC. The first type isbroadcast by propagators and is equivalent tothe reference broadcast in RBS: messages thattrigger timestamping of the receiving instant atthe sender side. It contains three fields: thepropagatorID identifying the sender, the sequen-ceNumber of the message, and the timeStampwhen the previous SyncBC arrived. These mes-sages are used to propagate the GNT the sameway as synchronization messages in FTSP [5].TimeUC messages are unicast messages used bytimestampers to notify the propagators about thetime when the last SyncBC arrived. They havethe following fields: the timeStamperID identify-ing the sender, and the corresponding propaga-torID, sequenceNumber, and timeStamp aboutwhich they are informing. Now we explain thesequence to synchronize the network in Fig. 1.We indicate the message sent specifying thefields: SyncBC (propagatorID; sequenceNumber;timeStamp) and timeUC (timeStamperID; propa-gatorID; sequenceNumber; timeStamp). The firsthop would be as follows:1. N6 initiates the synchronization process by

sending a SyncBC (N6; 0; void) message. Thenodes that receive the message (N1, N2, …)timestamp the arrival of the SyncBC fromnode 6 with sequence number 0; that is,TSN1{N6, 0}, TSN2{N6, 0}, ….

2. N1 informs N6 about the timestamp when itreceived the last SyncBC message: TimeUC(N1; N6; 0; TSN1{N6, 0}).

3. N6 sends the timestamp of the previousSyncBC message and sets a new referencepoint for timestamping: SyncBC (N6; 1;TSN1{N6, 0}). At this instant, all Ni neighborsof N6 have their respective sync-points fromthe first SyncBC: [TSNi{N6, 0}, TSN1{N6, 0}].N1 does not need it because it rules the GNT.From now on, steps 2 and 3 will be repeated,

causing all nodes in range of N6 to have a col-lection of sync-points. Then, using linear regres-sion, they get synchronized, calculating theiroffset and skew differences to the referenceclock. In that first hop, all nodes within thedashed ring will be synchronized to N1’s clock.

Note that N6 does not have the global time.To fix this and to propagate the clock one hopaway, any other propagator, such as N3, initiatesthe above described sequence. The subsequenthops will be performed following the same pro-cedure. Each propagator broadcasts SyncBCmessages including the timestamp of the previ-ous SyncBC message. Timestampers notify thepropagators about the last timestamp. All nodesin a range get a collection of sync-points thatallow them to synchronize to GNT.

Accuracy of the sync-points will be degradedas nodes are farther away from the clock genera-tor (N1), similar to sender-to-receiver methods.When synchronizing the nodes situated withinthe dotted rings, the timestampers used by thepropagators (N7 and N10) are one hop awayfrom the GNT (N1), which will increase theerror committed.

In case of near failure (i.e., low batteries),propagators and timestampers can transfer theirresponsibilities to neighboring nodes. In anycase, similar to the scheme proposed by Maróti

Figure 1. A grid network where propagator nodes send broadcast messages, andtime stamper nodes reply to them with the arrival time of the broadcast. Allthe nodes also have direct communication links with their neighbors.

Synchronization link

Global time provider

N1

N9

N4

N8

N11 N12

N16 N15 N14 N13

N2

N6 N5

N10

Timestamper

Propagator

Common node

N3

N7



et al., a network could be autonomous and self-healing in terms of synchronization using nodeidentifiers to automatically assign roles [5].

PRACTICAL ISSUESEarlier we described the MBS protocol withouttaking into account practical issues like how toinitialize the algorithm, how to determine therole to be performed by every node, and so on.To understand the general case, let us first con-sider several examples.

Let us assume that we have a node, N1, thatis already synchronized, and a couple of nodes,N2 and N3, that are not yet synchronized. Inorder to synchronize N2 and N3 with respect toN1, an intermediate node, N4, that can sendbroadcast messages to N1, N2, and N3 is need-ed. Thus, N4 will act as a propagator node to itsneighbors, and N1 will act as the timestamperfor N4. Thus, the general rule is that all theneighbors of the neighbors of N1 can be synchro-nized with respect to N1 provided that thesenodes have a one-hop estimation of the time ofN1. In the same manner, N1 will have a one-hopestimation of the time in the node with respectto which it is synchronized, which in turn willhave a one-hop estimation, and so on until theGTP is reached. The number of hops needed toreach the GTP can be used to define the qualityof a node synchronization, which we call HopId.

Now consider a node N1 that is not synchro-nized. In order to become synchronized, a nearby —synchronized — node must assume the role oftimestamper. Analogous to the previous case, all theneighbors of the neighbors of N1, which are alreadysynchronized, can behave as timestampers. There-fore, the one with the best GNT estimation (i.e., thelowest HopId) must be chosen as timestamper. Ifmore than one node have the lowest HopId, theone that has more synchronizable nodes will be cho-sen, and the node that is neighbor of both the times-tamper and N1 will act as the propagator.

In order to improve performance, if a nodelistens to more than one propagator, it must onlyuse sync messages from the propagator whosetimestamper has lower HopId. Also, if a node isacting as timestamper for a propagator, it cannotuse incoming SyncBC messages from this propa-gator to synchronize itself.

Now, the mechanism should be clear. First,all nodes but the GTP (with HopId equal to 0)are not synchronized. Nodes close to the GTPwill be synchronized with respect to the GTP,and they will be assigned a HopId equal to 1.Synchronization will spread across the entirenetwork following the rules discussed above.

Still, several criteria are possible to set theGTP, such as minimizing the total synchronizationerror expressed as the sum of the HopId of everynode, maximizing the number of nodes with lowerHopId, or setting an upper bound for the HopId.

MBS PROTOCOL EVALUATION ANDCOMPARISON

We have evaluated the MBS protocol in two dif-ferent architectures. Both are used to build mul-tihop WSNs following two standard protocols:ZigBee and Bluetooth. In the case of ZigBee,

we use a common platform: an Atmel microcon-troller (ATMega128) with the Chipcon’s CC2420transceiver [9]. Developing the entire stack tomake the network ZigBee-compliant requires alot of work, thus we used the EmberZNetembedded software. This way, we built a multi-hop, auto-routing and self-healing ZigBee net-work working at the application layer.

For the Bluetooth architecture, a MicrochipPIC16F876 microcontroller manages a MitsumiBluetooth module (WML-C20) through HCI,the standard Bluetooth Host Controller Inter-face. This design enables us to access low-powermodes and implement tree topology networksusing scatternets.

To evaluate MBS’s performance, we connectevery node to a wired bus, where one of themperiodically generates a pulse. All the nodes,which are synchronized with MBS, timestampthe receiving instant with their GNT estimationand send back the data to the PC, where thetimestamp differences with respect to the GTPare considered to obtain the synchronizationerror.

SINGLE-HOP SYNCHRONIZATIONWhen synchronizing wireless nodes, all methodsuse one of the following strategies: to timestampat the sender and receiver side, or use referencebroadcasts, timestamping only at the arrivingside. In Table 1 we compare the alignment errorsof some synchronization schemes presented inthe references.

The timing accuracy among nodes dependsmainly on the hardware and firmware architec-ture: how the sending and arriving moments aredetected, and which times (propagation, access,etc.) are affected and their uncertainty. Errorsshown in Table 1 determine the accuracy of eachsync-point that will be used to perform the linearregression in Eq. 2. Of course, the fewer theerrors, the more accurate the synchronization.Other factors that also affect the estimation arethe following:• The distribution of the errors (uniform, Gaus-

sian, etc.) will determine how the linearregression eliminates them and the quality ofthe clock estimation. The time differencebetween reception instants of broadcast mes-sages follows a Gaussian distribution with thearchitectures described before [3, 4].

• The local oscillator drift is influenced by theinitial accuracy (difference between the oscil-lator output frequency and the specified fre-quency at 25°C at the time of shipment by themanufacturer), temperature stability, andaging. Its behavior can also condition the pre-cision and the timing lifetime.

• Finally, the frequency and number of sync-points used in the estimation will determinethe expected accuracy.As stated by van Greunen and Rabaey, not

all sensor networks applications have the samesync needs in terms of accuracy [11]. In order toprovide the reader with some guidelines thatcould help in deciding on the best suited hard-ware (transceivers, crystals, etc.) and firmware(sync-message rate, number of data points toperform regression) in each application, we havecharacterized synchronization behavior in several

All the nodes, whichare synchronized

with MBS, timestamp the

receiving instant withtheir GNT estimation

and send back thedata to the PC,

where the timestampdifferences with

respect to the GTPare considered to obtain the

synchronization error.



scenarios. We compared the two architecturesdescribed above, Bluetooth and ZigBee, eachwith different alignment errors, as shown inTable 1. Both alternatives have been tested withtwo local oscillators with different frequency sta-bilities of 40 PPM and 150 PPM. We have alsochanged the synchronization rates (30 s and 300s) and number of sync-points used in the regres-sion (3, 6, 8, 12, 20, and 50). In Table 2 we showthe average and maximum synchronization errorwith 95 percent probability for both scenarios.

The first interesting result is that synchroniza-

tion precision depends mainly on the numberand accuracy of the sync-points used to estimateparameters by regression. Thus, if we have verylow computation resources or need low-accuracytiming we can use a few sync-points (i.e., alightweight method) [11]. In contrast, if the high-est accuracy is needed, we need powerful hard-ware to use the maximum number of points inthe regression.

Contrary to what people might expect andagreeing with what Maróti et al. have found out,synchronization rate and oscillator accuracy bare-ly affect precision [5]. This makes sense if we con-sider the local clock stable in the short to mediumterm, something that fortunately is so in mostcases. Aging has a negligible effect on stability:one to three PPM each year. Temperature influ-ences clock drift more severely: tens of PPMs inthe operating temperature range. In worst cases,this translates into an error of 1 μs/°C.

The reduced influence of synchronizationrate can be used in many ways. By increasing it,we can reduce the initial settling time in multi-hop networks, quickly synchronize a new node,or maintain accuracy in sudden temperaturevariations. Lowering the rate will be useful tominimize extra traffic and reduce power con-sumption. Despite the results shown in Table 2,the synchronization interval cannot be as long aswe want. Linear regression estimates the skewand offset of the local clock referred to theGNT, and there will always be errors.

According to Eq. 1, the more time from thelast resynchronization, the larger the error of

Table 1. Comparison of alignment error in synchronization methods.

Average error (μs) Worst case error (μs)

Sender — Receiver synchronization

ZigBee (Motes 2.4 GHz) [9] 14.9 61.0

TPSN (Motes 916 MHz) [8] 16.9 44.0

FTSP (Motes 433 MHz) [5] 1.4 4.2

Receiver — Receiver synchronization

RBS (Motes) [4] 21.9 93.0

MBS (Bluetooth) 4.5 18.0

MBS (ZigBee) 22.2 52.0

Table 2. Synchronization error in the MBS method with one hop (in μs).

ZigBee40 PPM

ZigBee150 PPM

Bluetooth40 PPM

Bluetooth150 PPM

Avg. Max.a Avg. Max.a Avg. Max.a Avg. Max.a

N = 3 tSync = 30 s 39.26 105.20 39.45 106.42 7.67 20.92 8.10 21.70

tSync = 300 s 39.99 108.09 39.59 109.05 8.00 21.61 7.73 20.88

N = 6 tSync = 30 s 21.12 52.65 22.06 55.23 4.17 10.30 4.34 10.77

tSync = 300 s 21.90 53.52 21.29 52.07 4.41 10.99 4.36 10.93

N = 9 tSync = 30 s 17.27 40.95 16.58 40.66 3.30 8.19 3.30 8.25

tSync = 300 s 16.10 39.38 15.72 38.90 3.28 8.22 3.36 8.39

N = 12 tSync = 30 s 13.33 32.83 14.39 35.13 2.82 6.84 2.79 6.89

tSync = 300 s 13.99 34.52 13.77 34.87 2.73 6.62 2.68 6.63

N = 20 tSync = 30 s 11.29 27.20 9.83 24.23 1.96 4.71 2.13 5.14

tSync = 300 s 10.61 26.13 10.46 25.51 2.19 5.29 2.13 5.30

N = 50 tSync = 30 s 7.03 17.46 6.77 15.58 1.23 2.93 1.27 3.10

tSync = 300 s 6.22 14.80 7.54 18.43 1.30 3.27 1.32 3.21

N is the number of sync-points used to perform regression, and tSync is the synchronization interval.a Maximum synchronization error with 95% probability.



this estimation. Figure 2 shows the synchroniza-tion error between two Bluetooth nodes and thereference node. Here we used sync messagesevery 30 s, and after 500 messages (about 4 h)the synchronization process was stopped. Wecan see how after the stopping instant, the syn-chronization error of each node grows depend-ing on the last estimation of the skew and offset.Additional results may be found in [3].

MULTIHOP GNT PROPAGATIONCumulative errors when propagating the net-work time only depend on the estimation’s accu-racy of the corresponding timestamper’s clock.When the timestamper is the one setting the ref-erence, the precision will be as described earlier.Differences arise when it owns an n-hop estima-tion of the network clock.

We have evaluated the behavior of MBS intwo different scenarios. With the same philoso-phy of the single-hop case, we have tested itsperformance with different numbers of sync-points. The first scenario is a four-hop Bluetoothscatternet. Bluetooth networks are made up ofpiconets, i.e., star networks with one master andup to seven slaves, where only the master cansend broadcast messages. Piconets form scatter-nets using nodes playing both master and slaveroles. Time-stampers are nodes that must beslaves in two different piconets.

In the case of ZigBee, we implemented amesh network similar to that of Fig. 1 but havingfour-hop depth. Synchronization errors for theBluetooth and ZigBee networks are shown inFigs. 3 and 4, respectively.

As in the case of one hop, synchronizationerror depends chiefly on the number of pointsused to perform regression and the alignmenterror in the establishment of sync-points. Wecan see how as the number of hops increases,synchronization error becomes sensitive to thenumber of sync-points used. Again, if the appli-cation needs higher synchronization accuracythan the alignment error between nodes, it ispossible to reduce error by increasing the num-ber of sync-points to perform regression.

We find no point in comparing precisionamong methods because it mainly depends onthe accuracy estimating sync-points and thenumber of pairs used. That is to say, architecture(communication transceiver, protocol, memoryavailable, etc.) will be much more relevant thanthe synchronization protocol used. On the otherhand, the amount of messages needed will havea big influence on the applicability of themethod. This is just one of the strongest pointsof MBS; it drastically reduces the number ofmessages compared to other receiver-to-receiverprotocols [4, 7]. Indeed, it is on the same orderof magnitude as FTSP (the most efficient sender-to-receiver method) [5].

CONCLUSIONS

Local clocks of nodes in wireless sensor net-works have different offset and accuracy. Syn-chronization is mainly achieved by collectingsync-points (pairs of timestamps collected at thesame time in the reference node and in the nodethat wants to be synchronized) and performing

linear regression to compensate for differencesamong nodes’ clocks.

Medium access time is a non-deterministicerror that hinders accurate timestamping oftransmission instants when working at the high-est layers of protocol stacks. In these cases, sev-eral techniques based on a receiver-to-receiverscheme have to be used. Nevertheless, thesemethods set a network time shared by all thenodes (including the broadcast senders) at theexpense of a high network load. This drawbackmay be inadmissible for large networks. In thisarticle we have presented MBS, a multihopbroadcast synchronization protocol that is ableto efficiently set a common global time. The keyissue of the technique is that each referencebroadcast informs about the timestamp of theprevious one. This way, message exchanging isminimized, similar to other sender-to-receivermethods.

We have implemented the MBS protocol intwo different architectures (Bluetooth and Zig-

Figure 2. Illustration of synchronization error vs. time with sync messages every30 s and 20 pairs used to perform regression. Synchronization process stopsafter 500 sync messages.

Time (hours) 1:00 0:00

20

0

Erro

r (μ

s)

40

60

80

2:00 3:00 4:00 5:00 6:00 7:00 8:00

Figure 3. Average synchronization error for Bluetooth network with 4 hops andsync messages every 30 s; here N is the number of points used to performregression, and the dashed bars represent maximum error with 95 percentprobability.

Sync level2 0

50

0

Sync

hron

izat

ion

erro

r (μ

s)

1 3 4

N = 3N = 6N = 9N = 12N = 20N = 50

100

150

200

250

300

350

400

450

500


Bee), evaluating its performance. Throughexhaustive experimentation we have identifiedthe factors that influence synchronization accu-racy the most. We have analyzed hardware(transceivers, crystals) and firmware (sync-mes-sage rate, number of data points to performregression) issues to provide the applicationdesigner with guidelines that could help in decid-ing on the best suited technology for each appli-cation, and we have characterizedsynchronization behavior in several scenarios.

These results allow the application designerto decide on the hardware architecture and pro-tocol scheme best suited to achieve the requiredsynchronization accuracy.

ACKNOWLEDGMENTSThis work was supported in part by the SpanishMCYT under AmbienNet project (TIN2006-15617-C03) and by the European Commissionunder the MonAMI project.

REFERENCES[1] I. F. Akyildiz et al., “A Survey on Sensor Networks,” IEEE

Commun. Mag., vol. 40, no. 8, 2002, pp. 102–14.[2] F. Sivrikaya and B. Yener, “Time Synchronization in Sen-

sor Networks: A Survey,” IEEE Network, vol. 18, no. 4,2004, pp. 45–50.

[3] A. Marco et al., “Multi-Hop Synchronization at the Applica-tion Layer of Wireless and Satellite Networks,” Proc. IEEEGLOBECOM ’08, New Orleans, LA, 2008, pp. 1–5.

[4] J. Elson, L. Girod, and D. Estrin, “Fine-Grained TimeSynchronization using Reference Broadcasts,” Proc. 5thSymp. Op. Sys. Design Implementation, Boston, MA,2002, pp. 147–63.

[5] M. Maróti et al., “The Flooding Time SynchronizationProtocol,” Proc. 2nd ACM SenSys ’04, 2004, pp. 39–49.

[6] J. Elson and D. Estrin, “Time Synchronization for WirelessSensor Networks,” Proc. 15th Int’l. Parallel & Distrib. Pro-cess. Symp., San Francisco, CA, 2001, pp. 1965–70.

[7] S. PalChaudhuri, A. Saha, and D. B. Johnson, “AdaptiveClock Synchronization in Sensor Networks,” Proc. 3rdIEEE IPSN, Berkeley, CA, 2004, pp. 340–48.

[8] S. Ganeriwal, R. Kumar, and M. B. Srivastava, “Timing-Sync Protocol for Sensor Networks,” ACM SenSys, LosAngeles, CA, 2003, pp. 138–49.

[9] D. Cox, A. Milenkovic, and E. Jovanov, “Time Synchro-nization for ZigBee Networks,” Proc. 37th South-East-ern Symp. Sys. Theory, Tuskegee, AL, 2005, pp. 135–38.

[10] R. Casas et al., “Synchronization in Wireless SensorNetworks using Bluetooth,” 3rd Int’l. Wksp. IntelligentSolutions in Embedded Sys., 2005, pp. 79–88.

[11] J. van Greunen and J. Rabaey, “Lightweight Time Syn-chronization for Sensor Networks,” Proc. 2nd ACMInt’l. Conf. Wireless Sensor Net. Apps., San Diego, CA,2003, pp. 11–19.

BIOGRAPHIESÁLVARO MARCO ([email protected]) received his degree inelectrical engineering in 2000 and his Ph.D. in electronicengineering in 2007, both from the University of Zaragoza,Spain. Currently, he is a senior researcher in the Depart-ment of Electrical Engineering and Communications at thesame university, and his research interests include sensornetworks, ambient intelligence, and assistive technology.

ROBERTO CASAS received his degree in electrical engineeringin 2000 and his Ph.D. in electronic engineering in 2004,both from the University of Zaragoza. Currently, he is anassistant professor in the Department of Electrical Engi-neering and Communications at the same university, andhis research interests include sensor networks, digital elec-tronics, and assistive technology.

JOSÉ LUIS SEVILLANO RAMOS ([email protected]) received hisPh.D. from the University of Seville, Spain, in 1993. Since1996 he has been an associate professor of computerarchitecture at the University of Seville. Currently, he iscoordinator of the Telefónica Chair on Intelligence in Net-works at the same university. He also serves as VP Mem-bership of the Society for Modeling & SimulationInternational, SCS.

VICTORIA’N COARASA is a technical researcher in the Depart-ment of Electrical Engineering and Communications at theUniversity of Zaragoza, Spain. His research interests includesensor networks, digital electronics, and assistive technolo-gy. He received his M.S. degree in wireless engineering in2007 from the University of Zaragoza, and is currentlyapplying to obtain a Ph.D. degree in electronic engineer-ing.

A’NGEL ASENSIO is a senior researcher in the Department ofElectrical Engineering and Communications at the Universi-ty of Zaragoza. His research interests include sensor net-works, digital electronics, and assistive technology. Hereceived his M.S. degree in electrical engineering in 2005from the University of Zaragoza and is currently applyingto obtain a Ph.D. degree in electronic engineering.

MOHAMMAD S. OBAIDAT [F‘05] ([email protected]) isan internationally known academic/researcher/scientist. Hereceived his Ph.D. in computer engineering from Ohio StateUniversity. He is currently a full professor of computer sci-ence and software engineering at Monmouth University.Among his previous positions are chair of the ComputerScience Department and director of the Graduate Programat MU. He has received extensive research funding, andauthored/co-authored 10 books and over 475 refereedscholarly journal and conference articles. He has served asa consultant for several corporations worldwide and is edi-tor of many scholarly journals including being Editor-in-Chief of Wiley’s International Journal of CommunicationSystems. He is President of the Society for Modeling & Sim-ulation International, SCS. He was awarded the distin-guished Nokia Research Fellowship and the DistinguishedFulbright Award. He has been invited to lecture and givekeynote speeches worldwide. He was the recipient of theBest Paper Award for one of his papers accepted for IEEEAICCSA 2009 and IEEE GLOBCOM 2009. He also receivedthe SCS prestigious McLeod Founder’s Award in recogni-tion of his outstanding technical and professional contribu-tions to modeling and simulation. He is a Fellow of SCS.


Figure 4. Average synchronization error for Zigbee network with 4 hops andsync messages every 30 s; here N is the number of points used to performregression, and the dashed bars represent maximum error with 95 percentprobability.

Sync level2 0

100

0

Sync

hron

izat

ion

erro

r (μ

s)

1 3 4

200

300

400

500

600

700

800

900

1000 N = 3N = 6N = 9N = 12N = 20N = 50


Wireless Communications IEEE Volume 18 Issue 1 February 2011 issue Volume 18 Issue 1 February 2011

Documents

Transcript of Wireless Communications IEEE Volume 18 Issue 1 February 2011 issue Volume 18 Issue 1 February 2011