<%BANNER%>

Behavioral Analysis, User Modeling, and Protocol Design Based on Large-Scale Wireless Network Traces

Permanent Link: http://ufdc.ufl.edu/UFE0022539/00001

Material Information

Title: Behavioral Analysis, User Modeling, and Protocol Design Based on Large-Scale Wireless Network Traces
Physical Description: 1 online resource (250 p.)
Language: english
Creator: Hsu, Wei-Jen
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2008

Subjects

Subjects / Keywords: aware, behavior, mobility, network, protocol, trace, wlan
Computer and Information Science and Engineering -- Dissertations, Academic -- UF
Genre: Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: In this dissertation we describe the TRACE framework, which is a five-step procedure for the environment-aware approach towards wireless mobile computer networks. As mobile computer networks become ubiquitous and deeply integrated with the daily lives, it is crucial to understand the network and design its protocols and services with the environment-aware approach: We first collect extensive network Traces that reflect truthfully the detailed behaviors of its users, and Represent the rich data sets in concise representations. Then we Analyze these constructed representations to Characterize the users. While many observed characteristics are interesting in themselves and reveal important differences between the realistic environment and commonly made assumptions in the literature, we further add values to these findings by employing them in various important tasks, including modeling the network users and designing routing protocols. The dissertation is centered around three major case studies, ranging from the microscopic, individual user behavior in the wireless networks to the macroscopic, global user encounter patterns. Specifically, in the case studies, we (1) observe the individual user mobility from the collected traces, identify skewed preferences and periodical re-appearance at the same location as prominent mobility characteristics, and propose the time-variant community (TVC) mobility model to capture such behaviors. The TVC model is flexible to match with many empirical traces while being mathematically tractable. (2) We construct an efficient way for mobile users to summarize their mobility preferences based on singular value decomposition (SVD) and calculate the distance metric between users. Based on this distance metric, we identify user groups in the population based on their mutual similarities, and design a profile-cast service to deliver messages to these behavioral groups without knowing their identities. (3) We further analyze the global encounter patterns between nodes, observe a fast-emerging Small World encounter pattern, and leverage such a network property to design an efficient message dissemination protocol named CSI, in which Communication relies on the Stable yet Implicit structures in mobile networks.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Wei-Jen Hsu.
Thesis: Thesis (Ph.D.)--University of Florida, 2008.
Local: Adviser: Helmy, Ahmed H.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2008
System ID: UFE0022539:00001

Permanent Link: http://ufdc.ufl.edu/UFE0022539/00001

Material Information

Title: Behavioral Analysis, User Modeling, and Protocol Design Based on Large-Scale Wireless Network Traces
Physical Description: 1 online resource (250 p.)
Language: english
Creator: Hsu, Wei-Jen
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2008

Subjects

Subjects / Keywords: aware, behavior, mobility, network, protocol, trace, wlan
Computer and Information Science and Engineering -- Dissertations, Academic -- UF
Genre: Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: In this dissertation we describe the TRACE framework, which is a five-step procedure for the environment-aware approach towards wireless mobile computer networks. As mobile computer networks become ubiquitous and deeply integrated with the daily lives, it is crucial to understand the network and design its protocols and services with the environment-aware approach: We first collect extensive network Traces that reflect truthfully the detailed behaviors of its users, and Represent the rich data sets in concise representations. Then we Analyze these constructed representations to Characterize the users. While many observed characteristics are interesting in themselves and reveal important differences between the realistic environment and commonly made assumptions in the literature, we further add values to these findings by employing them in various important tasks, including modeling the network users and designing routing protocols. The dissertation is centered around three major case studies, ranging from the microscopic, individual user behavior in the wireless networks to the macroscopic, global user encounter patterns. Specifically, in the case studies, we (1) observe the individual user mobility from the collected traces, identify skewed preferences and periodical re-appearance at the same location as prominent mobility characteristics, and propose the time-variant community (TVC) mobility model to capture such behaviors. The TVC model is flexible to match with many empirical traces while being mathematically tractable. (2) We construct an efficient way for mobile users to summarize their mobility preferences based on singular value decomposition (SVD) and calculate the distance metric between users. Based on this distance metric, we identify user groups in the population based on their mutual similarities, and design a profile-cast service to deliver messages to these behavioral groups without knowing their identities. (3) We further analyze the global encounter patterns between nodes, observe a fast-emerging Small World encounter pattern, and leverage such a network property to design an efficient message dissemination protocol named CSI, in which Communication relies on the Stable yet Implicit structures in mobile networks.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Wei-Jen Hsu.
Thesis: Thesis (Ph.D.)--University of Florida, 2008.
Local: Adviser: Helmy, Ahmed H.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2008
System ID: UFE0022539:00001


This item has the following downloads:


Full Text

PAGE 1

1

PAGE 2

2

PAGE 3

3

PAGE 4

Iwouldliketothankrstandforemostmyadviser,Dr.AhmedHelmy,forhisguidanceandenlightenmentthroughthecourseofmyworktowardsthePh.D.degree.Dr.Helmyisnotonlyagreatacademicadviser,butalsoagreatmentor,whohelpedmesignicantlyinkeepingthefaithandpursuingmydream,andagreatrolemodelfromwhomIlearnedsomanydierentthingsinlife.LookingbackattheyearsIworkedwithhim,theywerenotonlyintellectuallysatisfying,butalsoatremendouslyjoyfuljourneyinmylife.Iwouldlikealsoexpressdeepgratitudetomysupervisorycommitteemembers,includingDr.SartajSahni,Dr.AlinDobra,Dr.DapengWu,andDr.YeXia,whohavehelpedmesignicantlyintheprocessofformingmydissertation.Theirinputsbringnewperspectivesandunderstandingtotheproblemandmakethedissertationmorecomplete.Inaddition,Iwouldliketoexplicitlythankthemfortheirinputsduringtheexams,whichhavecollectivelymadethewholeprocessintellectuallychallengingandrewarding.IwouldliketoalsoextendmythankstomanyprofessorsandcolleaguesIhadtheprivilegetoworkwith.Theyhavehelpedmeinsomanydierentways.ForDr.DebojyotiDutta,Iappreciatehispatiencetospendtimewithmeformingtheresearchproblemsandidentifyingthepropertools,improvingmywriting,andingeneralmanyhelpfuldiscussions.ForProfessorKonstantinosPsounisandDr.ThrasyvoulosSpyropoulos,Ithankthemfortheirhelpduringthetime-variantcommunitymobilitymodelproject,whichtookmetotheappreciationofrigoroustheoreticalwork.ForDr.JabedFaruque,Iwouldliketothankrsthismanysupportivechallengestomyresearchproblems,whichmakemethinkmorecarefully,andsecondIamalsoindebttohimforthetimehesparedfromresearchtomaintainanicecomputingenvironmentforthegroup,withoutwhichmanyoftheresultsinthisdissertationwouldbeatleastdelayed.ForDr.FanBai,Shao-ChengWang,andDr.SaponTanachaiwiwat,Iwouldliketothankthemformanyhelpfulsuggestionsandreviewcommentsonmyresearchwork. 4

PAGE 5

5

PAGE 6

page ACKNOWLEDGMENTS ................................. 4 LISTOFTABLES ..................................... 10 LISTOFFIGURES .................................... 11 ABSTRACT ........................................ 15 CHAPTER 1INTRODUCTION .................................. 17 1.1EmergenceofMobileNetworks ........................ 17 1.2Behavior-awareNetworkApproach ...................... 18 1.3TheTRACEFramework ............................ 20 1.4StudyComponents ............................... 21 1.5Contributions .................................. 24 2RELATEDWORK .................................. 25 2.1TraceCollections ................................ 25 2.2TraceAnalysis .................................. 29 2.2.1GeneralStatistics ............................ 29 2.2.2DataMiningTechniques ........................ 32 2.2.3GraphAnalysis ............................. 34 2.3MobilityModeling ............................... 35 2.4MessageForwardingProtocolDesigninDTNs ................ 39 3DATASETS ..................................... 45 3.1DataSetsUsed ................................. 45 3.2TraceCollectionMethods ........................... 49 3.3DenitionofTerms ............................... 51 3.4DetailedDescriptionsofOurTraces ...................... 53 4CASESTUDYI:MODELINGINDIVIDUALUSERMOBILITY ........ 55 4.1OnModelingUserAssociationsinWirelessLANTraces ........... 55 4.1.1Introduction ............................... 55 4.1.2AnalysisofIndividualUserBehavior ................. 57 4.1.2.1Activenessoftheusers .................... 59 4.1.2.2Macro-levelmobilityofusers ................. 62 4.1.2.3Micro-levelmobilityofusers ................. 64 4.1.2.4Therepetitiveassociationpatternofusers ......... 68 4.1.3ConclusionsandFutureWork ..................... 72 6

PAGE 7

................................ 73 4.2.1Introduction ............................... 73 4.2.2Time-variantMobilityModel ...................... 75 4.2.2.1MobilitycharacteristicsobservedinWLANtraces ..... 76 4.2.2.2Constructionofthetime-variantcommunitymodel .... 78 4.2.3TheoreticalAnalysisoftheTVCModel ................ 82 4.2.3.1Nodalspatialdistribution .................. 84 4.2.3.2Averagenodedegree ..................... 85 4.2.3.3Hittingtime .......................... 86 4.2.3.4Meetingtime ......................... 90 4.2.4ValidationoftheTheorywithSimulations .............. 93 4.2.4.1Nodalspatialdistribution .................. 96 4.2.4.2Averagenodedegree ..................... 97 4.2.4.3Hittingtimeandmeetingtime ................ 98 4.2.5ApplicationI:GenerationofMobilityScenariosforSimulation ... 99 4.2.5.1MatchingmobilitycharacteristicswithWLANtraces ... 102 4.2.5.2Matchingmobilitycharacteristicswithvehiclemobilitytraces ............................. 104 4.2.5.3Matchingcontactcharacteristicswithencounter-basedtraces ............................. 105 4.2.6ApplicationII:UsingTheoryforPerformancePrediction ...... 107 4.2.6.1Estimationofthenumberofnodesneededforgeographicrouting ............................. 108 4.2.6.2Predictingmessagedeliverydelaywithepidemicrouting 108 4.2.7ConclusionsandFutureWork ..................... 111 5CASESTUDYII:MININGBEHAVIORALGROUPSINTHETRACES .... 113 5.1Introduction ................................... 113 5.2Preliminaries .................................. 116 5.2.1ChoiceofDataSetsandRepresentations ............... 116 5.2.2PreliminariesofClusteringTechniques ................ 118 5.3Challenges .................................... 119 5.4SummarizingtheAssociationPatterns .................... 121 5.4.1CharacteristicsofAssociationPatterns ................ 121 5.4.2SummarizationMethods ........................ 123 5.4.3InterpretingSingularValueDecomposition .............. 126 5.5ClusteringUsersbyEigen-BehaviorVectors ................. 129 5.5.1Eigen-BehaviorDistance ........................ 129 5.5.2SignicanceoftheClusters ....................... 130 5.6InterpretationoftheClusteringResults .................... 132 5.7PotentialApplications ............................. 136 5.8Prole-Cast:Behavior-AwareMobileNetworking .............. 137 5.8.1Prole-CastinginDelayTolerantNetworks .............. 139 7

PAGE 8

............... 141 5.8.3EvaluationandComparison ...................... 142 5.8.3.1Evaluationsetup ....................... 142 5.8.3.2Evaluationresults ....................... 144 5.8.4ExtensionsoftheProle-CastService ................. 148 5.9Conclusion .................................... 149 5.10AlternativeMethods .............................. 150 5.10.1VariousDistanceMetrics ........................ 150 5.10.2VariousDataRepresentations ..................... 152 6CASESTUDYIII:UNDERSTANDINGTHEGLOBALNODALENCOUNTERPATTERNS ...................................... 155 6.1Introduction ................................... 155 6.2EncountersbetweenNodes ........................... 158 6.3Encounter-RelationshipGraph ......................... 162 6.4TheReasonsunderneaththeSmallWorldEncounterPattern ........ 168 6.5CapturingUserFriendshipinWLANTraces ................. 170 6.6InformationDiusionusingEncounters .................... 175 6.6.1IdealScenarios .............................. 175 6.6.2SelshUsers ............................... 177 6.6.3RemovalofShortEncounters ...................... 178 6.7ConclusionsandFutureWork ......................... 180 6.8BiParetoDistributionandKolmogorov-SmirnovTest ............ 182 6.9AdditionalExperimentResults ........................ 184 7CASESTUDYIII:CSI:APARADIGMFORBEHAVIOR-ORIENTEDDELIVERYSERVICESINMOBILEHUMANNETWORKS ................. 190 7.1Introduction ................................... 190 7.2Background ................................... 193 7.2.1Mobility-BasedUserBehaviorRepresentation ............ 193 7.2.2Traces .................................. 194 7.3UnderstandingSpatio-TemporalCharacteristicsofUserBehavioralPatterns 195 7.4TheBehavior-DrivenCommunicationParadigm ............... 198 7.5ProtocolDesign ................................. 199 7.5.1AssumptionsandDesignRequirements ................ 199 7.5.2RelationshipbetweenBehavioralProlesandEncounters ...... 201 7.5.3CSI:TargetMode ............................ 203 7.5.4CSI:DisseminationMode ........................ 204 7.6SimulationResults ............................... 208 7.6.1CSI:TargetMode ............................ 209 7.6.1.1Simulationsetup ....................... 209 7.6.1.2Simulationresults ....................... 213 7.6.2CSI:DisseminationMode ........................ 216 7.6.2.1Simulationsetup ....................... 216 8

PAGE 9

....................... 217 7.7Discussions ................................... 219 7.7.1AdditionalOverhead .......................... 219 7.7.2PrivacyIssues .............................. 221 7.8ConclusionsandFutureWork ......................... 223 8CONCLUSIONSANDFUTUREWORK ...................... 224 APPENDIX:ObtainingMobilityInformationthroughSurveys ............. 227 A.1MobilitySurvey ................................. 227 A.2WeightedWaypointMobilityModelandItsImpactonAdHocNetworks 228 A.2.1GeneralDescriptionoftheWeightedWaypointModel ........ 228 A.2.2EstablishinganExampleWWPModelbasedonUSCCampus ... 229 A.2.3SimulationResults ........................... 232 A.2.3.1PropertiesofWWPmodel .................. 232 A.2.3.2ImpactoftheWWPmodelonnetworkperformance ... 233 A.3ACongestionAlleviationMechanismforWLANs .............. 234 A.3.1Flow-SwitchingMechanism ....................... 235 A.3.2SimulationResults ........................... 237 REFERENCES ....................................... 241 BIOGRAPHICALSKETCH ................................ 250 9

PAGE 10

Table page 3-1Statisticsofstudiedtraces .............................. 47 4-1Parametersofthetime-variantcommunitymobilitymodel ............ 79 4-2Parametersforthescenariosinthesimulation ................... 95 5-1Theaveragesignicancescoreforvarioussummariesofuserassociationvectors 125 5-2Jaccardindicesbetweenuserpartitions. ...................... 152 6-1EquationsfortheCCandPL. ............................ 166 6-2ThegraphpropertiesoftheERgraphswithselectedlinks. ............ 169 6-3Correlationcoecientforfriendshipindexesforalltraces. ............ 172 6-4BiParetodistributionttingtothetotalencountercurves. ............ 184 6-5Exponentialdistributionttingtothefriendshipindex. .............. 184 A-1Transitionprobabilitymatrix. ............................ 231 A-2Move-stopratio. ................................... 233 10

PAGE 11

Figure page 1-1IllustrationoftheTRACEframework. ....................... 20 1-2Componentsofthestudy. .............................. 22 3-1Illustrationofthetermdenitions. ......................... 51 4-1IllustrationofaMN'sassociationpatternwithrespecttotimeoftheday. .... 58 4-2CCDFofonlinetimefraction ............................ 59 4-3CCDFofnumberofassociationsessionsbyusers ................. 61 4-4CCDFofcoverageofusers. ............................. 63 4-5AveragefractionoftimeaMNassociatedwithAPs. ............... 65 4-6CCDFoftotalhandocountperMN. ....................... 66 4-7Sessiondurationsversushandocount. ....................... 67 4-8CDFforthecoecientofvariationofthehandorate. .............. 68 4-9Networksimilarityindexes. ............................. 70 4-10TwoimportantmobilityfeaturesobservedfromWLANtraces. .......... 77 4-11Illustrationofagenericscenarioofthetime-variantmobilitymodel ....... 81 4-12Anillustrationofasimpleweeklyschedule. .................... 81 4-13Illustrationoftheexpansionofthe\footage"ofcommunity. ........... 93 4-14IllustrationofthecommunitysetupforthegenericcasesofTVCmodel. .... 94 4-15Spatialdistributionofthenode. ........................... 96 4-16Theaveragenodedegreetheoryversussimulationresults. ............ 97 4-17Hittingtimeandmeetingtimetheoryversussimulationresults. ......... 99 4-18MatchingtheMITWLANtracewiththesynthetictrace. ............. 104 4-19Matchingthevehiclemobilitytracewiththesynthetictrace. ........... 105 4-20Matchinginter-meetingtimeandencounterdurationdistributionswiththehumanencountertrace. .................................... 107 4-21Geographicroutingsuccessrate. ........................... 109 4-22Packetpropagationwithepidemicrouting. ..................... 111 11

PAGE 12

.................. 118 5-2AMVDdistance:inter-clusterandintra-clusteruserpairs ............. 120 5-3Distributionofnumberofbehavioralmodesforusers. ............... 121 5-4Distributionofassociationvectorsintherstandthesecondbehavioralmodes 123 5-5ComplementaryCDFfortheratiooftherstbehavioralmodesizetothesecondbehavioralmodesize. ................................. 124 5-6Lowassociationmatricesdimensionality. ...................... 128 5-7Eigen-vectordistance:inter-clusterandintra-clusteruserpairs .......... 131 5-8Cumulativepowercapturedintopfoureigen-behaviorvectors:randomgroupingversusbehavior-basedgrouping. ........................... 132 5-9Usergroupsizefollowsapower-lawdistribution. .................. 133 5-10Twodierentviewsoftheprole-castserviceintheDTN. ............ 140 5-11Thechosenprotocolsforevaluationspanthespectrumofusergroupingknowledgeusedintheforwardingdecisionprocess. ....................... 144 5-12Relativeperformancemetricsofthegroup-castschemes. ............. 148 5-13Theoperationregionsofthecomparedprotocolsinthedeliveryrate-overheadspace. ......................................... 148 5-14Otherdistancemetrics:inter-clusterandintra-clusteruserpairs ......... 151 6-1Traces:CCDFofuniqueencounterfraction .................... 160 6-2Syntheticmodel:CCDFofuniqueencounterfraction ............... 161 6-3CCDFoftotalencountercount. ........................... 162 6-4Uniqueencountercountversustotalencountercount,USC. ........... 162 6-5ChangeintheERgraphmetricswithrespecttotraceperiod ........... 165 6-6Classicationofnodepairsintodierentcategoriesbasedontheirsimilaritymetricrange. ..................................... 169 6-7CCDFoffriendshipindexbasedontime. ...................... 172 6-8Metricsofencounter-relationshipgraphbytakingvariouspercentageoffriends. 174 6-9Unreachableratioofinformationdiusionusingtheepidemicrouting. ...... 177 6-10USCtrace:Unreachableratio. ............................ 177 12

PAGE 13

......................... 179 6-12Theunreachableratioafterremovingshortencounters. .............. 180 6-13Thedelayafterremovingshortencounters. ..................... 180 6-14IllustrationoftheD-statisticsandtheK-Stest. .................. 183 6-15ChangeintheERgraphmetricswithrespecttotraceperiod. .......... 185 6-16Dart-04trace:Unreachableratio. .......................... 187 6-17Dart-04trace:Averagemessagedelay. ....................... 187 6-18MITtrace:Unreachableratio. ............................ 187 6-19MITtrace:Averagemessagedelay. ......................... 188 6-20Dart-03trace:Unreachableratio. .......................... 188 6-21Dart-03trace:Averagemessagedelay. ....................... 188 6-22UFtrace:Unreachableratio. ............................. 189 6-23UFtrace:Averagemessagedelay. .......................... 189 7-1Illustrationoftheassociationmatrixtodescribeagivenuser'slocationvisitingpreference. ....................................... 193 7-2Illustration:considerthetrailingddaysofbehavioralproleattimepointsthatareTdaysapart. ................................... 195 7-3SimilaritymetricsforthesameuserattimegapTapart. ............. 196 7-4CorrelationcoecientofthesimilaritymetricsbetweenthesameuserpairattimegapTapart. ................................... 196 7-5Relationshipbetweenthesimilarityinbehavioralpatternandotherquantities. 201 7-6IllustrationoftheCSI:Tschemeinthehighdimensionbehavioralspace ..... 205 7-7DesignphilosophyoftheCSI:Dscheme. ...................... 206 7-8IllustrationoftheCSI:Dscheme ........................... 208 7-9PerformancecomparisonofCSI:Ttootherprotocols. ............... 214 7-10Splitperformancemetricsbythesimilaritybetweenthesenderandthetargetprole. ......................................... 215 7-11Illustrationsforthecomparisonbetweenonelongrandomwalkandmanyshortrandomwalks. ..................................... 216 13

PAGE 14

............... 218 A-1Thesurveyform. ................................... 227 A-2Thevirtualcampus. ................................. 229 A-3Markovmodeloflocationtransitionofmobilenodes. ............... 230 A-4Pausetimedistributionforlocations. ........................ 231 A-5Flowdurationdistributionforlocations. ...................... 232 A-6Mobilenodedensityversustime. .......................... 233 A-7UnevenowdistributionacrossAPs. ........................ 235 A-8Thecontrolowchartoftheproposedow-switchingmechanism. ........ 235 A-9Flowsre-distributedacrossAPs,relievingcongestionatlibrary1. ......... 238 A-10AllAPs:AverageAPcongestedtimeratio. ..................... 239 A-11ThemostcongestedAP:AverageAPcongestedtimeratio. ............ 240 A-12Averagequalitytimeratioofallows. ....................... 240 14

PAGE 15

15

PAGE 16

16

PAGE 17

3 ].Bydenition,MANETsareself-organized,infrastructure-lessnetworks,andconsideredasstand-alonenetworksinwhichtheparticipantsexchangeinformationamongthemselves.Typically,MANETsconsistofautonomousdevices,andeachdeviceplaysbothrolesofanend-hostandarouteratthesametime.Whilethecommunicationrangeofindividualnodesislimitedtoitsclosevicinityduetothenatureofthewirelessmedium,theend-to-endconnectivityinthenetworkisprovidedbythecooperationofitsparticipants,throughmulti-hopforwarding,sometimesinvolvingtemporarystorageofthemessagesinthenon-volatilememoryofintermediatenodes(inasub-caseofMANETsgenerallyknownastheDelayTolerantNetworks,orDTNs[ 4 ]).MANETsprovideanattractivealternativewayofcommunicationwherethesetupofaninfrastructureisinfeasibleortoocostly,orwhenthedisseminatedinformationismeantforonlylocalparticipantssothereisnoneedtoreachtheInternet.PotentialapplicationsofMANETsincludevehicularnetworks(VANET)[ 7 30 ],disasterrelief[ 5 ],wild-lifetracking[ 32 33 ], 17

PAGE 18

6 ],tonameafew.Theemergenceofpersonalizedportablewirelesscommunication/computingdevices(e.g.,PDAs,smartphones)alsoopensthedoorforcreatingamobilevirtualsocialnetworkbetweenpeople.WeenvisionsuchaMANETwouldfacilitatesocializingapplications(e.g.,matchingpeoplewithsimilarinterests,informationsharingamongsmallgroups,etc.)inthefuture. 18

PAGE 19

1. Byanalyzingmultipledatasetscollectedinrealisticenvironment,theunderstandingswegainshedlightsonfundamentaluserbehaviors,suchaspreferencesandperiodicity.Thecommonalitiesanddierencesfoundfromdierentenvironmentsalsohelpustodistinguishcommonuserbehaviorsfromthespecicsofagivenenvironment. 2. Thendingsfromtheanalysisprovideasetofmoresuitableassumptionstobeusedlaterintheevaluationoftheproposednetworkprotocolsandservices.Inparticular,ithelpstoavoidmakingunrealisticassumptionsintheevaluationstage,sothattheresultscanbemoremeaningful. 3. Theinsightgainedfromtheanalysisoftheenvironmentusuallyprovidesagoodbasistobuildbehavior-awareprotocolsandservices.Forexample,agoodunderstandingofnetworkusagepatternmayleadtoagoodtrendpredictionandabnormalitydetectionservice. 4. Withathoroughanalysisofuserbehaviorfrommultipleenvironments,onecanidentifytheimportantcommonalitiesandbuildtheprotocolswiththesefactsasmajorconsiderations.Inthefollowingsectionweintroducethegenericframework,abbreviatedasTRACE,forthebehavior-awareapproachtowardscomputernetworksinthisdissertation.ThedissertationfeaturesseveralcasestudiesoftheTRACEframework,withdierentfocusesonvarioussub-problemsofcomputernetworkdesign.Theywillbeintroducedinthesubsequentsection. 19

PAGE 20

1-1 .Theindividualstepsareintroducedbelow. Figure1-1. IllustrationoftheTRACEframework. 1. 1 2 ]).PartofthetracecollectioneortisstillongoingattheUniversityofFlorida.Wewillexplainthedetailsaboutthetracesinchapter 3 2. 3. 20

PAGE 21

4. 5. 1-2 .Asshowninthegure,eachcasestudyhastwoseparateparts,theobservationandtheapplication.Theow 21

PAGE 22

Figure1-2. Componentsofthestudy. Werstdiscusstherelatedworkinchapter 2 topositionourworkintheliterature.Inchapter 3 ,weintroducethedatasetsweusethroughthedissertation,addressthestrengthsandtheshortcomingsofthecurrentlyavailabledatasetsinthecommunity,anddiscussabouttheongoingdatacollectioneortatUniversityofFlorida.Wethenmoveontoshowourthreecasestudiesinchapter 4 5 ,and 6 / 7 ,respectively.Finally,weconcludethedissertationinchapter 8 .Chapter 4 focusesontheobservedindividualusermobility(i.e.,themicroscopicuserbehavior)fromthetraces.Westartbydisplayingseveralinterestingcommonobservationsfromthetraces,includingtheon-ouserbehavior,theskewedlocation

PAGE 23

5 .Thecentralquestionweaddressinthischapteriswhetheritispossibletoidentifygroupsofusersfollowingsimilartrendsfromthecollectedtraces,withoutanyassumptionsoftheexistenceofgroupsintheenvironment.Weapplyunsupervisedlearningtechniques(e.g.,clustering)inthischaptertothedatasets,anddisplaythatwithinthelargepopulation(intheorderofthousands),withcarefulselectionofthefeaturesandthedistancemetricbetweenuserswecanclassifyusersbasedonthepreferencesintheirmobilitypatterns.Suchagroupidenticationtechniquehasawide-rangeofapplications,fromnetworkmanagementtointelligentadvertisingtobehavior-awareprotocoldesign.Wechoosebehavior-awareprotocoldesignastheapplicationforthiscasestudyandproposeaprole-castservicewhichtargetsagroupofusersdenedimplicitlybytheirbehaviorpatternsasthedestinationnodes.Thesalientfeatureoftheprole-castserviceisthatthesenderdoesnothavetoknowthereceiver'snetworkidentitieswhensendingthemessages.Wefurthermoveuponeleveltounderstandthemacroscopicstructureofuserinteractioninchapter 6 .Inthischapter,weseektounderstandtheencounterpatterns(i.e.,thepatternofmobilenodesmovingintothecommunicationrangeofeachother)realisticallybyrepresentingtheinformationobtainedfromthetracesasgraphs.Weobservedtheemergenceofaspecialgraphicstructure,knownastheSmallWorld[ 8 ],frommultipletraces.Thissuggestsapotentialcorrespondencebetweentheexistingsocialnetworkstructureandthecommunicationopportunitiesbetweenthemobileusersinthe 23

PAGE 24

5 ,andwenamethisprotocolasCSI,sinceitsanewCommunicationprotocolbasedontheStableyetImplicitstructureinhumanmobilenetworks.Wepresentitsdetaileddesigninchapter 7 1. Overthecourseofseveralyears,wehavecollectedusertracesfromWLANsintheUniversityofSouthernCaliforniaandtheUniversityofFlorida.Withthehelpfromthecorrespondingnetworkadministratorsintheschools,weareabletoobtainextensive,campus-widemeasurementsoftheactivitiesofthecampusnetworkusers.Wehavegainedtheunderstandingofwhatinformationtocollecttomaximizetheusageofthedata.Inthefuture,thecollectedtraceswillbesharedwiththeresearchcommunitythroughourprojectwebsite[ 1 ]. 2. Wehavebuiltarichsetofdierentrepresentationsandanalysistoolstoinvestigatevariousaspectsofthetraces.Asmentionedearlier,thesetoolsrevealvariousbehaviorsofusersinthetrace,rangingfrommicroscopicindividualmobilitytomacroscopicnetwork-wideencounterpatterns. 3. Webuildthetime-variantcommunitymobilitymodelbasedontheinsightgainedbystudyingthetraces.ThismodelprovidesaexibleandscalableplatformonwhichresearcherscansetupawiderangeofscenariosforMANETprotocolandserviceevaluations.Thecodeforthemobilitytracegeneratorisavailableat[ 9 ]. 4. Weproposeamatrixrepresentationbasedonlong-runusermobilitypreferenceandasummarizationtechniquetoextractimportantfeaturesfromthematrix.Thenweconstructadistanceorsimilaritymeasurebetweenusersbasedonthesefeatures.Thedistancemetriccanbeusedtoclassifyusersintodistinctgroups,andtheidenticationofsuchgroupsprovidesusefulinformationforthenetworkadministrators. 5. Weproposetheprole-castingserviceformessagedeliveryinmobilenetworks,anewcommunicationparadigminwhichthepropertiesofindividualusers,insteadofthenetworkidentities,areusedtoidentifythedesireddestinationnodes.Webelievetheprole-castapproachismoresuitablethanthetraditionalidentity-centricapproach,especiallywhenthenetworkishighlydynamic.Theprole-castserviceincorporatestheunderstandinggainedfromdetailedstudiesoftheenvironment,suchasthesimilaritymetricmentionedaboveandtheSmallWorldencounterpatterns,intotheprotocoldesignphase. 24

PAGE 25

2.1 .Wefurtherdiscussexistinganalysesdoneonthecollecteddatasetsinsection 2.2 toputourresearchincontext.Wealsointroducetheworkrelatedtoourtwomajorapplications,mobilitymodelingandmessageforwardingprotocoldesigninthedelaytolerantnetwork(DTN)framework,insections 2.3 and 2.4 respectively. 18 ]andtheidenticationofpower-lawdistributionsinthenodedegreefromnetworktopologytraces[ 19 ].Inbothcases,theworkwasmadepossiblebyextensivecollectionofrelevantdatasets.FortheresearchinMANETs,nodalmobilityisoneofthemajorcomponentstounderstandasthemobilitychangesthenetworkconnectivityandhenceimpactsthesystem-wideperformanceonmanyfronts.Therefore,therehasbeenextensiveeortstocollectusermobilitytracesthroughvariousmethods.Onestraight-forwardwaytoobtainmobilityinformationisthroughcloseobservations[ 22 ]ofthemovingusersorsurveys[ 21 ].Theseapproaches,althoughbenecial,haveseverelimitationsintermsofitsscalability{ifhumaneortisinvolved,itisdiculttorepeatthetracecollectionprocesstoincludealargepopulation. 25

PAGE 26

23 ]andtypicaldailylife(i.e.,city-wideaccesspoints)[ 24 ].Theireorthasbeenfollowedbymanyresearchers,includingBalazinskaandCastro[ 10 ],McNettandVoelker[ 11 ],KotzandHendersonetal.[ 12 13 ],andPapadopoulietal.[ 14 ],eachcollectingWLANtracesfrominfrastructureswithdierentsizesanduserpopulations.Amongtheseeorts,BalazinskaandCastrofocusonWLANusersinthreecorporatebuildings[ 10 ],McNettandVoelkercollectusagetracesspecicallyforhand-helddevices(i.e.,PDAs)[ 11 ],andtheothertracecollections[ 12 { 14 ]areobtainedfromgenericusersonuniversitycampuses.WehavealsocollectedWLANtracesfromUniversityofSouthernCalifornia[ 15 ]andUniversityofFloridacampuses.Mostofthesetracesarecollectedpassively,i.e.,thereisnoneedfortheWLANuserstoactivelyparticipateinreportinganydata.Theaccesspoints(APs)andsometimesotherloggingserversinthenetworkpassivelymonitortheassociationandusageofindividualusers.Theonlyexceptionis[ 11 ]wheretheresearchersinstallreportingsoftwareonthePDAstokeeptrackofallAPsinitscommunicationrange.Thepassivetracecollectionapproachisusuallymorescalableasitdoesnotrequiresoftwareinstallationorproactiveparticipationfromtheusers.Withthismethod,traceswithmorethanthousandsofusersarenotuncommon.TheseeortsleadtorichdatasetstounderstandusermobilityinWLANs,especiallyonuniversitycampusesfromwheremostofthetracesareobtained.InadditiontothelocationinformationrevealedbytheassociationwithAPs,theWLANusageofeachuser(i.e.,theamountoftracsent/received)isusuallyalsologged,wideningthepotentialusageofthetraces.InadditiontoWLANs,otherpossibilitiesarealsoleveragedtocollecttraces.IntheRealityMiningproject[ 16 ],EagleandPentlandprogramthecellphonesofthe 26

PAGE 27

16 ]),henceitsscaleisusuallynotcomparabletothepassivelycollectedWLANtraces.However,itisnote-worthythatlarge-scaleuserlocationtracescollectedfromthecellularphonenetworksdoexist,andasofthedissertationwriting,studiesofcellularuserbehavioralsoemergeintheliterature(e.g.,[ 25 ]).Thedatasets,atthismoment,areonlyavailabletothecellularsystemoperators.Alongadierentline,therearealsoeortsincollectingvehiclemovementtraces,inmostcasesthroughGPSpositioningsystem.OneexampleistheCabSpottingproject[ 17 ]whichlogsthelocationinformationofparticipatingtaxisinthegreaterSanFranciscoarea.Projectsofthisnaturealsorequireactivereportingfromthemonitoredvehicles,hencetheparticipantsareusuallyintheorderofhundreds.Morerecentlythereareseveraltestbedsdeployedtocollectencounterevents(i.e.,whendevicesmoveintothecommunicationrangeofeachother)betweenmovingobjects.Theobjectiveoftheseprojectsistounderstandtheemergenceofcommunicationopportunitiesbetweenthedevicescarriedbymovinghumanbeings.TheHaggleproject[ 26 ]focusesonthescenariosnamedasthepocket-switchnetworks,i.e.,theuserscarryminiaturedevicesequippedwithshort-rangeradiointheirpockets,andthesedeviceslogtheirmutualencountersaspotentialcommunicationopportunities.TheyhavecarriedoutexperimentsinconferencesettingsatINFOCOM2005and2006[ 27 ]andinresearchlabs.ExperimentswithasimilarobjectivearealsoperformedbySuetal.[ 29 ]inauniversitycampussettingandbyLeguayetal.inacollegetown[ 28 ]setting.Thesedatasetscanbeleveraged,forexample,toevaluateroutingprotocolperformancesinDTNsinempiricalenvironments.Whilehumanmobilityhasreceivedrelativelymore 27

PAGE 28

30 ])andwildanimals(e.g.,TurtleNet[ 31 ],ZebraNet[ 32 ],andwhaletracking[ 33 ]).Duetothefastemergenceoftraces,theresearchcommunitytriestoorganizewebsitesforarchivingormaintainingpointerstotherelevanttraces.Thesewebsiteshelptoprovidebetteraccessibilityforresearcherstolocatetheresourcesinthecommunity.Twoprominentexamplesofsuchwebsitesare[ 1 ]and[ 2 ].Mostofthetracesweuseinthisdissertationcanbefoundoneitherwebsite.Inadditiontoutilizingexistingdatasetsinthisdissertation,wehavealsoconductedeortsincollectingdatasetsourselves.WehavebeencollectingWLANtracesfromtheUniversityofSouthernCaliforniasincesummer2005,andpartofthetraceshasbeenmadeavailablethroughtheMobiLibwebsite[ 1 ]establishedbyDr.Helmy.ThisdatasetconsiststhelocationinformationofwirelessusersonUSCcampus,andthenetowinformation(theirusageofthenetwork)

PAGE 29

102 ]).Furthermore,throughthecourseoftheresearchwork,wehaveobtainedagoodcollectionofpointerstoexistingdatasets.Weputthesepointersonlinefortheeaseofreferencesforusandtheresearchcommunityinthefuture.PleaserefertotheMobiLibwebsite[ 1 ]formoredetails.Wehavealsotriedoutothertechniquesforthecollectionofmobilityrelatedinformation.Oneexampleisgivingoutsurveys[ 21 77 ]andaskingpeopleabouttheirmovementpatterns.Thisapproachishelpfulinbuildingtheinsightsaboutmobility,butdoesnotscaleverywell.SomeofmyearlyresearchworkbuiltontopofthesurveysissummarizedintheAppendix. 10 { 13 23 24 ].Inadditiontounderstandingindividualusers,theresearchersalsoconsidertheWLANfromwhichthetraceiscollectedasasystem,observehowusersutilizethesystemasawhole,anddisplaytherelevantstatistics,suchasaveragenumberofusersperaccesspoint(AP),thedistributionofAPpopularity,anduserhandofrequencybetweenaccesspoints.Suchanapproachisnaturalasmostofthecurrenttracecollectioneortsobtainthetracesfromasingleadministrativeentity(inmostcases,fromuniversitycampusnetworks,e.g.,[ 11 { 13 23 ]).Thesestatisticshelptheresearcherstogainunderstandingofuserbehaviors. 29

PAGE 30

3 fordetails)inordertogeneralizethendingsbeyondaspecicenvironment,enablingustodiscussaboutgenericuserbehaviors.Inordertoachievethat,wehavetoapplyappropriatenormalizationstomakethetracescollectedwithdierenttechniquescomparable,andidentifyrelevantmetricsofuserbehaviorsfromdierentcontexts.Withthesetracesavailable,laterresearchworksfocusoncharacterizinguserbehaviorsinwirelessLANs.Oneparticularimportantaspect,asmentionedearlier,isthemobilityofusers.BalazinskaandCastroprovideananalysiswithspecialfocusonusermobility,deningthenotionofhomelocationandtwoquantitativemeasures,persistenceandprevalence,togaugeusermobility[ 10 ].Thisprovidesabroadclassicationofusersbasedonthedegreeoftheirmobility,butdoesnotcompletelydescribetheirdetailedbehavior(e.g.,howtheuserssplittheironlinetimetovariouslocations).Alongthelineofmodelinguserassociationtoaccesspoints(APs),in[ 34 ]theauthorsproposetoclusterAPsbasedonthetimeofpeakuserarrivals.In[ 35 ]thefocusisonthearrivalpatternsofusersatAPsandtheauthorsproposetousetime-varyingPoissonprocessestomodelthearrivalpatterns,andfurtheridentifyclustersofAPsbasedontheparametersinitsarrivalprocess.ThesemodelingeortsfocusmorespecicallyoncapturingthechangesofnumbersofusersassociatedwiththeAPsbymodelingthearrivalanddepartureprocesses,hencetheresultingmodelscapturethedynamicsoftheusersofanaccesspointinaggregation(i.e.,thevariationofthetotalnumberofusersassociatedwithaparticularAP)ratherthanthedynamicsoftheindividualuser.Incontrast,wetakeaholisticviewatmodelingassociationsofindividualusers,andobserveseveralaspects 30

PAGE 31

4 ,weproposeageneralframeworkwhichisappliedtocapturefundamentalaspectsofuserassociationbehaviorsintheWLANtraces(e.g.,Useractiveness,preferencesinassociation,hando,andrepetitiveperiodicpatternsinassociation)thatcanbeusedtobuildmodelsforWLANusers.ThemajorndingsweobservefrommultipleWLANtracesathandsincludetheon-obehavior,theskewedlocationvisitingpreferences,thehand-o,andtheperiodicalre-appearanceofnodes.Weusethesemetricstoprovideacompletedescriptionforusers'mobility-relatedbehaviorsinwirelessnetworks.Bystudyingmultipletracesfromdierentenvironmentscollectedatdierenttimes,weareabletoestablishthatmosttracesdisplaysimilartrends,butthedetailsdierduetodierencesinuserpopulation,environment,time,andmethodologiesoftracecollection.Webelievethatthedesignandevaluationofthenextgenerationwirelessnetworksshouldgohand-in-handwithdeep,insightfulunderstandingoftherealisticenvironmentsinwhichtheywillbedeployedandused.However,theWLANtracesstudiedinthisdissertationdonotprovidedirectlynodalmobilitymodels,astheyrepresentthecombinedeectsofcoarse-grained(i.e.,per-APgranularity)nodalmobility,plustheon-ousagepatternsofthedeviceownersandtheinuencesofwirelesssignalpropagationintheenvironments.Inthatsense,onemayenvisionthatall-encompassingmodelsmaybebuiltbystudyingthetraces.Understandingofsuchrealisticscenariosshedslightsonsometimesfalselytakenassumptionsinover-simpliedrandommobilitymodels(suchasnodesholdingthesameprobabilitytovisitalllocationsorbehavingsimilarlythroughthewholesimulationperiod),andquantiesthedetailedbehaviorsofuserssothatfuturemodelscanincorporatethem.Wewillfurtherdiscussthispointregardingmobilitymodelsinsection 2.3 below.Asasidenote,therealsoexistsanalysisofotheraspectsthantheusermobilitybasedonthecollectedtraces.Forexample,in[ 37 ]theauthorsproposemodelstodescribetracowsgeneratedbyWLANusers.Thispointsoutthewideapplicabilityofthetracesfor 31

PAGE 32

5 weattemptadataminingapproachtounderstandtherelationshipbetweenusersinthelarge-scaledatasets.Morespecically,wedenesimilaritymetricsbetweenindividualusersbasedontheirpreferencesofassociation,andleverageunsupervisedlearningtechniques(i.e.,clustering)toidentifygroupsofcoherentbehaviorfromthediverseuserpopulation.Alongthislineofresearch,thereareonlyafewpreviousworksthatusedataminingtechniquestoclassifyusers.TheearliestexampleisbyTangandBaker[ 24 ],wheretheyclassifyMetricomusersintogroupswithatwostepprocedure.Therststepclassiestheusersbasedonmobility-relatedstatistics,suchasnumberoflocationsvisitedanddistancemoved.Eachgroupidentiedintherststepisfurtherclassiedinthesecondstepbasedtheactivenessoftheuser(i.e.,quantiedbytheeventsgeneratedbytheuser)duringtheday.Anotherexampleweareawareofis[ 38 ],whereKimetal.classifyusersbasedontheperiodicityandthemovementrange.Specically,intheirpaper,usersareclassiedbasedonthedominantperiodsintheirmovement(i.e.,classiedintogroupsthatdisplaystrongdailyorweeklymovementpatterns)andtheirlongestmovementranges.Theyclassifyusersbasedondierentrepresentations,hencetheresultshavedierentinterpretationstoours.Inparticular,theirclassicationofusersisbasedonhigh-levelbehavioralstatistics,whileourclassicationofusersisbasedthene-grainedlocationpreferenceshencemore 32

PAGE 33

39 ]theauthorsapplyclusteringtechniques(K-meansandGaussianmixturemodel)tothetraceoflocationcoordinatesofthesameuseratmanydierenttimeinstantstodiscoversignicantplacesfortheuser,buttheyhavenotfocusedonclassifyingusers.In[ 78 ],theauthorsusethemutualencounterfrequenciesbetweennodestoidentifytheunderlyingcommunities,wherethenotionofacommunityreferstoagroupofnodeswhoremainincontactforlongperiodsoftime.Theclusteringisdonebasedoncommunicationopportunitiesavailablebetweenthedevices,notthesimilaritybetweenuserbehaviors(However,wealsonotethathighsimilarityinnodalmobilitydoesleadtobettercommunicationopportunitybetweenthesimilarnodes,hencethereisacorrelationbetweenthetwoapproaches).Inthisworkwerepresenttheassociationhistoryofeachuserinamatrixform,andutilizesingularvaluedecomposition[ 41 ]toobtaintheassociationfeaturesfromusers.Singularvaluedecomposition(SVD)iswidely-appliedtodiscoverlineartrendsinlargedatasets.Itiscloselyrelatedtoprincipalcomponentanalysis[ 40 ].In[ 42 ],theauthorsutilizePCAtodecomposethetracowmatricesforISPnetworksandunderstandthemajortrendsinthenetworktracowmatrices.OurapplicationofSVDtoindividualuserassociationmatricesissimilarinspirittotheirwork.Notethatitistypicalforpeopletofollowdominantroutinesinlives,henceweexpecttheSVDapproachtobeapplicabletovarioushumanbehavioraldatasets.In[ 16 ],theauthorsalsousePCAtodiscovertrendsinacellphoneusergroup,whichissimilartoouranalysisonindividualusers.Inthisdissertation,inadditiontoanalyzingmuchlargerdatasetsthanthedatasetusedin[ 16 ],wefurtherquantifyusersimilaritybydeningdistancemetricstoclassifywirelessnetworkusersintogroupswithrobustvalidation.Notethatinordertomaketheeigen-behaviorvectorsobtainedfromalluserscomparable,weneedtokeeptheorigin 33

PAGE 34

40 ]wherethemeanofeachdimensionisnotsubtracted.Ithasbeenusedtostudythediversityofspeciesatvarioussites[ 43 ]inbiologyliterature. 6 welookintotheencounterpatterns(i.e.,thepatternsofwirelessdevicesmovingintocommunicationrangeofeachother)oftheusersinthetraces.Weseektounderstandtheglobalstructureoftherelationshipsbetweenusersinthetraces.Specically,weprovidenewperspectivestostudytheWLANtracesbylookingintoencounterdistributionsandutilizingtheSmallWorldtheorytodescribetheencounterrelationshipbetweenusersasagraph.TheSmallWorldgraphmodelisproposedin[ 8 ]andwidelyutilizedtodescribevariousnetworksinmanyareas,suchassocialnetworks,Internettopology,andelectricalpowernetworks[ 44 ].In[ 45 ]theauthorappliedtheconceptofSmallWorldtodeviseacontact-basedresourcediscoveryschemeinwirelessnetworks.TwoprominentfeaturesofSmallWorldgraphsarehighclusteringcoecientscomparabletotheregulargraphsandlowaveragepathlengthscomparabletotherandomgraphs.TheemergenceoftheSmallWorldpropertiesindicatesthereisacorrespondencebetweentheencountersofdevicesinthistracesandthefundamentalsocialrelationshipbetweentheirowners,astheSmallWorldnetworkpropertyisanimportantcharacteristicofthesocialnetworksofhumanbeings.In[ 46 ]BaiandHelmyndthat,undermobilitymodelswithhomogeneousbehaviors(i.e.,Eachnodefollowsexactlythesamemodelwithsomerandomness),eventuallyeachnodeencounterswithallothernodesinthenetwork(i.e.achieving100%encounterratio).However,theempiricalobservationsfromlargeWLANtracesshowverydierentbehaviors,withmostnodesencounteringonlyaverysmallportionofthewholepopulation,duringatimeframeaslongasamonth.Thisobservationindicatesthattheuserpopulationsinlargerenvironments,suchasuniversitycampuses,areactually 34

PAGE 35

2.3 below.Similargraphanalysisofpotentialcommunicationopportunitieshasalsobeendonebasedonstudentclassschedulesinauniversityin[ 47 ].Theauthorscreateagraphinwhicheachstudentisrepresentedasanode,andassumethatstudentsregisteredinthesameclassformlinksbetweenthenodescorrespondingtothesestudents.TheyshowthatthisgraphalsodisplaysSmallWorldproperties.Wemustnote,however,thattheclassregistrationinformationisanindirectindicatorofthephysicallocationsofthestudents,andhencedoesnotdirectlytranslateintoagraphofcommunicationopportunities.Inaddition,theclassscheduledoesnotcapturethemobilitypatternsoutsideoftheclasses.Usingthetracesforactuallocationinformationofthedevices,byourdiscretion,seemstobeabetterinformationsourcetoconstructthecommunicationopportunitiesgraph(albeititiscoarse-grainedlocationinformation). 48 ].Morerecently,mobilityhasbeenutilizedastheenablingfactorformessagedeliveryindelaytolerantnetworks(DTNs[ 4 ]),whereacompletepathfromthesourcenodetothedestinationnodedoesnotexistatanytimeinstant,broadeningthescenariosinwhichcommunicationnetworkscanbeestablished.Itisalsoanimportantsystemvariabletoconsiderforprotocolperformanceanalysis,asitisshownthatdierentunderlyingmobilitymodelschangetheperformanceorderingofvariousMANETrouting 35

PAGE 36

20 ].Thereforedesigninggoodmobilitymodelshasbecomeatopicthatattractssignicantattentionfromcomputernetworkresearchers.Awidevarietyofmobilitymodelsareavailableintheresearchcommunity.See[ 49 50 ]foragoodsurvey.Amongallmobilitymodels,thepopularityofrandommo-bilitymodels(e.g.,randomwalk,randomdirection,andrandomwaypoint)rootsinitssimplicity.Theyarenotonlyeasytogenerate,tuneandscale,butalsoamenabletomathematicalanalysisthatrevealsimportantfundamentalpropertiesinmobility,suchasthestationarynodaldistribution[ 51 ],thehittingtime,themeetingtime[ 52 ],andthemeetingduration[ 53 ].Thesequantitiesinturnenableroutingprotocolanalysistoproduceperformancebounds[ 55 { 57 ].However,randommobilitymodelsarebasedonover-simpliedmovementrules,andaswewillshowinchapter 4 ,theresultingmobilitycharacteristicsareverydierentfromreal-lifescenariosobservedfromtherealtraces.Henceitisdebatablewhetherthendingsunderthesemodelswilldirectlytranslateintoperformancesinreal-worldimplementationsofMANETs.Morerecently,anarrayofsyntheticmobilitymodelsareproposedtoimprovetherealismofthesimplerandommobilitymodels.Morecomplexrulesareintroducedtomakethenodesfollowapopularitydistributionwhenselectingthenextdestination[ 21 ],stayondesignatedpathsformovements[ 59 ],ormoveasagroup[ 58 ].Morevariantsofmobilityrulescanbefoundinvariousmodels[ 49 50 ].Theserulesenrichthescenarioscoveredbythesyntheticmobilitymodels,butatthesametimemaketheoreticaltreatmentofthesemodelsdicult.Inaddition,mostsyntheticmobilitymodelsarestilllimitedtoi.i.d.models(inwhicheverynodebehavesstatisticallythesame),andthemobilitydecisionsarealsoindependentofthecurrentlocationofnodesandtimeofsimulation.Adierentapproachtomobilitymodelingisbyempiricalmobilitytracecollection.Alongthisline,researchershaveexploitedexistingwirelessnetworkinfrastructure,suchaswirelessLANs(e.g.,[ 10 13 ])orcellularphonenetworks(e.g.,[ 16 ]),totrackusermobilitybymonitoringtheirlocations.Suchtracescanbereplayedasinputmobility 36

PAGE 37

60 61 ].Morerecently,DTN-specictestbeds[ 27 29 30 32 33 ]aimatcollectingencountereventsbetweenmobilenodesinsteadoftheactualmobilitypatterns.However,mostoftheseworks(except[ 27 ])donotincludedetailedmathematicalanalysisforthemobilitycharacteristics.Also,duetotheexperimentalnatureofthesestudies,thesizeofthetracesandtheenvironmentsinwhichtheexperimentsareperformedcannotbeadjustedatwillbytheresearchers.Toimprovetheexibilityofthetraces,theapproachoftrace-basedmobilitymodelshavealsobeenproposed[ 62 { 65 ].Basedonthecollectedtraces,thesemodelsdiscovertheunderlyingmovementrulesthatleadtotheobservedproperties(suchasnodaldistribution,durationofstayatlocations,arrivalpatterns,etc.)inthetraces.Statisticalanalysisisthenusedtodetermineproperparametersofthemodeltomatchitwiththetrace.Ideally,agoodmobilitymodelshouldachieveanumberofgoals:(i)itshouldrstcapturerealisticmobilitypatternsofscenariosinwhichonewantstoeventuallyoperatethenetwork;(ii)atthesametimeitisdesirablethatthemodelismathematicallytractable;thisisveryimportanttoallowresearcherstoderiveperformanceboundsandunderstandthelimitationsofvariousprotocolsunderthegivenscenario,asin[ 27 48 56 57 ];(iii)nally,itshouldbeexibleenoughtoprovidequalitativelyandquantitativelydierentmobilitycharacteristicsbychangingsomeparametersofthemodel,yetinarepeatableandscalablemanner;designinganewmobilitymodelforeachexistingornewscenarioisundesirable.Mostexistingmobilitymodelsexcelinoneor,lessoften,twoaspectsoftheaboverequirements,butnonesatisesallofthematthesametime.Themostwidelyusedmobilitymodelsarerandommobilitymodelssuchasrandomwalk,brownianmotion,randomdirection,andrandomwaypoint[ 49 50 ].Theirstrengthisthetheoreticaltractabilitybuttheirweaknessisthelackofrealism.Morecomplicatedsyntheticmo-bilitymodels(e.g.,[ 21 58 59 ])improvetherealism,butmostofthetimeattheexpenseoftheoreticaltractability.Morerecently,alargenumberofempiricalmobilitytracesfrom 37

PAGE 38

10 11 13 16 27 ].Althoughonecanusesuchtracesdirectlyinanevaluationwithexcellentrealism,thesetracesareusuallyratherinexibleandprovideonlyasinglesnapshotoftheunderlyingmobilityprocess.Toaddressthesetwoissues,trace-basedmobilitymodels[ 62 { 65 ]havebeenproposed(i.e.larger,moreexiblesynthetictracescreatedfromthesmallerempiricallycollectedones).Yet,mostofthesemodelsdonotpossessthenecessaryexibilitytomatchmobilitycharacteristicsoftracesotherthantheonesonwhichtheyarebased.Asanapplicationoftheobservationswemakeontheindividualusermobilitycharac-teristicsinchapter 4 ,wecombinethestrengthsofvariousapproachestomobilitymodelingmentionedaboveandproposearealistic,exible,andmathematicallytractablesyntheticmobilitymodel.Large-scaledeploymentsofWLANsinuniversity[ 11 13 ]andcorporate[ 10 ]campusesprovideexcellentplatformsinwhichhugeamountofuserdatacanbecollectedandanalyzed.Weleveragethesetracestounderstandempiricalusermobility,andproposeatime-variantcommunitymobilitymodelbasedontheprominentmobilitycharacteristicsobserved.Wedierentiateourworkandothertrace-basedmodels([ 62 { 65 ])inseveralaspects.First,whilethepreviousworksemphasizethecapabilitytotruthfullyrecreatethemobilitycharacteristicsobservedfromthetraces,wegobeyondthatandemphasize,inadditiontotherealism,themathematicaltractabilityofthemodel.Thisadditionalfeaturefacilitatestheapplicationofourmodeltoperformancepredictionofvariouscommunicationprotocols.Second,weabstracttheobservedmobilitycharacteristicsfromWLANtraces,andproposeamobilitymodelthathaswiderapplicability{inadditiontoWLANs,itcanbetunedtomatchwithothertypesoftraces,suchasavehiclemobilitytrace[ 17 ],andevenwithothercharacteristicsinothertracesofhumanmobility(e.g.,theencounterdurationandtheinter-encountertimein[ 27 ]).Ourtime-variantcommunitymobilitymodel(inshort,theTVCmodel)isbuiltuponourpreviouswork[ 66 ]presentedinsection 4.1 ,inwhichweidentifyseveralprominentpropertiesthatarecommoninmultipleWLANtraces.TheTVCmodelextendsthe 38

PAGE 39

52 ]byintroducingtime-dependentmobilityandhenceinducingperiodicalbehaviorofthenodes.Althoughcapturingtime-dependentbehaviorissuggestedin[ 65 ],ithasnotbeenincorporatedintheirmodel.Amongalleortsofprovidingrealisticmobilitymodels,toourbestknowledge,thisistherstworktoexplicitlycapturetime-variantmobilitycharacteristics.TheTVCmodelpresentedinthisdissertationisageneralizationofthepreviousconferenceversion[ 67 ].Theconceptofcommunityisalsomentionedin[ 68 ]inadierentcontext.Theauthorsassumetheattractionofacommunity(i.e.,ageographicalarea)toamobilenodeisderivedfromthenumberoffriendsofthisnodecurrentlyresidinginthecommunity.Inourpaperweassumethatthenodesfollowlocation-basedpreferencetomakemovementdecisions,andeachnodemovesindependentlyoftheothers.Mobilitymodelswithinter-nodedependencyrequireasolidunderstandingofthesocialnetworkstructure,whichisanimportantareaunderdevelopment.Wechoosetoleavethisasfuturework. 4 ]).InDTNs,packetroutingreliesonnotonlythespatialconnectivity,butalsotemporalchangeofnodalpositions(i.e.,mobility)tobesuccessful[ 69 ].Mostofthepreviousworkinthisareafocusondesigningpacketforwardingheuristics[ 54 56 57 61 71 74 93 94 ].Ingeneral,dierentdegreesofknowledgeofmobilitypatternisassumed[ 54 ],orani.i.d.randommobilitymodelisused[ 56 57 ].Someprotocols(e.g.,[ 61 74 ])seektodiscoverpromisingleadstothedestinationnodebasedonnodal 39

PAGE 40

93 94 ].Ananalysisontheseprotocolsshowsthateachnodehastoencounterwithahighpercentage(i.e.,morethan30%)ofothernodesbeforetheselectedpathsbecomestable[ 46 ].AswediscoverempiricallyfromtheWLANtracesinchapter 6 ,thisisusuallynotachievedinadiverse,large-scaleenvironmentsuchasuniversitycampuses,whereonaverageagivennodeencountersonlyaround6%ofthewholepopulation.Henceitbecomesanissueworthinvestigatingthat(1)wouldmessagedeliverybesuccessfulinsuchasparse(intermsoftheavailableencounterevents)network?and(2)howtodesigngoodmessageforwardingstrategiesinsuchenvironments?Inthisdissertationwetakeabehavior-awareapproachtomessageforwardingintheDTNs.Ingeneral,ourgoalistomaketheforwardingprotocoltobeawareofthebehavioralpatternsoftheindividualuserswhenmakingtheforwardingdecisions,andleveragetheencounterpatternstofacilitatethemessageforwarding.Wesplitthetaskintoseveralcomponentsinthedissertation{(a)weincorporateuserbehavioralpatternsinaDTNmessageforwardingprotocol,designedforaspecialcaseofsendingmessagestouserswhoaresimilartothesenderinchapter 5 ,(b)wediscussaboutunderstandingofglobaluserencounterpatternsinchapter 6 ,and(c)weshowhowtoleveragetheencounterpatternsinprotocoldesignwithmoregenericscenariosinchapter 7 .Inchapter 5 ,weproposeanewserviceparadigmnamedprole-cast.Inprole-cast,thedestinationnode(s)arenotidentiedbytheirnetworkidentities(e.g.,networkaddresses),butbytheiraliationsandbehavioralpatterns(i.e.,theproleofthenode).Prole-castisrelatedtomulti-castasbothofthemtargetgroupsofreceivers.However,inprole-casttheintendedreceiversaredenedbytheirintrinsicproperties,andtherewouldbenoexplicitjointosubscribetoagroupasinmulticast.Managinggroupmembership 40

PAGE 41

70 ]butitisstillahardproblemtosolve.Thegoalofourprole-castserviceistoleverageunderlyingbehavioralpatterns(i.e.,theproles)toguidemessagedelivery,whichtiesnaturallytomanycontext-centricservicesinmobilenetworks,suchassearchingandtargetedannouncement/advertisement.Weleveragemobility-basedprole-castasanexampleinthecasestudypresentedinchapter 5 .Thegoalofourapplicationistodeliveramessagetothenode(s)whohavesimilarmobilityprolesasthesenderitself,withoutknowingtheirnetworkidentitiesbeforehand.Theforwardingprotocolusesthecharacteristicsofnodalmobility,whichwereferredtoasthemobilityprolesofusers,toguidethepropagationofmessagesamongthenodes.Notethatthisapplicationisdierentfromgeo-cast[ 72 ],whichtargetsatthenodescurrentlywithinageographicalregionasthereceivers.Ourtargetreceiversarenodeswithacertainmobilityprole,regardlessoftheiractuallocationsatthetimethemessageissent.Withthecasestudyofmobility-basedprole-castweshowthatunderstandinguserbehavioralpatterncanbehelpfulindesigningroutingprotocolsorservices.Thissuccessisdirectlybasedonthefactthatmobilityprolecanbeusedasadistinguishingfeatureofthemobileusers,asdiscussedinthersthalfofchapter 5 (basedonthedataminingapproachtotraceanalysis).However,thiscasestudyappliesonlytosendingtoagroupwithsimilarmobilityproletothesender,averyspecializedcase.Wewishtofurtherenlargethescopeoftheprole-castparadigmtoincludeothertypesofuserprolesasdescriptorsforpotentialdestinations.Insomecases,thetargetprolescouldbeevenindependentofthenodalmobilitypatterns.Thefore-mentionedgoalleadsusnaturallytotheideaofleveragingtheencounterpatternstodisseminatecopiesofmessagesinthenetworkeciently.Weseekawaytoecientlyspreadthemessagetothewholenetworksothatthepotentialrecipientnodescaneasilyretrieveacopyofthemessage. 41

PAGE 42

6 ,werststartfromanempiricalpointofview,andinvestigatetheissueofwhetherthestore-and-forwardmodelispotentiallyfeasiblewithasimpleforwardingstrategyunderthecurrentencounterpatternsofwirelessdevicesderivedfromtheWLANtraces.Ourinitialndingsareencouraging:Werstusetheepidemicrouting[ 71 ]totestthereachabilityofthenetwork(i.e.,ifthecurrentencounterpatternsleadtoanetworkinwhichmostnodesarereachable).Itturnsout,notonlymostnodesarereachable,buttheencounterpatternsleadtoarobustnetwork{evenifsomenodesareuncooperative,orencounterswithshortdurationsareconsiderednotuseable,messagesstillpropagatewellinthenetwork.ThissuggeststhepossibilityofdesigningalearningprotocoltoidentifynodeswithdierentrolesintheunderlyingSmallWorldencounterpattern,andmakethemessagedisseminationmoreecient(i.e.,reducingthehighoverheadassociatedwiththeepidemicrouting).Inchapter 7 wethendiscussthedesignofthismoregenericmessagedisseminationprotocol.ThedierencebetweentheCSIprotocolsinchapter 7 andthecasestudyinchapter 5 isthefollowing.Insection 5.8 wefocusononlysendingmessagestouserswithsimilarbehavioralproletothesender.InCSIweintroducethenotionofthetargetproletodecouplethebehavioralproleofthesenderfromthedestinationproleinthemessage.Thissignicantlyenhancesthecapabilityofthemessagedisseminationschemes,byallowingthesendertospecifytargetbehavioralprole(inCSI:Tmode),orevensometargetprolesthatareorthogonaltothebehaviorbasedonwhichwemeasurethesimilaritybetweenusers(inCSI:Dmode).InthedesignprocessoftheCSIprotocols,weconducttherstdetailedsystematicstudyonthespatio-temporalstabilityofuserbehaviorsinmobilesocieties,anewdimensionthathasnotbeenconsideredbefore.Oureortontheextractionofbehavioralprolesandbehavior-baseduserclassicationisrelatedtotherealityminingproject[ 16 ]andtheworkofGhoshetal.[ 111 ].Weleveragetherepresentationofmobilitypreferencematrixdenedinchapter 5 ,whichrevealsmoredetaileduserbehaviorthantheve 42

PAGE 43

16 ]andthepresence/absenceencodingvectorusedbyGhoshetal.[ 111 ].Themajorapplicationconsideredinchapter 7 istodesignamessagedisseminationschemeindecentralizedenvironments.Whileseveralpreviousworksexistinthedelaytolerantnetworkeld,mostofthem(e.g.[ 61 71 74 76 107 ])considerone-to-onecommunicationpatternbasedonnetworkidentities.Theprole-castcommunicationparadigmtargetedatabehavioralgroupisanewparadigmindecentralizedenvironments.Someofthepreviousworksassumeexistinginfrastructure:PeopleNet[ 110 ]usesspecializedgeographiczonesforqueriestomeet.Thequeriesaredeliveredtorandomlychosennodesinthecorrespondingzonethroughtheinfrastructure.Others(e.g.,[ 74 107 ])relyonpersistentcontrolmessageexchanges(e.g.,thedeliveryprobability)foreachnodetolearnthestructureofthenetwork,evenwhenthereisnoon-goingtrac.Fromthedesignpointofview,ourapproachdiersfromthembyavoidingsuchpersistentcontrolmessageexchangestoachievebetterenergyeciency,animportantrequirementindecentralizednetworks.ThespiritofourdesignismoresimilartotheworkbyDalyetal.[ 76 ],inwhicheachnodelearnsthestructureofthenetworklocallyandusestheinformationformessageforwardingdecisions.However,thelearningprocessproposedin[ 76 ]stillinvolvesmessageexchangesaboutpastencounters,evenintheabsenceofactualtrac.Ourgoal,ontheotherhand,istodesigntheprotocolsothenodesrelyontheintrinsicbehavioralpatternofindividualusersto\position"themselvesinthebehavioralspaceinalocalizedandfullydistributedmanner,withoutanymessageexchangebetweennodes.TheuseofuserbehavioralprolestounderstandthestructureofthespaceissimilartothemobilityspaceroutingbyLeguayetal.[ 61 ]andtheutility-basedroutingbyAiklasetal.[ 105 ].Themajordierencesbetweenourworkand[ 61 105 ]aretwofold:First,wedesigntheCSI:Dmode,inwhichthetargetproleneednotberelatedtothebehavioralprolebasedonwhichthemessagedisseminationdecisionsaremade.Second,wealsoprovidea 43

PAGE 44

61 105 ]. 44

PAGE 45

13 ],UCSD[ 11 ],andMIT[ 10 ]traceswerecollectedbyotherresearchgroups.WesummarizetheimportantcharacteristicsofthesetracesinTable 3-1 andexplainthemajorissuesbelow.Forcomparisonpurposes,wealsoincludeonetraceofencountersbetweenportablewirelesssensorsdeployedatINFOCOM2005bytheresearchersintheHaggleproject[ 27 ],andonevehiclemobilitytraceobtainedfromtheCab-spottingproject[ 17 ].ForlongertracessuchastheDartmouth[ 13 ]trace,theUSCtrace,andtheUFtrace,wetakepartsofthedatasetsforeachcasestudyinthedissertation.Wemakesuchchoicestofacilitatetheprocessingofdataandfocusouranalysisonsmaller,tractablepartsofthedatasets.Thechosenpartsarerepresentativeofthedatasetasawhole,andsimilarconclusionscanbedrawnifwehadchosenotherpartsofthedataset.Wewillexplainourchoicesfurtherineachcasestudy.AsshowninTable 3-1 ,weusedierenttagstorefertovariouspartsofdatasetsfromthesametrace.Insomecases,weapplymultiple 45

PAGE 46

4 and 6 ,weanalyzeselectedone-monthperiodsfromthelongerDartmouth,USC,UF,andUCSDtraces.FortheUCSDtrace,wechoosetherstmonth,astheuseractivitydecreasedduringtheirstudyduetolossofinterestinparticipationandsomeminorproblemsintracecollection[ 11 ].Weselecttwoone-monthperiodsfromtheDartmouthtrace:July2003(Dart-03,duringthesummervacation)andApril2004(Dart-04,duringthespringquarter).FortheUSCtrace,wepicktherstavailablemonthforthedetailedtrace;fortheUFtrace,wepicktherstmonthofspring2008andrandomlypick10;000usersfromtherelativelylargeuserpopulation(seedescriptionsforUSCandUFtracesinsection 3.4 ).TheMIT,Dartmouth,USC,andUFtracescollectmeasurementsofgenericwirelessnetworkusers,includingbutnotlimitedtolaptops,PDAs,andVoIPdevices.TheUCSDtraceisfromaspecicprojecttostudythebehaviorsofPDAusers.Tofurthercomparetheassociationbehaviorsofsmaller,handhelddevices(e.g.,PDA,VoIPdevices)withgenericwirelessdevicesinthesameenvironment,wealsoseparatethePDA(Dart-PDA)andVoIPdevice(Dart-VoIP)usersfromtheDartmouthtraceduringApril2004,andstudytheirbehaviorspecically.However,accordingtothedevicetypeinformationprovidedinDartmouthtracearchive[ 85 ],thereareonly25PDAusersand63VoIPdeviceusersduringthistimeperiod.Theresultswegetfromthesesmallsamplesizesmayneedtobeveriedbystudiesinlargerscale. 46

PAGE 47

Statisticsofstudiedtraces 1,3661733Jul.20'02toWLAN3EngineerPollingWholetrace1,366MIT-consMIT[ 82 ]Aug.17'02GenericbuildingsMIT-rel 10,296623188Event-basedJul.20032,518Dart-03 WLANApr.20045,582Dart-04GenericUniversityDart-relApr.'01tocampusDart-cons Jun.'04PDAonly25Dart-PDA VoIPdevice63Dart-VoIP WLAN04/05/04-06/04/046,582Dart-04springDartmouth[ 81 ]Generic(Springquarter04) 275518N/ASep.22'02toPDAonlyUniversityPollingSep.22'02to275UCSDUCSD[ 83 ]Dec.8'02campusOct.21'02 4,54879N/AEvent-basedApr.20,'05to4,528USCportsDec03-Dec04(trap)WLANUniversityMay.19'05 25,481137Apr2005-Now(detail)Genericcampus01/25/06-04/28/065,000USC-06springUSC[ 80 ]ports(Springsemester06) UF44,751728N/AAugust2007toGenericUniversityEvent-basedJan.14,'08to10,000UFNowcampusFeb.13,'08 41internalN/AN/AMar.7'05toiMOTEconferencePollingWholetrace41Cambridge-INFOCOM05Cambridge[ 84 ]nodesMar.10'05 549taxisN/AN/ASep.22'06toGPSdevicesGreaterSanPollingWholetrace549Vehicle-traceCab-spotting[ 17 ]Nov.1'06installedontaxisFranciscoarea

PAGE 48

5 and 7 ,weuselongertracestostudythetrendsinuserassociationpreferencesandhowwecouldutilizesuchinformationtoclassifyusers.Weusetwosemester-longtracesfromthelongerDartmouthandUSCtraces,sinceasemester(oraquarter)isthetypicallongesttimeperiodforwhichthebehaviorofusersonuniversitycampusesremainconsistent.ForDartmouthwepickthespringquarterof2004,whichincludestheactivitiesof6;582usersduring61days.ForUSCwepickthespringsemesterof2006,whichincludestheactivitiesof25;481usersduring94days.Tomakethetaskmanageable,weanalyzeonlythemostactive5;000usersfortheUSCtrace.InordertocomparetheWLANuserbehaviorswiththebehaviorsofusersinotherenvironments,wealsoincludetwoadditionaltracesinthestudy.TheCambridge-INFOCOMtraceiscollectedfromparticipantsofanexperimentcarriedoutatINFOCOM2005conference[ 27 ].Eachparticipantisgivenawirelesssensor(anInteliMote)andaskedtocarrythisdevicethroughouttheconference.ThesedeviceskeeptrackofotherBluetoothdevices(includingall41iMOTEsdeployedfortheexperimentandotherBluetoothdevicesintheconferencevanueaswell)andrecordthedevicesthatarewithinitscommunicationrangeatregularintervals.Inotherwords,thistracecollectstheencountereventsbetweenthewirelessdevices.Theothertrace,Cab-spotting,isavailablefromtheCab-spottingproject[ 17 ].ThisisaprojectfortrackingthelocationsofparticipatingtaxisinthegreaterSanFranciscoareausingGPSdevicesinstalledonthe 48

PAGE 49

4 ,thesetwoseeminglyverydierentgroupssharesomesimilarmobilitycharacteristics.Inchapter 6 ,wecomparetheencounterpatterns(seeitsdenitionbelowinsection 3.3 )betweentheWLANuserswiththeusersintheencountertracecollectedatINFOCOM2005. 82 ])orassociationtrackingsoftwareontheMNs(intheUCSDtrace[ 83 ]),and(ii)Event-basedmethodswhichrecordMNonline/oineeventsusingloggingserver(e.g.syslog)[ 80 81 ].FortheDartmouthtraceweusethederivedassociationhistorytracefromtheirsyslogtrace[ 81 ],andforUSCtracethelogsarecollectedfromtheswitch(i.e.,theswitchcreatesalogwhenaMNassociates/disassociateswithoneoftheAPsconnectedtooneoftheswitchports).FortheUFtrace,thelogsarecollectedatboththeAPlevel(foruserassociationstoAPs)andtheauthenticationserverlevel(forusersign-inandsign-out).Itisgenerallyacceptedthattheevent-basedapproachprovidesmoreaccuraterecordsofMNassociationwiththeAPsinthenetwork.However,thereisnoin-depthstudytoquantifythedierencesbetweenthesetwoapproaches.Inordertofurtherunderstandtheeectsofdierentmethodsoftracecollectionontheuserbehaviormetricsobtainedfromthetraces,wealsocreateanemulatedpollingtraceasfollows:Foranevent-basedtrace,weobservethetraceatregulartimeintervalsandemulatewhatwouldberecordedifthetraceweretakenbypolling-basedmethod.Wethenprocesstheemulatedpollingtraceaswedotoanormalpolling-basedtrace,andcomparethendingswiththeoriginalevent-basedtrace.WeusetheApril2004Dartmouthtrace(Dart-04)tocarryoutthisexperiment,obtaining 49

PAGE 50

50

PAGE 51

Illustrationofthetermdenitions. AP,respectively.FortheUCSDtracethepollingintervalis20seconds.WeuseonlytherelaxedapproachtoprocesstheUCSDtrace. 3-1 )wewilluseinthesubsequentchapters.Weusethenotiononline(orinshort,on)torefertothestatewhenaMNisassociatedwithanyAPinthenetwork(orequivalently,itscurrentlocationisknown,oritis\present"atthemoment).Onthecontrary,thenotionoine(orinshort,o)referstothestateofaMNbeingabsentatthemoment(i.e.,itisnotassociatedwithanyAPcurrently).Relatedtotheabovedenitions,anonlineeventisdenedasaMNstartinganewassociationtoanyAPfromtheostate.AnoineeventisdenedasaMNdisassociatingitselffromthecurrentAPandchangingtoostate(i.e.,itdoesnotroamtootherAPs,butgoesoinedirectly).AhandoeventoraroamingeventisdenedasaMNchangingitsassociationfromoneAPtoanotherwithnooinetimeinbetween.Anassociationsessionisdenedasthedurationbetweenanonlineeventtothenextoineevent.Therecanbemanyhandoeventswithinoneassociationsession.Thetotalonlinetimeisthesumofthelengthsofassociationsessions(i.e.thesumofthe\shaded"intervalsinFig. 3-1 ),andtheexistencetimeisthetimedierencebetweenaMN'srstonlineeventanditslastoineeventinthestudiedtrace.WeusetheexistencetimeasaconservativemeasureofthetimedurationforwhichtheMNisapotentialuserofthenetwork,sincetheMNsarenotalwaysonlineandtheuserpopulationcanchangewithrespecttotimeonuniversitycampuses(for 51

PAGE 52

52

PAGE 53

27 29 30 32 33 ]). 53

PAGE 54

54

PAGE 55

4.1 .Weidentifyseveralmajormobilitycharacteristics,suchasthenodesnotbeingalwayson,theskewedlocationvisitingpreferencesandtheperiodicalre-appearanceofnodesatthesamelocation,asprominentmobilitycharacteristicsfromthetraceswestudy.Then,weproposeamobilitymodelinsection 4.2 ,whichisexibletocapturethespatialandtemporaldependencyoftheprominentmobilitycharacteristicsweobserveempiricallyfromthetraces.Inaddition,thismodelisalsomathematicallytractable,henceitfacilitatestheoreticalanalysisofprotocolsinmobilenetworks. 4.1.1IntroductionRecently,wirelessnetworkshavebeendeployedubiquitouslyinvariousenvironments,especiallyinuniversitycampusesandcorporations,andgainedpopularityrapidly.Withmoreusersswitchingtowirelessnetworks,theimportanceofunderstandinguserbehaviorinsuchenvironmentsisbecomingclearer.FromthevastamountofwirelessLAN(WLAN)tracesavailabletotheresearchcommunity,onecanobtainimportantandfundamentalknowledgeaboutitsusers.Amongthevastspaceforpotentialinvestigation,wefocusonthefollowingquestion:Howdowerealisticallymodeluser

PAGE 56

4.2 .Additionally,analysisofuserbehaviorandnetworkusagepatternsenablesaccurateassessmentofwirelessnetworkutilizationandaidsthedevelopmentofbettermanagementtechniquesandcapacityplanningdecisions.Asnewtechnologiesevolve(e.g.,variantsof802:11WLANs,oradhocnetworks),fundamentalunderstandingofuserbehaviorbecomesessentialforthesuccessfuldeploymentofsuchemergingtechnologies.SeveralstudieshavebeenpreviouslyconductedontheanalysisofWLANtraces[ 10 ],[ 11 ],[ 13 ],andweborrowfromthesetracesandstudies.Thesestudiesarequitehelpful,buteachofthemisbasedonasinglecampuswithadierentfocus,andhenceitbecomesunclearwhethertheirndingsgeneralizebeyondthestudiedcampus.Inourstudy,wegobeyondpreviousworktocompareuserbehavioracrossdierenttraces,andtrytoobservethegeneraltrendsandquantifythedetaileddierencesamongthem.WelookintotheaspectsthatweconsiderimportanttomodeluserbehaviorinWLANs,andreasonaboutthecommonalitiesanddierencesoftheseaspectsbetweencampuses.Forthemetricswestudy,wendthatingeneralmostofthecampusesfollowsimilartrends,suchas(1)Mostnodesdisplayon-ousagepattern.Theyareoinefornon-negligibleamountoftime,andswitchbetweenonlineandoinestatesoften.ThisfactislargelyoverlookedbypreviousresearchesonmodelingWLANusersalthoughitisanomnipresentphenomenonfromall 56

PAGE 57

4.2 .Asoneexpects,thedetailsoftheseuserbehaviormetricsdependontheunderlyingcampusenvironmentanduserdevicetype;wewillcommentonthendingsthroughoutthesection.Inadditiontothat,inthisworkwealsocomparetwodierenttracecollectionmethodologies,polling-based(e.g.SNMP)andevent-based(e.g.syslog).WeshowthedierencebetweenthesetwotracecollectionmethodsbygeneratinganemulatedSNMPtraceinpost-processingfromsyslogtraces(whichhavebettertimeresolutionthanSNMPtraces),andcomparethedierencesamongthetwotraces.Sometimes,majordierencescanbeattributedtodierenttracecollectionmethodsused.Thissuggeststheneedforastandardmethodologyfortracecollectiontomakedatafromdierentenvironmentscomparable. 57

PAGE 58

IllustrationofaMN'sassociationpatternwithrespecttotimeoftheday. associationbehaviorsinaWLAN.WeshalluseFig. 4-1 toillustrate.OnecouldseetheassociationpatternofaMNasasequenceofassociatedAPs(shownbyshadesinFig. 4-1 ),potentiallywithtimesegmentsduringwhichtheMNisoine(e.g.notassociatedwithanyAP)betweenassociations.Welookintofourmajorcategoriestounderstanduserbehaviorasfollows:(a)Activenessofusers:Thiscategorycapturesthetendencyofausertobeonline(i.e.,HowactivelytheusershowsupinWLAN?).Ingeneralwirelessnetworkusersarenotalwayson,butshowupinthetraceintermittently,asopposedtothealways-onnodesassumedinthesyntheticmodels.(b)Macro-levelmobilityofusers:ThiscategorycaptureshowwidelyaMNmovesinthenetworkinthelongrun(i.e.,forthewholetraceduration),andhowitsonlinetimeisdistributedamongtheAPs.Theintentionistocaptureoveralllong-runstatisticsandpreferenceofaMNvisitingAPs.(i.e.,HowaretheshadesdistributedinFig. 4-1 ?DoweneedmanydierentintensitiesofshadesforeachuserasitassociateswithmanyAPs?Canwendafew\dominant"APsforeachMN?)(c)Micro-levelmobilityofusers:ThiscategorycaptureshowMNsmoveinthenetworkwhileitremainsassociatedwithsomeAP(i.e.,hando).TheintentionhereistocapturethemobilityofaMNwhileusingthewirelessnetwork,adierentobjectivefromthemacro-levelmobility.(i.e.,HowoftendoestheMNchangeassociationswithoutleavingthenetwork?) 58

PAGE 59

CCDFofonlinetimefraction (d)Repetitiveassociationpatternofusers:Thiscategorycapturestheuserassociationbehaviorwithrespecttotime.Weexpectuserstoshowrepetitivestructureinassociationpatternsduringsimilartimesofdierentdays,astheirmobilitypatternsaredictatedbytheirdailyschedule.ThisideaisalsoillustratedinFig. 4-1 :theuserappearsatAP1duringlateeveningsinbothdays.Weproposethenetworksimilarityindex(NSI)asametrictoquantifythetendencyofuserstoshowrepetitivepatternsintheirassociations.

PAGE 60

4-2 .FromFig. 4-2 weobservethatinalltracesonlyasmallportionofusersarealwaysoneventhoughbydenitiontheuseractivenessisalreadyover-estimated,exceptfortheDart-04trace.Theaverageonlinetimefractionis87:68%forDart-04trace,andbetween36:44%(Dart-03)and14:12%(UCSD)forothertraces.Thestandarddeviationforonlinetimefractionislarge,varyingfrom0:24to0:36foralltraces.Theseobservationsarguestronglythatusershaveon-ousagepatterns,wheresomeoftheusersareheavyusers(withhighontime)whilemanyarelightusers.Thedistributionsoftheon/otimesseemtodependheavilyontheenvironments(i.e.,campus)andthedevicetypesinthetraces.UCSDtrace,whichfocusedonlyonPDAusers,istheleastactiveoneamongalltraces.Theothertraces(MIT,USC,Dart-03)arenotverydierentinonlinetimefractiondistribution.TheactivenessofMNsincreasesignicantlyfrom2003to2004inDartmouthtrace,whichagreeswiththendingsin[ 13 ].BycomparingthecurvesofDart-04,Dart-rel,andDart-cons,weobservethatonlinetimefractionisconsistentforthesametraceunderdierenttracecollection(ortracereconstruction)methods.ComparisonbetweenDart-04withDart-PDAandDart-VoIPshowsthatduringthesametraceperiod,thehandhelddevicesarelessactivethantheaverageofthetotalpopulation.However,handhelddevicesintheDartmouthtracearemuchmoreactivethantheUCSDtrace,butthereasonisnotclearatthispointandwarrantsfurtherinvestigation.WealsocheckwhetherthesignicantlyhigheronlinetimefractioninDart-04traceiscausedbyuserswithonlyoneshortassociationsession(henceitsonlinetimefractionisover-estimatedbyourdenition).ItturnsoutthatthehighonlinetimefractioninDart-04traceiscausedbysignicantincreaseofalways-onusers.InDart-04trace,thereare27:5%ofusersthatinitiateonlyoneassociationsessionwhichlastsforthedurationof30days,thewholetraceperiod.ThesamenumberforDart-03traceislessthan0:04%.Therearetwopossiblereasonsfortheverydierentbehaviorinthetwotimeperiods.(1)July2003wasduringsummervacation,hencetheactivitywassignicantlylower, 60

PAGE 61

CCDFofnumberofassociationsessionsbyusers or(2)ThewaypeopleuseWLANhaschangedbetweenthesetwotraceperiodattheDartmouthCollege.UsersinDart-04tracetendtousewirelessLANasareplacementforwirednetwork,andkeeptheirdeviceassociatedwithWLAN,insteadofestablishingtheconnectiononlywhenitisneeded.Ifthelaterspeculationistrue,asweseethisparadigmshiftfromusingWLANasatemporaryconnectiontoanalways-on,permanentconnection,itispossiblethattheonlinetimefractionwillalsoincreasesignicantlyforotherdeployments.WefurthercomparetheCCDFofthenumberofassociationsessionsgeneratedbyusersinthesetracesinFig. 4-3 .WeobservethatthePDAusersintheUCSDtracegeneratemoreassociationsessionsthanusersinothertraces(exceptMIT-contrace,explainedbelow),whichincludegenericwirelessnetworkdevices(mainlylaptopusers)duringcomparabletraceduration.Thisfact,togetherwiththelessonlinetimefractioninFig. 4-2 ,indicatesthattheUCSDPDAusersaremorelikelytousethedevicesforshorterbutmorefrequentsessions.However,thisobservationdoesnotapplytothehandhelddevicesinDartmouth.BothPDAsandVoIPdevicesinitiatelesssessionsthanthegeneraldevicesinDartmouth.Fromthegurewealsoobservethatcountofassociationsessionsissensitivetothetracecollectionmethod.Theemulated 61

PAGE 62

4-3 ,weseethattheemulatedpollingtracesobserveonlyonefthofsessionsfortheMNwiththelargestnumberofsessions(200versus1000).Anothertechnicaldicultyhereistoadequatelytranslateasampleseeninthepolling-basedtracestothedurationofassociationappropriately,aswendthecurvesofMIT-consandMIT-reldrasticallydierent.AcloserinvestigationintotheMITtracerevealsthatalthoughSNMPpollingintervalsaretypically5minutes,sometimesrecordsofMNassociationareobtainedatlongerintervals,leadingtobogusterminationsandre-initiationofassociationsessionsiftheconservativeassumptionisusedandhencethehighassociationsessioncountsshownbythecurveMIT-cons. 62

PAGE 63

CCDFofcoverageofusers. WedenethecoverageofauserasthepercentageofAPsonthecampustheuserassociateswithduringthetraceperiod.FortheUSCtraceweuseswitchportsinplaceofAPs.ThedistributionsofthecoverageofusersinthetracesareshowninFig. 4-4 .Thismetriccaptureshowwidelyausermovesforthewholeperiodoftraceinthestudiednetworkenvironments.Weobservethatusershavesmallcoverageinallenvironments.Theaveragecoverageisbetween4:52%(UCSD)and1:10%(Dart-cons/rel).Noneofthesetraceshasevenasingleuservisitingmorethan35%ofallAPs.IntheUCSDtrace,thePDAusersseemlikelytovisitalargerportionofcampusthanthegenericusersdointheothercampus-widetraces,duetotheportabilityofPDAs.SimilarobservationappliestotheVoIPdevicesintheDartmouthtrace,whichisthemostmobilesub-usergroupintheDartmouthtrace.However,PDAsintheDartmouthtracearelessmobilethanthegenericusersintheperiodwestudied.Wesuspectthattheresultmaybeinuencedbyafewextremeusers(thereareonly25PDAusersidentiedduringthisperiod,andhalfofthemvisitonly4orlessAPs).TheMITtraceiscollectedfromonlythreebuildings,hencetherelativecoverageofusersisabithigher.Itisimportanttonotethatthecoverageseemstoremainstablewithrespecttotimechange,althoughtheactivenessofuserschangessignicantly(compareDart-03andDart-04).Thecoverageissensitive

PAGE 64

4-5 .Fromthegureweobservethatforallenvironments,thegeneraltrendisthateachuserhasveryfewAPsatwhichitspendsmostofitsonlinetime.Inparticular,forallthetraces,aMNspendsonaveragemorethan65%ofitsonlinetimewithoneAP,andmorethan95%ofonlinetimeatasfewasthetop-5APscombined.Theleft-endofthecurvesaresimilar,butthetailsvary.ThehighermobilityoftheUCSDPDAuserstranslatesintoalongertail,whereinadditiontothosefewmostvisitedAPs,theusersalsoaccessthewirelessnetworkatmuchmorelocationswithasmallfractionoftheuser'sonlinetimeascomparedtoothertraces.SimilarobservationsapplytoDart-VoIPandDart-PDAtraces.ItisinterestingwhyDart-PDAtraceshowssmallcoverageinFig. 4-4 buthighaveragefractionoftimeassociatedwithlesspopularAPshere.Thesetwopoints,however,donotcontradicteachother.Acloserinvestigationrevealsthatalthoughthereareasmallfractionofwidely-visitedPDAs(fromFig. 4-4 ),thosewhovisitmanyAPscontributemoreoftheironlinetimetolesspopularAPs.Thismetricisrobusttodierenttracecollectionmethodsandassumptionsoftracepost-processing,asthecurveforDart-04isclosetoDart-consorDart-rel.SimilarobservationsaremadefortheMITtrace. 64

PAGE 65

AveragefractionoftimeaMNassociatedwithAPs.ForeachMN,theAPlistissortedbasedonassociationtimebeforetakingtheaverageacrossusers. totheprevioussection:Howmobiletheuseriswhileusingthenetwork.Weusehandostatisticsasameasureofusermobilitywhileusingthenetwork.However,aftertheinvestigationofthehandostatistics,wediscoveralotofhandoeventsareduetoso-called\ping-pongeect"ratherthanrealmovements.Theterm\ping-pongeect"referstothephenomenonofexcessivehandoeventsduetodisturbanceinwirelesschannelswhiletheMNitselfmightbestationary.Hence,wecannotdirectlylinkthehandostatisticstothemicro-levelmobilityoftheusers.Developmentofbetterltersforping-pongeectsisneededbeforewecanreallyunderstandthemicro-levelmobilityformtheWLANtraces.FirstweshowtheCCDFcurvesforthetotalhandoeventcountduringthewholetraceperiodinFig. 4-6 .Ourrstintuitionisthatusermobilityshouldbedependentonthedevicetype,andhandhelddevicesshoulddisplayhighermobilitythanusersinothertraces.ThisistruefortheDart-VoIPtrace,astheVoIPdeviceshavethemostper-userhandocountamongalltraces.However,thePDAsinbothUCSDandDartmouthtracedonothavemorehandoeventsthanothertraces.FortheUCSDtrace,thismayberelatedtothefactthatthePDAsareusuallyusedforshortsessions,hencetheyexperiencelesshandoevents.FortheDart-PDAtrace,someofthePDAsareonlinefor 65

PAGE 66

CCDFoftotalhandocountperMN. longdurations,buttheydonothavemanyhandoevents.Thereasonisnotclearatthispoint.FromFig. 4-6 weobservethattheexactnumberofhandocountdependsheavilyonthenetworkenvironment(e.g.,thedeploymentoftheAPs,etc).IntheUSCtrace,thecoarselocationgranularitydirectlyleadstothelowerhandocounts.Ontheotherhand,theDartmouthtraceshavemuchmorehandoeventsthanothertraces.WealsoobservethatthehandocountsinFig. 4-6 aresensitivetothetracecollectionmethod,asthecurveforDart-04dierssignicantlyfromDart-relandDart-cons.Thisisagainbecausethepolling-basedmethodoverlooksquickchangesofuserassociationsbetweenpollingintervalsandhencemanyhandoeventsarenotcaptured.Inadditiontotheabove,wealsoobservethatforallthetraces,handocountsvarysignicantlyamongtheusers-Therearesomeuserswithmanyhandoeventsandsomewithfew.Tobetterunderstandthecauseofhandoevents,welookintotherelationshipbetweensessionlengthsandhandoeventsinthesessionforeachtrace.Asanexample,weshowascatterplotforsessionlengths(inminutes)andhandocountsforallsessionsintheUSCtraceinFig. 4-7 .Fromthegraph,weseethatthereisnocleartrendbetweenthesessionlengthsandthehandocounts.Insomecases,weseeextremelylongsessions 66

PAGE 67

Scatterplot:SessiondurationsversushandocountinthesessionfortheUSCtrace. withoutanyhandoevents,orextremelymanyhandoeventsinasessionwithshortduration.Thecorrelationcoecientsbetweensessionlengthsandhandocountsforallthestudiedtracesarebetween0:377and0:030.Sowecanseethatthesessionlengthandthehandocounthaveaweaklinearcorrelationtoeachotherinalltraces.Wefurtherlookintothefollowingstatisticstoobservewhetherthesessionswithhighhandocountsareallfromasmallsetofextremelymobileusers:Foreachuser,wecalculatetheaveragehandoeventperunittime(i.e.thehandorate)foreachofitssessions,andthencalculatethemeanandvariancefortheuser'shandoratefromallthesessionstheuserinitiates.Ifahighdegreeofmobilityleadingtothehighhandocountisanintrinsicpropertyforsomeusers,weshouldseethatthoseusersshowhighaverageandlowvarianceintheirhandorates.Weusethecoecientofvariation(thestandarddeviationdividedbythemean)tounderstandthedegreeofvariationinthehandoratesforusers.InFig. 4-8 ,weshowtheCDFofthecoecientofvariationofthehandorateforthestudiedtraces.Onlytheuserswithmorethanonesessionandonehandoeventareconsideredinthegraph,sinceuserswithonlyonesessionautomaticallyresultin0varianceforitshandorate.Fromthegure,weseethatthehandoratedisplayshighvarianceformostoftheusers.Inalltraces,morethan60%ofusershaveits 67

PAGE 68

CDFforthecoecientofvariationofthehandorateoftheusers.Notethatforalltraces,coecientofvariationislargerthan1:0foratleast60%ofMNswithmorethanonesession. coecientofvariationofthehandoratelargerthan1:0(i.e.,thestandarddeviationbeinglargerthanthemean).ThisindicatesevenforagivenMN,handoratevariesdrasticallyfromsessiontosession.Combiningtheobservationsintheprecedingparagraphs,weconcludethathandoeventsnotonlydistributeunevenlybetweenusers,butalsohappenunevenlybetweenthesessionsforthesameuser.ThisindicatesthatthehandoeventsaregreatlyinuencedbytheenvironmentalconditionwhenasessionisestablishedratherthanthepropertyoftheMNwhoinitiatesthesession.WeevenobservethatsomeMNshavehundreds,sometimeseventhousands,handoeventsbetweenlessthan5APswithinasession.Suchascenarioismuchmorelikelyduetoping-pongeectratherthantrueusermobility.Thereductionofping-pongeectisanimportantissuetomakebetterinterpretationaboutthemicro-levelusermobilityfromtheWLANtracesandwarrantsfurtherstudy. 68

PAGE 69

4-9 weshowtheNSIforallthetraces.Toseethedetailsbetter,wesplitthegureintotwoparts:curveswithsmallerabsoluteNSIvaluesareshowninFig. 4-9 (a),andcurveswithbiggerabsoluteNSIvaluesareshowninFig. 4-9 (b).WewilldiscussthephysicalmeaningoftheabsolutevalueofNSIlaterinthissection.FromFig. 4-9 (a),weseethatinmostofthesetraces(i.e.,USC,MIT,Dart-03)weobservenoticeablyhighernetworksimilarityindexifthetimegapisclosetointegermultiplesofaday.Thisisanindicationthatusershavethestrongesttendencytoshowrepetitiveassociationpatternatthesametimeofeachday.Itisalsointerestingtoobservethatforthesetraces,thenetworksimilarityindexforthegapof7days(i.e.,aweek)isthesecondhighest,onlyslightlylowerthanthatforthegapof1day.Thisindicatesweeklyrepetitivepatternisalsostronginthesetraces.Ontheotherhand,theUCSDtraceshowslittlerepetitivepatternasthereisalmostnoobviousspikesinitsNSIcurve.ThiscanbeattributedtoitsuserpopulationbeingPDAusers.Unlikelaptops,whicharemorerelatedtowork,PDAsareusuallyused 69

PAGE 70

(b)Figure4-9. Networksimilarityindexes.Thepeaksrepresentintervalsforwhichthereishighsimilarity.(a)NSIcurveswithsmallerabsolutevalues(lessalways-on,stationaryusers),(b)NSIcurveswithlargerabsolutevalues(morealways-on,stationaryusers). inamorecasualwayinshort,scattereddurations.HenceitisexpectedthatPDAusersshowlessrepetitivenessintheirusagepattern. 70

PAGE 71

4-9 (b),theNSIcurvesfortheDart-04traceoritssub-groupsofusers 4.1.2.1 ).Inthe2004trace,wehavemorealways-on,stationaryusersusingWLANasareplacementofwirednetworks.ThisisreectedbythehigheraveragevalueoftheNSIcurves,indicatinglargerfractionofusersalwaysstayatthesamelocation.ThismaybeattributedtothefactthatDartmouthtracesincludeusersinstudentdormitories,whicharemainlystationaryusersandhavecontributedtohighlocationsimilarityindexes.WefurthercomparetheNSIcurveoftheDart-04traceinFig. 4-9 (b)totheNSIcurveofDart-04-March(onlyusedinthisexperiment)inFig. 4-9 (a).FortheDartmouthCollege,themonthofMarchcontainsthespringbreak,whensomeofthestationaryusersindormsareabsent,andweseethattheperiodicityofassociationbehaviorismorevisibleintheMarchtrace.Fromtheaboveexperiment,wearguethattheperiodicbehaviorintheaverageNSIcurvecomesfromnon-stationaryusers(e.g.,thosewhocometoworkorclassesduringdaytimeandfollowaregularschedule),notthestationaryuserswhouseWLANsasareplacementofwiredLANs.Thispointispartlysupportedbythendingsin[ 36 ]:Mostusersdisplayingperiodicityinassociationhavehomelocationsatacademicbuildings.TheUSChasnotdeployedWLANindormitoriesyet(fortheone-mothtracewehavechosenforthisstudy),andtheMITtraceismainlyfocusedonbuildingsforwork.Thatmaybethereasonwhyperiodicassociationbehaviorsaremoreobviousinthosetraces.

PAGE 72

10 ],[ 13 ],[ 11 ]onanalysesofWLANtracesbyconsideringtracesfrommultiplecampusesandmultipleaspectstomodeluserbehavior.WealsomakeourownWLANtracescollectedatUSCcampusavailableat[ 1 ],togetherwithmanypointerstoexistingWLANtracearchives.Tosummarize,thendingsfromthetracespointoutimportantcommonfeaturesinallstudiedenvironments.Wirelessnetworkusersinuniversitycampusesandcorporatenetworkarecharacterizedby(1)limitednumberofvisitedAPsinthenetworkandalargeproportionofonlinetimespentatveryfewofitsmostvisitedAPs.Thecoverageofusersneverexceeds35%inalltraces,andusersspendmorethan95%oftheironlinetimewithasfewasveAPs.Furthermore,thesenumbersseemtoremainrelativelystableinagivenenvironment,eveniftheWLANgainspopularityandusersbecomemoreactive.(2)Periodicassociationpatternswithstrongdaily/weeklypattern.Webelievethatthesemetricscaptureimportantcharacteristicsaboutusersinwirelessnetworksthatarelargelyoverlookedbyearlierworkonmobilitymodelingandwirelessnetworksimulation.(3) 72

PAGE 73

4.2.1IntroductionInthemobileadhocnetworks(MANETs)[ 3 ],asthedevicesareusuallyeasilyportableandthescenariosofdeploymentareinherentlydynamic,mobilitybecomesoneofitskeycharacteristics.IthasbeenshownthatmobilityimpactsMANETsinmultipleways,suchasnetworkcapacity[ 48 ],routingperformance[ 20 ],andclustermaintenance[ 79 ].Inshort,theevaluationofprotocolsandservicesforMANETsseemstobeinseparablefromtheunderlyingmobilitymodels.Itis,thus,ofcrucialimportancetohavesuitablemobilitymodelsasthefoundationforthestudyofadhocnetworks. 73

PAGE 74

66 ].Asweshowintheprevioussection,oneofthesalientcharacteristicsislocationpreference.IntheTVCmodel,weextendtheconceptofcommunitiesfrom[ 52 ]toserveaspopularlocationsforthenodes.Anotherimportantcharacteristicisthetime-dependent,periodicalbehaviorofmanynodes.Tocapturethis,weimplementtimeperiodsinwhichthenodesmovedierently[ 67 ].Toourbestknowledge,thisistherstsyntheticmobilitymodelthatcapturesnon-homogeneousbehaviorinbothspaceandtime.Inadditiontotheimprovedrealism,theTVCmodelcanbemathematicallytreatedtoderiveanalyticalexpressionsforimportantquantitiesofinterest,suchasthenodalspatialdistribution,theaveragenodedegree,thehittingtime(timerequiredforamobilenodetohitarandomlyselectedcoordinate)andthemeetingtime(timerequiredfortwomobilenodestocomewithincommunicationrangeofeachother).Thesequantitiesareoftenfundamentaltotheoreticallystudyissuessuchasroutingperformance,capacity,connectivity,etc.Weshowthatourtheoreticalderivationsareaccuratethroughsimulationcaseswithawiderangeofparametersets,andadditionallyprovideexamplesofhowourtheorycouldbeutilizedinactualprotocoldesign.ToestablishtheexibilityofourTVCmodelwealsoshowthatwecanmatchitstwoprominentproperties,locationvisitingpreferencesandperiodicalre-appearance,withmultipleWLANtraces,collectedfromenvironmentssuchasuniversitycampuses[ 80 81 ]andcorporatebuildings[ 82 ].Moreinterestingly,althoughwemotivatetheTVCmodelwiththeobservationsmadeonWLANtraces,ourmodelisgenericenoughtohavewiderapplicability.WevalidatethisclaimbyexamplesofmatchingourTVCmodelwithtwoadditionalmobilitytraces:avehiclemobilitytrace[ 17 ]andahumanencountertrace[ 84 ].Intheformercase,weobservethatlocationvisitingpreferencesandperiodical

PAGE 75

4 ])protocols.Despitethesecharacteristicsarenotexplicitlyincorporatedinourmodelbyitsconstruction,theycanbestillrealisticallyreproduced.Toourbestknowledge,thisistherstsyntheticmobilitymodelproposedthatmatchesmeasurementsets(traces)collectedfrommultiplescenarios,andhasalsobeentheoreticallytreatedtotheextentpresentedhere.Duetoitsstrengthsinbothexibilityandtheoreticaltractability,theTVCmodelhastwomajorapplications:togeneraterealisticmobilitypatternsunderawiderangeofdierentscenariosandtofacilitateperformanceanalysisandprediction.WealsomakethecodeoftheTVCmodelavailableat[ 9 ].Inthissection,werstre-iteratethemobilitycharacteristicswediscoveredfromthetraces,discusshowweconstructamobilitymodeltocapturethem,andthenformallyintroduceourTVCmodelinsection 4.2.2 .Then,inSection 4.2.3 ,weembarktopresentourtheoreticalframeworkandderivegenericexpressionsofvariousquantitiesundertheTVCmodel.TheaccuracyoftheseexpressionsisvalidatedagainstsimulationsinSection 4.2.4 .FinallyweshowthetwomajorapplicationsoftheTVCmodel:inSection 4.2.5 ,weshowhowtogeneraterealisticmobilityscenarioswithmatchingmobilitycharacteristicsinvarioustraces;inSection 4.2.6 ,wemotivateourtheoreticalframeworkfurther,byapplyingouranalysistoprovideguidelinesandperformancepredictionsinprotocoldesign. 75

PAGE 76

1 ]or[ 2 ]).ThereasonforthischoiceisthatWLANtracesloginformationregardinglargenumbersofnodes,andthusaremorereliableforstatisticalanalysis.Afteranalyzingalargenumberoftraces,wehaveobservedtwoimportantpropertiesthatseemtoberecurrentinallofthem:skewedlocationvisitingpreferencesandtime-dependentmobilitybehavior[ 66 ].First,bylocationvisitingpreferencewemeantheamountoftimethatanodespendsassociatedwithagivenaccesspoint(AP).InFig. 4-10 (a)wecalculateforvarioustracesthefractionofthetotalonlinetimeanaveragenodespendswithitsmostfavoriteAP,itssecondfavoriteAP,etc.,uptoitsleastfavoriteAP.(ThisisessentiallytheprobabilitydensityfunctionoftheassociationtimeofanodewithanAP,withtheAPssortedindescendingorderoftotalassociationtime.)Itisclearfromtheplotthatanodeonaveragespendsmorethan65%ofitsonlinetimeassociatedwithitsfavoriteAP,andmorethan95%ofitsonlinetimeatonlyveAPs.Werefertothisbehaviorbysayingthatthelocationvisitingpreference(orinshort\locationpreference")ofnodesisskewed.Second,bytime-dependentmobilitybehaviorwerefertothefactthatnodestendtobehavedierentlyduringdierenttimesoftheday(orevenduringdierentdays),andmostspecicallytoexhibitsomeamountofperiodicityintermsofthelocationstheychoosetovisit.InFig. 4-10 (b)weplottheprobabilityofanodeappearinginthesamelocationatsometimeinthefuture,asafunctionofthedierenceintime.ItisevidentfromtheplotthatnodesappearatthesameAPwithahigherprobabilityafteratime-gapofintegermultiplesofdays.Thiscreatesthesaw-toothpatterninthecurves.Aslightly 76

PAGE 77

(b)Figure4-10. TwoimportantmobilityfeaturesobservedfromWLANtraces.(a)Skewedlocationvisitingpreferences.(b)Periodicalre-appearanceatthesamelocation.Labelsoftracesused:MIT:tracefrom[ 82 ],Dart:tracefrom[ 81 ],UCSD:tracefrom[ 83 ],USC:tracefrom[ 80 ]. strongerweeklycorrelationcouldalsobeobservedinsomeplots(seeforexampletheslightlylargepeakintheMITcurveforatimegapofsevendays).Itisthusclearthatnodesbehavedierentlyindierentperiodsintime,andthatsimilarbehaviorstendberepeatedonadailybasis.Unfortunately,mostexistingmobilitymodelsfailtocapturethesetwoproperties.Forsimplerandommodels,likerandomdirection,randomwaypoint,randomwalk,etc.,thereisobviouslynopreferenceinbothspaceandtime.ThisisdemonstratedinFig. 4-10 byastraightline(uniformdistribution)fortheRandomDirectionmodelfortherespectiveprobabilities.Evenformoresophisticatedmodelsthattrytocaptureotheraspectsofmobility,suchasgroupmobilityintheRPGMmodel[ 58 ]oramodelconsideringobstaclesandpathways[ 59 ],thesetwopropertieswouldalsobestraightlinesintheplotsasspatialandtemporalpreferenceisnotapartofthesemodels 21 52 62 { 64 ])thataimatcapturingspatialpreferenceexplicitly.Anexampleofsuchamodelisthesimplercommunitymodelof[ 52 ].AsisshowninFig. 4-10 (a),withappropriatelyassignedparametersthismodelisabletocapturetheskewed

PAGE 78

4-10 (b).Itisourgoaltodesignamobilitymodelthatsuccessfullycapturesbothofthesetwoproperties,observedinthemajorityoftraces.OnecouldarguethatapotentialshortcomingofthisapproachisthatWLANtracesdonotregistercontinuousmovementofthedevices,butratherassociationsofusers/nodeswithspecicAPs.Whatismore,somedevicesarenotalwayson,andtypicallytherearesomegapsinthecoverageofaccesspointsinthesenetworks.However,webelievethatthetwomainpropertiesweobserved,namelyskewedlocationpreferenceandtime-dependency,areprevalentinreal-lifemobility.Thisbeliefisfurthersupportedbyobservingtypicaldailyactivitiesofhumans:mostofustendtospendmostofthetimeatahandfuloffrequentlyvisitedlocations,andarecurrentdailyorweeklyscheduleisaninseparablepartofourlives.Asaresult,amodelsupportinglocation-preferenceandtime-dependentmobilityshouldbeabletocapturehumanmobilityinmanycontexts,ifcarefullydesigned.Comparisonwithnon-WLANtracesinSection 4.2.5 conrmsourargument. 78

PAGE 79

Parametersofthetime-variantcommunitymobilitymodelForallparameters,wefollowtheconventionthatthesubscriptofaquantityrepresentsitscommunityindex,andthesuperscriptrepresentsthetimeperiodindex. vMinimum,maximum,andaveragespeed6Dmax;j; DjMaximumandaveragepausetimeaftereachepoch6 4-11 .Wealsousethisexampletointroducethenotationsweuse(seeTable 4-1 )intherestofthesection.Asshownin 79

PAGE 80

4-11 ),oronecommunitycanevencontaintheother(asinTP2inFig. 4-11 ).Finally,thenumberofcommunitiesineachtimeperiodmayvary.Forexample,thereare3communitiesintotalintherstperiod,2inthesecondone,and4inthethird 4-11 orFig. 4-12 )thatcorrespondstoadailyorweeklyschedule.Wenowdescribehowanodemovesinsidetheaboveconstruction.Nodemovementconsistsofasequenceofepochs.EachoftheseepochsisaRandomDirectionmovement 4-11 ).8

PAGE 81

Illustrationofagenericscenarioofthetime-variantmobilitymodel,withthreetimeperiodsanddierentnumbersofcommunitiesineachtimeperiod. Figure4-12. Anillustrationofasimpleweeklyschedule,whereweusetimeperiod1(TP1)tocaptureweekdayworkinghour,TP2tocapturenighttime,andTP3tocaptureweekenddaytime. chosendistance;attheendoftheepoch,thenodepicksapausetimerandomlyandthenproceedstothenextepoch.ThedierencebetweenourcommunitymodelandtheRandomDirectionmodelisthatinadditiontoallotherparameters,inthecommunitymodelthenodealsochoosesrandomlythecommunityinwhichthenextepochwillbeperformed.Thatis,withprobabilityjthenextepochtakesplaceinsidethenode'sj-thcommunity(PStj=1j=1),ratherthanmovingaroundthewholesimulationarearandomly,asinthestandardRandomDirectionmodel.(Notethatweusuallyaddsuperscripttinthenotation,i.e.tj,todenotethattheseprobabilitiesmightchangebetweentimeperiods.)Wesaythenodeisinstatejwhenithasanepochinthej-thcommunity.Further,toensurethatalocalmoveiscompatiblewiththelocalcommunitysize,wealsoscalethelocalepochlengthbydrawingitfromanexponentialdistributionwithaveragelength 81

PAGE 82

4-11 ).Werefertosuchepochsasroamingepochs.Finally,afteranepoch,anodepausesforatimeuniformlychosenin[0;Dmax;j],wherethemaximumpausetimeisagaindependentonthecommunity.Asanalnote,onemayarguethatcapturinglocationpreferenceandtime-dependenciescouldplausiblybeachievedwithamobilitymodelconstructedwithdierentwaysthantheonewepropose.However,ourchoicesarelargelyguidedbythefactthatmostofthebuildingblocksweutilizetocreateourmobilitymodel(e.g.randomdirectionepochs,communities,etc.)areeasytounderstand,andhavebeenshowntobeamenabletotheoreticalanalysis[ 52 ].ThebenetswillbecomeevidentinSection 4.2.3 .Atthesametime,wewillalsoshowinsection 4.2.5 thatthesechoicesdonotcompromiseourmodel'sabilitytoaccuratelycapturereallifemobilityscenarios.

PAGE 83

86 87 ].Finally,wederivetheexpectedhittingandmeetingtimesforourmodel.Thehittingtimeisthetimeittakesanode,startingfromthestationarydistribution,tomovewithintransmissionrangeofaxed,randomlychosentargetcoordinateinthesimulationeld.Themeetingtimeisthetimeuntiltwomobilenodes,bothstartingfromthestationarydistribution,moveintothetransmissionrangeofeachother.ThesetwoquantitiesareofinterestduetotheircloserelationshiptotheperformanceofDTNroutingprotocols,oringeneraltheperformanceofprocessesthatrelyonnodeencounters.Knowingthemeetingtimeforamobilitymodelis,forexample,crucialwhenusinga\mobility-assisted"or\store-carry-and-forward"protocoltodeliveramessage[ 55 { 57 ],whilehittingtimesmightbeneededifsomenodesinthenetworkarestatic(e.g.sensors,basestations,etc.).WenotethatapreliminaryversionofsomeofthetheoreticalderivationspresentedhereappearforaspecialcaseofourTVCmodelonlyin[ 67 ](thatmodelincludedonecommunityandtwotimeperiodsonly).Here,wegeneralizeallderivationsforanycommunityandtime-periodstructure.Moreover,wepresentsomeadditionalresultsregardingthespatialdistributionandtheaveragenodedegreethatarerelevanttovariouswirelesscommunicationprotocols,asweshowinSection 4.2.6 .Westartwithausefullemmathatcalculatestheprobabilityofanodetoresideinaparticularstate. vtj)=StXk=1tk( vtk+ 83

PAGE 84

vtk+ vtj)andthepausepart( Notethattheabovestationaryprobabilitiescanbecalculatedforeachtimeperiodandnodeseparately.WeusePtj(i)todenotetheprobabilitythatnodeiisinstatejduringtimeperiodt(i.e.,Ptj(i)=Ptmove;j(i)+Ptpause;j(i)). 84

PAGE 85

Proof. FollowingthesameprincipleinLemma 4.3 ,weincludeallcommunitypairsandarriveatthefollowingTheorem. 4{6 )issimplyaweightedaverageofthenodedegreeofnodeaconditioningonitsstates.ForeachstatewithprobabilityPtj(a),theexpectednodedegreeisasum 85

PAGE 86

4{6 )degeneratestoX8bX8Commtk(b)Ptk(b)K2 Similartothenodalspatialdistribution,theaveragenodedegreecanbecalculatedforeachtimeperiodseparately. 86

PAGE 87

52 ]thatthelatterprovidesagoodapproximationforthecontinuousversion,andiseasiertoanalyzeforourpurposes.(3)Theexpectedhittingprobabilityforawholetimeperiod,PH,isthencalculatedforeachsub-casefromtheunit-stepprobability,byassuming\hitting"occursindependentlyineachtimestep 4-11 ,thecommunitiesareoverlapped,henceifthetargetiswithinComm21itmustbewithinComm22. 52 ]tobeagoodapproximation,whentheexpectedlengthofanepochisintheorderofthesquarerootoftheareaofthecommunitytheepochtakesplacein.

PAGE 88

4{8 ),weneedtocalculateP(w1;:::;wV)andHT(w1;:::;wV)foreachpossiblesub-case(w1;:::;wV). Proof. TherststepforcalculatingHT(w1;:::;wV)istoderivetheunit-timehittingprobabilityintimeperiodtundertargetcoordinate-communityrelationshipwt,denotedasPth(wt). vtj=Ctj2;(4{10)whereI()istheindicatorfunction. Proof. vjinunittime.Sinceanodefollowingrandomdirectionmovementsvisitstheareaitmovesabout

PAGE 89

vj=Ctj2[ 52 ].Hencethecontributiontotheunit-timehittingprobabilitybymovementsmadeinstatejisPtmove;j2K vtj=Ctj2,i.e.,whenthenodemovesincommunityjandthetargetisinthenewlycoveredareainthetimeunit. Notethatthemovementmadeineachtimeunitdoesnotincreaseordecreasetheprobabilityofhittingthetargetinthesubsequenttimeunits,thereforeeachtimeunitcanbeconsideredasanindependentBernoullitrialwithsuccessprobabilitygiveninEq.( 4{10 ).Thecorollarybelowimmediatelyfollows.

PAGE 90

4{13 ),sinceineachcycleoftimeperiodsfollowsthesamerepetitivestructure.TherstterminEq.( 4{14 )correspondstotheexpecteddurationoffulltimeperiodcyclesuntilthehittingeventoccurs.SinceforeachcyclethesuccessprobabilityofhittingthetargetisP,inexpectationittakes1=Pcyclestohitthetarget,andthereare1=P1fullcycles.ThesecondterminEq.( 4{14 )isthesumofdurationoftimeperiodsbeforethetimeperiodtinwhichthehittingeventoccursinthelastcycle.Finally,thethirdtermisthefractionofthelasttimeperiodbeforethehittingeventoccurs.Notethatthelastpartisanapproximationwhichholdsifthetimeperiodsweconsideraremuchlongerthanunit-time. 4.10 .SimilartoLemma 4.8 ,weaddupthecontributionstothemeetingprobabilityfromallcommunitypairsfromnodeaandbinthefollowingLemma.

PAGE 91

v A(atj;btk)A(atj;btk) v A(atj;btk)A(atj;btk) v A(atj;btk)A(atj;btk) 4{15 )consistsoftwoparts:(I)Bothofthenodesaremovingwithintheoverlappedarea.ThisaddstherstterminEq.( 4{15 )tothemeetingprobability.Thetworatios,A(atj;btk) v A(atj;btk),whichreectsthecoveredareainunittime.Weusethefactthatwhenbothnodesmoveaccordingtotherandomdirectionmodel,onecancalculatetheeective(extra)areacoveredbyassumingthatonenodeisstatic,andtheotherismovingwiththe(higher)relativespeedbetweenthetwo.Thisdierenceiscapturewiththemultiplicativefactor^v[ 52 ].(II)Onenodeismovingintheoverlappedarea,andtheotheronepauseswithinthearea.ThisaddstheremainingtwotermsinEq.( 4{15 )totheunit-timemeetingprobability.Thesetermsfollowsimilarrationaleasthepreviousone,withthedierencethatnowonlyonenodeismoving.Thesecondtermcorrespondstothecasewhennodeamoves(andbisstatic),andthethirdtermcorrespondstothecasewhennodebmoves(andaisstatic).Thederivationoftheunit-timemeetingprobabilitybetweennodesaandbfortimeperiodtincludesallpossiblescenariosofcommunityoverlap.IfnodeahasSt(a)communitiesandnodebhasSt(b)communities,therecanbeatmostSt(a)St(b)community-overlappingscenariosintimeperiodt. 91

PAGE 92

4{15 )isthegeneralformofEquation(13)and(14)in[ 67 ].Ifweassumeperfectoverlapandasinglecommunityfrombothnodes,wearriveat(14).Ifweassumenooverlap,weresultin(13).Alsonoteinthegeneralexpressionspresentedinthiswork,thewholesimulationareaisalsoconsideredasacommunity.Thereforewedonothavetoincludeaseparatetermtocapturetheroamingepochs. 67 ].TheLemmaisre-producedbelowforcompleteness. 4-13 ,whenamobilenodemoveswithinitscommunity,theareacoveredbythenode(i.e.,theareathatcouldfallinthecommunicationrangeofthenode)actuallyextendsoutofthecommunitybythetransmissionrangeofthenode.Hence,the\footage"ofthecommunityislargerthanCtj2.Weapproximatethisareaby(Ctj+2K)2,ignoringthesmalldierencesatthecorners.Finally,sinceeachnodeselectsitscommunityatrandomwithinthesimulationarea,theprobabilitythatpartofthefootageofthecommunityofnodeaischosenaspartofthecommunityofnodebissimply(Ctj(a)+2K)2 Finally,similarlytoTheorem 4.10 ,theexpectedmeetingtimecanbecalculatedusingtheresultsintheLemmasinthissection. 92

PAGE 93

Illustrationoftheexpansionofthe\footage"ofcommunity. Proof. 4.10 andisomitted. 4-2 .Table 4-2 (a)liststheparametersweuseforasimpliedmodel(twotimeperiodswithtwocommunitiesineachtimeperiod,whereoneofthecommunitiesisthewholesimulationeld).Formorecomplexmodels,wetryoutthesetupoftieredcommunitiesandmultiplerandomly

PAGE 94

(b)Figure4-14. IllustrationofthecommunitysetupforthegenericcasesofTVCmodel.(a)Concentricmultiple-tiercommunitiessetting.(b)Multiplerandomlyplacedcommunitiessetting. 4-14 (a),arandomlychosenpointinthesimulationeldservesasthecenterofthecommunities,andmultipletiersofcommunitieswithdierentsizessharethesamecenter.Thisconstructionissuggestedbyacommonobservationfromourdailylives:Peoplevisitthevicinityareaoflocationsthatbearimportancetothemmoreoftenthanroamfaraway.Whenweassignthetieredcommunitystructure,itnaturallymakessensetohavethenodevisittheoutertierslessfrequentlythantheinnertiers,althoughthisisnotrequiredforthetheoreticalderivation.Inthesimulations,weusetwoalternativetimeperiodswithatwo-tierlocalcommunityineachtimeperiod,andtheparametersarelistedinTable 4-2 (b).Inthemultiplerandomlyplacedcommunitieslayout,asillustratedinFig. 4-14 (b),multiplecommunitiesareinstantiatedrandomlytoshowthatourtheoryisnotlimitedtoasinglecommunity.Weusetwotimeperiodswithtworandomlyplacedcommunitieseachforthisscenario.Otherthanthedierenceincommunitysetupandsizes,weagainusetheparametersinTable 4-2 (b)forthiscase.Ourdiscrete-timesimulatoriswritteninC++,andnodesmoveasdescribedinSection 4.2.2 .Moredetailsaboutthesimulator,aswellasthesimulatorcode,canbefoundat[ 9 ]. 94

PAGE 95

ParametersforthescenariosinthesimulationCommonparameters:Forsimplicity,weusethesamemovementspeedforallnode:vmax=15andvmin=5inallscenarios.Inallcasesweusetwotimeperiodsandtheyarenamedastimeperiod1and2forconsistency.Inthesimplemodelweuseasinglelocalcommunity(withsubscriptl)ineachtimeperiod.Forthegenericmodel,wetestwithtwodierentcongurations:(1)Atwo-tiercommunityineachtimeperiod-inthisscenariotheinnertiercommunityandtheoutertiercommunityhasedgelengthCl1andCl2,respectively.(2)Tworandomlyplacedcommunitiesineachtimeperiod-inthisscenariothecommunitiesbothhaveedgelengthCl1,buttheparameterscorrespondtothetwocommunitiesaredierent(i.e.,correspondtosubscriptl1andl2inthetable).Inallcases,thereisalsoaroamingstate(withsubscriptr)inwhichthenodemovesaboutthewholesimulationarea(i.e.thewholesimulationareaisacommunity). (a)Thesimplemodel. ModelnameDescriptionNC1lC2lDmax;lDmax;r (b)Thegenericmodel. ModelnameNC1l1C1l2C2l1C2l2D1max;l1D1max;l2D2max;rD2max;l1D2max;l2D2max;r Modelname

PAGE 96

(b) (c)Figure4-15. Spatialdistributionofthenode(shownastheprobabilityforanodetoappearineach50x50gridblock).(a)Randomlyplacedcommunity.(b)Single-tiercommunitycenteredat(300;300)or(700;700)withonehalfprobability.(c)Two-tiercommunitycenteredat(300;300)or(700;700)withonehalfprobability. 4-15 (a).Theminordiscrepancyisduetothenitenumberofsamples.Tomakethescenariomoreinteresting,wealsogeneratethespatialdistributionfornodeswhenthecommunitiesarexed.WeusetheparametersetsofModel-1(onecommunityineachtimeperiod)andModel-5(two-tiercommunityineachtimeperiod)fromtable 4-2 ,andassignthecenterofthecommunityateither(300;300)or(700;700)withonehalfprobability.TheresultingnodalspatialdistributionsareshowninFig. 4-15 (b)and(c),respectively.Thenodeappearswithhigherprobabilitywherethecommunitiesareassigned.FromEq.( 4{3 ),forthescenarioinFig. 4-15 (b),thenodeappearsinthecommunitywithprobability0:0864andinotherareawithprobability0:0008,respectively.ForthescenarioinFig. 4-15 (c),thenodeappearsintherst-tiercommunity,thesecond-tiercommunity,andtheother 96

PAGE 97

(b) (c)Figure4-16. Comparisonoftheoreticalandsimulationresults(theaveragenodedegree).(a)Randomlyplacedcommunity.(b)Relativeerrorforscenarioswithxedsingle-tiercommunity.(c)Relativeerrorforscenarioswithxedtwo-tiercommunities. areawithprobability0:0759,0:0039,and0:0004,respectively.Inbothcasesthesimulationresultsfollowthetheoreticalresultsreasonablywell,withinabout10%errorfortheareainthecommunities. 4.5 ,whenthecommunitiesarerandomlychosen,theaveragenodedegreeturnsouttobetheaveragenumberofnodesfallinginthecommunicationrangeofagivennode,asifallnodesareuniformlydistributed.Hencetheaveragenodedegreedoesnotdependontheexactchoicesofcommunitysetup(i.e.single,multiple,ormulti-tiercommunities)orothermobilityparameters.InFig. 4-16 (a),wecomparetheevolutionofthetheoreticalaveragenodedegreeversusthecommunicationrange(K)tothesimulationresultsforsomeofthemodelslistedinTable 4-2 .Thesimulationcurvesfollowthepredictionofthetheorywell.Othercongurationswetried(notlistedhere)alsoshowsimilartrends.Again,tomakethescenarioabitmorerealistic,wesimulatesomemorescenarioswhenthecommunitiesarexed.Amongthe50nodes,wemake25ofthempickthe 97

PAGE 98

4-2 .Models1through4correspondtoscenarioswithsingle-tiercommunitiesineachtimeperiod,andmodels5through7correspondtoscenarioswithmulti-tiercommunities.Weshowtherelativeerrors,calculatedasError=(TheorySimulation)=Simulation,inFig. 4-16 (b)and(c).Apositiveerrorindicatesthetheoreticalvalueislargerthanthesimulationresult,whileanegativeerrorindicatestheconverse.Inthesimulations,whenthecommunicationrangesaresmallascomparedtotheedgeofthecommunities,therelativeerrorsarelow,typicallybelow10%exceptforModel-3,indicatingagoodmatchbetweenthetheoryandthesimulation.However,asthecommunicationrangeincreases,theareacoveredbythecommunicationdiskbecomescomparabletothesizeofthecommunityandEq.( 4{5 )isnolongeraccuratesincethecommunicationdiskextendsoutoftheoverlappedareainmostcases.Thatisthereasonforthediscrepanciesbetweenthetheoryandsimulation.BesidesModel-3,weobserveatmost20%ofrelativeerrorwhenthecommunicationdiskislessthan20%thesizeoftheinner-mostcommunity,indicatingthatourtheoryisvalidwhenthecommunicationrangeisrelativelysmall. 4-2 ,andcomparetheaverageresultswiththetheoreticalvaluesderivedfromthecorrespondingequations(i.e.( 4{8 )and( 4{18 )).Tondoutthehittingorthemeetingtime,wemovethenodesinthesimulatorindenitelyuntiltheyhitthetargetormeetwitheachother,respectively.AgainweshowtherelativeerrorsbetweenthetheoreticalvaluesandthesimulationresultsforvariousscenariosinFig. 4-17 .Weseethatforallthescenarios,therelativeerrorsarewithinacceptablerange.Theseresultsdisplaytheaccuracyofourtheoryunderawiderangeofparametersettings.Theabsolutevaluesfortheerrorarewithin16%forthehittingtimeandwithin20%forthemeetingtime.Formorethan70%ofthetested 98

PAGE 99

(b) (c) (d) (e) (f)Figure4-17. Relativeerrorbetweentheoreticalandsimulationresults(thehittingtimeandthemeetingtime).(a)Hittingtime,simplemodel.(b)Hittingtime,multi-tiercommunities.(c)Hittingtime,multiplerandomcommunities.(d)Meetingtime,simplemodel.(e)Meetingtime,multi-tiercommunities.(f)Meetingtime,multiplerandomcommunities. scenarios,theerrorisbelow10%.Theerrorsbetweenthetheoreticalandsimulationresultsaremainlyduetosomeoftheapproximationswemadeinthevariousderivations.Forexample,thereexistsomebordereectswithrespecttothehittingandmeetingprobabilitieswithinacommunity.Whenanodeisclosetotheborderofacommunity,itcouldalso\see"someothernodesoutsideofthecommunityifitstransmissionrangeislargeenough.However,wehavechosentoignoresuchoccurrencestokeepouranalysissimpler.Furthermore,theapproximationofthehittingandmeetingprocesseswithdiscreteBernoullitrialsisvalidonlyfortheepochsthatarelargeenough(intheorderofcommunitysize).Nevertheless,asshowninthegures,theerrorsarealwayswithinacceptableranges,justifyingoursimplifyingassumptions. 99

PAGE 100

9 ].Thetoolprovidesmobilitytracesinbothns-2[ 88 ]compatibleformatandtime-location(i.e.,(t;x;y))format.Inthissection,ouraimistwofold:(i)rst,wewouldliketodemonstratethemodel'sexibilityandhowitcanbeconguredtogeneratemobilityinstancesthatarerepresentativeofvarioustargetwirelessnetworkssuchasWLANs,VANETs,etc.;(ii)atthesametime,wewouldliketovalidatethemodel's\realism"or\accuracy"byexplicitlycomparingmobilityinstancesproducedbyourmodelwithrealmobilityinstancescapturedinwell-known,publicly-availabletraces.However,itisimportanttonotethattheuseofsuchamodelisnotmerelytomatchitwithanyspecictraceinstanceavailable;thisisonlydoneforvalidationandcalibrationpurposes.Rather,thegoalistobeabletoreproduceamuchlargerrangeofrealisticmobilityinstancesthanasingletracecanprovide.Werstoutlineheresomegeneralguidelinesabouthowtousethemodelinordertoconstructspecicmobilityscenarios.Then,weshowhowtoexplicitlyconguretheTVCmodelinordertomatchthemobilitycharacteristicsobservedinthreecasestudies:awirelessLANtrace,avehiculartrace,andahuman-encountertrace.Alltheparametervaluesweuseintheexamplesinthissectionarealsoavailableat[ 9 ].STEP1:Determinethestructureinspaceandtime:TherststeptoconstructtheTVCmodelforagivenscenarioistosetupthecommunitiesandthetimeperiodstructure.Ifthemapofthetargetenvironmentisavailable,oneshouldobservethemap 100

PAGE 101

4-10 )asguidelinestoassignthestructureinspaceandtime.Forexample,fromlocationpreferencecurvesliketheoneinFig. 4-10 (a),onecandeterminethenumberofcommunitiesoneneedstoexplicitlycreate;asaverysimpleexample,ifinmostWLANtracesitisobservedthatthetypicalnodespendssay95%ofitstimeataround2to5preferredlocations(dependingonthenode),thenonecouldassigneachnodetohavefrom2to5localcommunitiesinthenetwork(withtheactualnumberandlocationsofcommunitiesrandomlychosenforeachnode),withalarger(roaming)communityrepresentingtherestofthe5%ofmobilitytime 4-10 (b),onemayobservethere-appearanceperiodicityanddecideonthetimeperiodstructureaccordingly.Ifanertimegranularityisnecessary(e.g.time-of-day)onecouldadditionallyobservethemobilitycharacteristics(e.g.locationpreferences)onanhour-by-hourbasisandidentifyclearchangesinanode'sdailybehavior.STEP2:Assigncommunity-relatedparameters:Ideally,foragivenenvironment,oncethecommunitiesareidentied,therelatedparameters(e.g.,tj; Dtj; Ltj,whichrepresenttheprobability,averagepausetime,andaverageepochlength,atcommunityjduringtimeperiodt)couldbeassignedaccordingtothemobilenodes'behaviorineachcommunity(e.g.,howlongdoesatypicalpersonspendatthecafeteriaforlunch?).Nevertheless,inmostcasessuchinformationisnotavailable,orextremelydiculttoobtain.Hence,onecouldagainresorttomeasuredstatisticsfromtypicaltracestoguide

PAGE 102

4-10 (a).WecalculatethesecommunitystayprobabilitieslaterinLemma 4.1 .STEP3:Assignuseron-obehavior:NotethatthemobilitytracegeneratedbytheTVCmodelisan\always-on"mobilitytrajectoryofthemobilenode(i.e.,thenodeisalwayspresentsomewhereinthesimulationeld).Dependingonthetargetenvironment,thisalways-onbehaviormaynotberealistic.Inmanyempiricallycollectedtraces,notallnodesarepresentallthetime(i.e.,someofthenodesare\o"ornotintheobservedareasometimes).Thisisthecaseintwoofthescenarioswediscussbelow-intheWLANtraces,nodesare\on"onlywhentheyarenotmoving;inthevehiclemobilitytrace,nodesare\on"whentheyaremoving.Thus,beforeproducingthenalsynthetictrace,assumptionsaboutwhentheuserisconsidered\on"shouldsometimesbemadeandsuperimposedtotheTVCmobilitytracesgenerated.Wehaveappliedthissteptothetracesofthetwoscenariosmentionedabove.Next,welookintothreespeciccasestudies,namelyasetofWLANtraces,avehiculartrace,andatraceofinter-nodeencounters.Weshowhowtoapplythefore-mentionedprocedureineachcase,andshowthatsyntheticmobilitytracesproducedbytheTVCmodelsuccessfullymatchthecharacteristicsobservedintherealtraces. 102

PAGE 103

4-10 (a))tocalculatetheattractionfromthecommunities(tj)andthepausetimesofthenode( 4-10 (b)representtheprobabilityofan\on"nodeassociatedwiththesamecommunityafterthegiventimegap,andthepeaksappearwhentheconsideredpointsintimeareinthesametypeoftimeperiod.Therefore,thepeakvalueisPStj=1(tj)2(Pton;j)2,wherePton;jdenotestheprobabilityanodeisconsidered\on".Hence,thefractionoftimenodesspendonmoving( v)andpause( v))tochangethepeakvaluesinthecurveofperiodicalre-appearancepropertytomatchwiththecurvesinFig. 4-10 (b).WeusetheMITWLANtrace[ 82 ]asanexampletodisplaythematchbetweenthesynthetictracederivedfromtheTVCmodelandtherealtrace.WealsoachievedgoodmatchingwiththeUSC[ 80 ]ortheDartmouth[ 81 ]traces,butdonotshowithereduetospacelimitations(see[ 9 ]).Weshowtheskewedlocationvisitingpreferencesandtheperiodicalre-appearancepropertiesinFig. 4-18 (a)and(b),respectively.Wersttryasimplesyntheticmodel(labeledasmodel-simplied,usingtheparametersofModel-1inTable 4-2 )withonecommunityintwotimeperiods.Whilethissimplemodelcapturesthemajortrendsinthemobilitycharacteristics,thereareseveralnoticeabledierences.First,sincethereisonlyonecommunity,thetailinthemodel-simpliedcurveinFig. 4-18 (a)is\at"asopposedtotheexponentiallydiminishingtailoftheMITcurve(noticetheY-axisinFig. 4-18 (a)isinlogscale).Second,thepeaksinthemodel-simpliedcurveinFig. 103

PAGE 104

(b)Figure4-18. MatchingtheMITWLANtracewiththesynthetictrace.(a)Skewedlocationvisitingpreferences.(b)Periodicalre-appearanceatthesamelocation. 4-18 (b)areofequalheights,duetothesimpletwo-alternating-time-periodstructure,asopposedtovaryingpeakvaluesoftheMITcurve.Wecanimprovethematchingbetweenthesynthetictraceandtherealtracebyaddingcomplexityinbothspace(usingmorecommunities)andtime(usingmorecomplexschedule,suchastheweeklyscheduleshowninFig. 4-12 ).InarenedmodellabeledasModel-complexinFig. 4-18 ,weshowthattheresultingmobilitycharacteristicsmatchverycloselywiththeMITtrace.Thisalsodemonstratestheexibilityofourmodel-theusercanadjustitscomplexitybychoosingthenumberofcommunitiesandtimeperiodsneededtoachieveadesiredlevelofmatchingwiththemobilitycharacteristics. 17 ],awebsitethattracksparticipatingtaxisinthegreaterSanFranciscoarea.Weprocessa40-daytraceobtainedbetweenSep.22,2006andNov.1,2006for549taxis.Weobtainthemobilitycharacteristicsofthetaxisbythefollowingsteps.Foreachtaxi,werstidentifyitsmovementrangewithinthe40-dayperiod,thendrawarectangularareathatboundsthemovementofthetaxi,anddividethisareaintoequal-sized10-by-10grids.Wetallythemobilitystatisticsofthetaxisusingthese100gridsaslocations,andshowtheresultsinFig. 4-19 (a)and(b),respectively,withthe 104

PAGE 105

(b)Figure4-19. Matchingthevehiclemobilitytracewiththesynthetictrace.(a)Skewedlocationvisitingpreferences.(b)Periodicalre-appearanceatthesamelocation. labelVehicle-trace.ItisinterestingtoobservethatthetrendofvehicularmovementsisverysimilartothatofWLANusersintermsofthesetwoproperties.Wefurthershowthat,usingtheoutlinedprocedures,wecangenerateasynthetictracewithsimilarmobilitycharacteristicsasthevehiclemobilitytrace.Afterobservingthetraceclosely,wediscoverthatthetaxisareoine(i.e.,notreportingtheirlocations)whennotinoperation.Henceinthesynthetictracewemakethecorrespondingassumption(inSTEP3)thatthenodesareassociatedwiththecurrentgridtheyresideinonlywhentheyaremoving;wethenconsiderthepausetimesasbreaksinthetaxioperation(hencePton;j=( v)=( v)inthiscase),fromwhichwecancalculateoradjusttherespectivemodelparameters.ThecurvesinFig. 4-19 withlabelModelcorrespondtothemobilitycharacteristicsofthesynthetictrace.Asanalnote,althoughvehicularmovementsaregenerallyconstrainedbystreetsandourTVCmodeldoesnotcapturesuchmicroscopicbehaviors,designatedpathsandotherconstraintscouldstillbeaddedinthemodel'smap(forvehicularorhumanmobility)withoutlosingitsbasicproperties.Wedeferthisforfuturework. 84 ],bysettingupitsparametersproperly.Specically,we 105

PAGE 106

27 ].Inthisexperiment,wirelessdevicesweredistributedto41participantsoftheconference,withappropriatesoftwareinstalledthatcouldlogencountersbetweennodes(i.e.comingwithinBluetoothcommunicationrange),astheymovedaroundthepremisesoftheconferencearea.Theinter-meetingtimeandtheencounterdurationdistributionsofall820pairsofusersobtainedfromthistraceareshowninFig. 4-20 withlabelCambridge-INFOCOM-trace.TomimicsuchbehaviorsusingourTVCmodel,weobservetheconferencescheduleatINFOCOM,andsetupadailyrecurrentschedulewithvedierenttypesoftimeperiods(STEP1):technicalsessions,coeebreaks,breakfast/lunchtime,evening,andlatenight(see[ 9 ]forthedetailedparameters).Foreachtimeperiodwesetupcommunitiesastheconferencerooms,thediningroom,etc.Wealsogenerateacommunitythatisfarawayfromtherestofthecommunitiesforeachnodeandmakethenodesometimesisolatedinthiscommunitytomimicthebehaviorofpatronsskippingpartoftheconference.Itisinterestingtonotethattheinter-meetingtimedistributionhasasharpdrop(the\knee"inthecurve)at16hours,whichisapproximatelythetimegapbetweentheendofthedayandthebeginningofthesubsequentdayattheconference.Thissuggeststhenodes(naturally)meetwithlowerprobabilityduringthenights,andthusthetime-dependentmobilityprovidedbyourTVCmodelisappropriate.Wecannaturallyachievethisbyassigningnodestodisjointcommunities(i.e.,the\hotelrooms")duringthenights.InSTEP2,weusethetheorypresentedinsection 4.2.3 toadjusttheparametersandshapetheinter-meetingtimeandencounterdurationcurves.Forexample,astrongertendencyfornodestochooseroamingepochs(settinglargertr)wouldincreasethemeetingprobability(see,e.g.,Eq.( 4{15 )),hencereducinginter-meetingtimes.Sincethedevicesusedtocollecttheencountertracesarealways-on,wedonotapplyanychangestothesynthetictraceinSTEP3.Werandomlygenerate820pairsofusersandobtaintheircorrespondingdistributionsoftheinter-meetingtimeandtheencounter 106

PAGE 107

(b)Figure4-20. Matchinginter-meetingtimeandencounterdurationdistributionswiththehumanencountertrace.(a)Inter-meetingTime.(b)Encounterduration. duration.ThesedistributionsareshowninFig. 4-20 withlabelModel.ItisclearthatourTVCmodelhasthecapabilitytoreproducetheobserveddistributions,evenifitisnotconstructedexplicitlytodoso.Thisdisplaysitssuccessincapturingthedecisivefactorsoftypicalhumanmobility.Itisclearfromthecasesstudiedherethat,onceweobservethetargetenvironmentcloselyandcomeupwiththerightunderlyingparameters,theTVCmodelisabletocapturetheconsequentmobilitycharacteristicswell.Inaddition,withtherespectiveconguration,itispossibletogeneratesynthetictraceswithmuchlargerscale(i.e.,morenodes)thantheempiricaloneswhilemaintainingthesamemobilitycharacteristics.Itisalsopossibletogeneratemultipleinstancesofthesynthetictraceswiththesamemobilitycharacteristicstocomplementtheoriginal,empiricallycollectedtrace.Althoughotherproposedmodelshavealsomanagedtomatchsomesetsofcollectedmeasurements[ 62 { 65 ],noneoftheexistingworkshasbeenshowntocapturethevarietyofqualitativelydierenttraces(e.g.WLAN,vehicles,inter-contacts)thattheTVCmodeldoes. 4.2.3 areinterestingintheirownmerit,theyareparticularlyusefulinpredictingprotocolperformance,whichinturncanguidethedecisionsofsystemoperation.Weillustratethispointwithtwoexamplesinthissection. 107

PAGE 108

87 ].Thus,usingtheresultsofSection 4.2.3.2 wecanestimatethenumberofnodes(asafunctionoftheaveragenodedegree)neededtoachieveatargetperformanceforgeographicrouting,foragivenscenario.WeconsiderthesamesetupasinSection 4.2.4.2 ,wherehalfofthenodesareassignedtoacommunitycenteredat(300;300)andtheotherhalfareassignedtoanothercommunitycenteredat(700;700).Weareinterestedinroutingmessagesacrossoneofthecommunities,fromcoordinate(250;250)tocoordinate(350;350)withsimplegeographicrouting(i.e.,greedyforwardingonly,withoutfacerouting[ 89 ]).Usingsimulationsweobtainthesuccessrateofgeographicroutingundervariouscommunicationrangeswhen200nodesmoveaccordingtothemobilityparametersofModel-1(Table 4-2 ).ResultsareshowninFig. 4-21 (eachpointisthepercentageofsuccessoutof2000trials).Ifweassumenowthatthemobilitymodelwasdierent,sayModel-3,wewouldliketoknowhowmanynodeswewouldneedtoachievesimilarperformance.UsingEq.( 4{6 )wendthat760nodesareneededtocreateasimilaraveragenodedegreeforModel-3.Tovalidatethis,wealsosimulategeographicroutingforascenariowhere760nodesfollowModel-3.Comparingtheresultingmessagedeliveryratioforthisscenariototheoriginalscenario(200nodeswithModel-1)inFig. 4-21 ,weseethatsimilarsuccessratesareachievedinbothscenariosunderthesametransmissionrange,whichconrmstheaccuracyofouranalysis. 71 ].Ithasbeenshownthatmessagepropagationunderepidemicroutingcanbemodeledwithsucientaccuracy(assumingthenumberofnodesislargeenough)usingasimpleuid-basedmodel[ 91 ].(Notethatitsperformancehasalsobeenanalyzed 108

PAGE 109

Geographicroutingsuccessrateunderdierentmobilityparametersetsandnodenumbers. usingMarkovChain[ 33 92 ]andRandomWalk[ 56 ]models.)ThisuidmodelhasbeenborrowedfromtheMathematicalBiologycommunity,andisusuallyreferredtoastheSI(Susceptible-Infected)epidemicmodel.ThegistoftheSImodelisthattheratebywhichthenumberof\infected"nodesincreases(\infected"nodesherearenodeswhohavereceivedacopyofthemessage)canbeapproximatedbytheproductofthreequantities:thenumberofalreadyinfectednodes,thenumberofsusceptible(notyetinfected)nodes,andthepair-wisecontactrate,(theimplicitassumptionthereofcoursebeingthatnodesmeetindependently).ThiscontactrateintheSImodelisequivalenttotheunit-stepmeetingprobabilitiescalculatedinSection 4.2.3.4 .Thus,onecouldinessenceplug-inthesemeetingprobabilitiesintotheSImodelequationsandcalculatethedelayforepidemicrouting.Yet,intheTVCmodel(andofteninreallife)therearemultiplegroupsofnodeswithdierentcommunities,andthusdierentpair-wisecontactratesthatdependonthecommunitysetup.Forexample,nodeswiththesameoroverlappingcommunitiestendtomeetmuchmoreoftenthannodesinfarawaycommunities.Forthisreason,weextendthebasicSImodeltoamoregeneralscenariothatisapplicabletotheTVCmodel.Weconsiderthefollowingsetupinthecasestudy:WeuseModel-3(Table 4-2 )forthemobilityparameters.AtotalofM=50nodesaredividedintotwogroupsof25nodeseach.Onegrouphasitscommunitycenteredat(300;300)andtheotherat(700;700).Onepacketstartsfromarandomlypickedsourcenodeandthetimeneededuntilitreachesall 109

PAGE 110

ovI2(t)S1(t)dI2(t) ovI1(t)S2(t)S1(t)+I1(t)=M=2S2(t)+I2(t)=M=2:(4{21)whereSx(t)andIx(t)denotethenumberofsusceptibleandinfectednodesattimetingroupx,respectively.Parametersovandno ovrepresentthepair-wiseunit-timemeetingprobabilitywhenthecommunitiesareoverlapped(i.e.,fornodesinthesamegroup)andnotoverlapped(i.e.,nodesindierentgroups),respectively.WeuseEq.( 4{15 )toobtainthesequantities.ThismodelisanextensionfromthestandardSImodel[ 91 ]andsimilarextensionscanbemadeformorethantwogroups[ 90 ].Therstequationgovernsthechangeofinfectednodesintherstgroup.Noticethattheinfectiontosusceptiblenodesinthegroup(S1(t))cancomefromtheinfectednodesinthesamegroup(I1(t))ortheothergroup(I2(t)).Wecansolvethesystemofequationsin( 4{21 )togettheevolutionofthetotalinfectednodesinthenetwork.AscanbeseeninFig. 4-22 ,thetheorycurvecloselyfollowsthetrendinsimulationcurve(thenon-perfectmatchingbetweenthetwocurves,isduetothefactthatuidmodelsbecomemoreaccurateapproximationsoftheactualstochasticspreadingforlargenumbersofnodes).Thisindicatesrstthatscenariosgeneratedbyourmobilitymodelarestillamenabletouidmodelbasedmathematicalanalysis(SI),despitetheincreasedcomplexityintroducedbytheconceptofcommunities.Italsoshowsthatresultsproducedthuscanbeusedbyasystemdesignertopredicthowfastmessagespropagateinagivennetworkenvironment.Thismight,forexample,determineifextranodesareneededinawirelesscontentdistributionnetworktospeedupmessagedissemination. 110

PAGE 111

Packetpropagationwithepidemicrouting.Thetotalpopulationisdividedinto2groupswithdierentcommunity. Asanalnote,inadditiontotheepidemicrouting,thetheoreticalresultsforthehittingandmeetingtimescouldbeappliedtopredictthedelayofvariousotherDTNroutingprotocols(seee.g.[ 52 56 91 ]),foralargerangeofmobilityscenariosthatcanbecapturedbytheTVCmodel. 9 ].Inadditiontoprovidingrealisticmobilitypatterns,theTVCmodelcanbemathematicallyanalyzedtoderiveseveralquantitiesofinterest:thenodalspatialdistribution,theaveragenodedegree,thehittingtimeandthemeetingtime.Throughextensivesimulationstudies,wehaveveriedtheaccuracyofourtheory.TheTVCmodelcanbeeasilygeneralizedtoprovidescenariosinwhichnodesdisplaymoreheterogeneousbehavior.Nodesmayhavedierentsetofparametersandeventhetimeperiodstructurecanbedierentfordierentnodes.Withtheseextensions,wehaveamobilitymodeltodescribeanenvironmentincludinguserswithdiversemobilitycharacteristics.Webelievesuchamodelisaveryimportantbuildingblock 111

PAGE 112

56 57 ])underthetime-variantcommunitymobilitymodel.Wealsowouldliketoconstructasystematicwaytoautomaticallygeneratethecongurationles,suchthatthecommunitiesandtimeperiodsofnodesaresettocapturetheinter-nodeencounterpropertiesweobserveinvarioustraces(forexample,theSmallWorldencounterpatternsobservedinWLANtraces[ 60 ],seechapter 6 forthedetails). 112

PAGE 113

113

PAGE 114

4 ,aswepreviouslydonotseektodistinguishusersassimilarordissimilarbasedontheirbehavioralpatterns,whichwewilldointhischapter.Specically,wetakearststeptowardunderstandingandcharacterizingthestructureofbehavioralpatternsofuserswithinlargeWLANs.Wedevelopmethodstoidentifygroupsofusersthatdemonstratesimilarandcoherentbehavioralpattern.Thisisimportantforseveralreasons:(1)Fromtheapplicationorserviceperspective,thegroupsidentifydierentexistingmajorbehavioralmodesinthenetwork,and,hence,canbepotentiallyutilizedtoidentifytargetsforgroup-awareservices.(2)Fromthenetworkmanagementperspective,ithelpsustounderstandthepotentialinterplayoftheusergroupswiththenetworkoperationandrevealsinsightpreviouslyunavailablebylookingatthemereaggregatenetworkstatistics.(3)Fromasocialsciencesperspective,theresultsunraveltherelationshipsbetweenusers(i.e.,their\closeness"intermsofnetworkusagebehavior)whentheyembraceanewlifestyle.Weapplyouranalysisframeworkonlong-termWLANtracesobtainedfromtwouniversitycampuses[ 80 81 ]acrossthecoastsofUSA.Werepresentauser'sbehavioralfeaturesbyconstructinganormalizedassociationmatrixtowhichweapplyouranalysis.WhiletheapplicabilityofourmethodsisnotspecictoWLANs,thesearethemostextensivewirelessuserbehavioraltracesavailabletoday.Weleverageunsupervisedlearning(i.e.,clustering)techniques[ 95 ]todeterminegroupsofusersdisplayingsimilarbehavior.Whileclusteringhasbeenwidely-appliedinotherareasandinsome 114

PAGE 115

24 38 ]toWLANtraces,themaincontributionofthestudyistoconstructproperrepresentationsforourdatasetsanddesignnoveldistancemetricsbetweenusers.Thesetwoaspectsarefundamentalintheapplicationofclusteringtechniquesanddeterminethequalityoftheresultsweobtain.Thekeychallengeindesigningagooddistancemetricistoaccuratelyandsuccinctlysummarizethetrendsinthedata,sothedistancesarenotinuencedbynoiseandcanbeevaluatedeciently.Weshowthatasingular-valuedecomposition(SVD)basedschemenotonlyprovidesthebestsummaryofthedata,butalsoleadstoadistancemetricthatisrobusttonoiseandiscomputationallyecient.Thesuccinctsummariesalsohelptoreducetheprocessing,storage,andexchange(i.e.,whennodescommunicatewitheachothertoconveytheirbehavioralsummaries)overhead.Furthermore,wevalidateourmethodsandexplainitssignicance.WeleverageourTRACEapproach(outlinedinthechapter 1 )tounderstandusergroupinginthisstudy.Specically,theworkstartswiththeWLANTracesthatcapturerealisticuserbehavior.WethenfocusonaspecicRepresentationdistilledfromthetracesthatcapturesimportantaspectsofuserbehavior,asweintroduceinsection 5.2 .WethenAnalyzetheclusteringoftheusersbasedonthechosenrepresentation,normalizedassociationvectors,fromsection 5.3 tosection 5.6 .Werstshowtheneedforagooddistancemetricforclusteringinsection 5.3 .Toachievethatgoal,weconductfurtheranalysistounderstandthenatureofuserassociationpatterns,andevaluateandcontrastvarioussummariestocaptureitsmajortrendinsection 5.4 .Wethenutilizeafeature-basedapproachtoachievemeaningfuluserclusteringinsection 5.5 anddiscussitsinterpretationinsection 5.6 ,whereweshowtheCharacteristicsofusergroupsweobservedfromthetraces.WebrieydiscusshowtoEmploytheofthemethodsandndingswedevelopintheuser-clusteringeortinsection 5.7 ,andtakeoneapplication,thebehavior-awaremessagedelivery,asthemajorfocusinsection 5.8 toshowtheusefulnessoftheunderstandingofusergroupingandoursimilaritymetric. 115

PAGE 116

3-1 .WhilethedevicesloggedintheWLANtracesaremainlylaptopcomputers,wenotethatourmethodsarenotlimitedtothespecicdatasetswechoose,anditwouldbeofgreatinteresttostudytracesfromothermobiledevices(e.g.,cellphones,iPods),ifavailableforalargepopulation.Tounderstanduserbehaviorfromwirelessnetworktraces,therstfundamentaltaskistochoosearepresentationoftherawdata.Thischosenrepresentationshouldhavesignicancetothenetworkandinthegreatersocialrelationshipcontext.WechoosethepatternsofusersvisitingWLANaccesspoints(APs)fortheanalysis.VisitingpatternisimportanttoWLANsasmobilityisoneofitsdeningcharacteristics.WhenaWLANusermoveswithinthecampusandassociateswithAPsacrossthenetwork,thesetofAPswithwhichtheuserassociatesisconsideredanindicatoroftheuser'sphysicallocation.Fromasocialcontext,theplacesapersonvisitsregularlyandrepeatedlyusuallyhaveastrongerconnectiontoheridentityandaliation.Itisperhapsoneoftheimportantdistinguishingfactorsforpeoplewithdierentsocialattributes. 116

PAGE 117

5.10 .Wechoosetouseadayasthetimeslotsinceitrepresentsthemostnaturalbehaviorcycleinourlives.Theassociationvectorforeachtimeslotisann-entryvector,(x1;x2;:::;xn),wherenisthenumberofuniquelocations(i.e.,buildings 5-1 ,i.e.,weconcatenatetheassociationvectorsforeachtimeslot(day).Iftherearendistinctlocationsandthetraceperiodconsiststtimeslots,theassociationmatrixforauserisat-by-nmatrix.

PAGE 118

Illustrationofassociationmatrixrepresentation. 95 ].Weusethehierarchicalclustering,inwhicheachelementisinitiallyconsideredasaclustercontainingonemember.Then,ateachstep,basedonthedistancesbetweentheclusters

PAGE 120

(b)Figure5-2. Cumulativedistributionfunctionofdistancesforinter-clusterandintra-clusteruserpairs(AMVDdistance).(a)USC.(b)Dartmouth. chosenthresholds,andtheparticularchoiceisdata-dependent.Weexperimentwithvariousthresholds,anddiscoverthatfortheUSCtrace,wecangroupthepopulationsinto200clusterswithaclearseparationbetweeninterandintraclusterdistancedistributions(Fig. 5-2 (a)),whichisaqualitativeindicatorforatightclustering.However,thedistancemetricworkspoorlyfortheDartmouthtrace,asshowninFig. 5-2 (b).Theseparationbetweeninterandintraclusterdistancedistributionsisnotclear,regardlessoftheclusterthresholdsweuse(wehavetriedseveral).OneproblemwiththeAMVDmetricisthatitconsidersallassociationvectors,i.e.,itincludesnotonlytheimportanttrends,butalsothenoisevectorswhentheusersdeviatefromthedominanttrend,leadingtobadclusteringresults.Ameaningfuldistancemetricshouldcapturethemajortrendsofuserbehaviorandberobusttonoiseandoutliers.AnotherproblemoftheAMVDmetricisitscomputationalcomplexity.Wehavetocalculatethedistancesbetweenallt2pairsofassociationvectorsforeachuserpair.IfthereareNusersthecomputationrequirementisoforderO(N2t2).Furthermore,itrequiressignicantspacetostoretassociationvectorsforallNusers.Thuswewouldliketodesignametricthatisboth(1)robusttonoiseand(2)computationandstorageecient.Inordertoachievebothgoals,westartbystudyingthecharacteristicsoftheassociationpatternsofasingleusertovalidatetherepetitivepatternsormodesofbehavior.Weshowthatthisstudyleadsustotheappropriatedistancemetric. 120

PAGE 121

(b)Figure5-3. Distributionofnumberofclusters(behavioralmodes)forusers.(a)Clusteringthreshold=0.2.(b)Clusteringthreshold=0.9. 5-3 .InFig. 5-3 (a),weuseasmallclusteringthreshold(0:2),withwhichonlyverysimilarassociationvectorsare 121

PAGE 122

5-3 (b)thatusersstillshowmultiplebehavioralmodes.Onaverage,with0:9astheclusteringthreshold,thenumberofbehavioralmodesforUSCandDartmouthusersare5:57and4:32,respectively,andtheuserswiththemostbehavioralmodeshave32clustersinbothcases.Mostofthoseuserswithtwobehavioralmodeshaveaconsistentassociationpattern:Onemodecorrespondstotheassociationvectorswhentheuserisoine,andtheotheronecorrespondstotheassociationvectorswhentheuserisonline.Theseusersswitchbetweenonlineandoinebehaviorsfromdaytoday,andwhentheyareonline,theassociationvectorsareconsistentandfallinasinglebehavioralmode.Werefertotheseusersassingle-modalusers.Ontheotherhand,wealsoobservemanymulti-modalusers.Theseusersshowamorecomplexbehavior:theirassociationvectorsformmorethantwoclusters,whichindicatethattheydisplaydistinctbehavioralmodeswhentheyareonline.71:9%ofusersinUSCand59:4%ofusersinDartmouthareclassiedasmulti-modalwhentheclusteringthresholdis0:9.Hence,weconcludethatalthoughusersinWLANsarenotextremelymobile,theydomoveanddisplayvariousassociationpatternsoveraperiodoftime.Toexaminethedegreeofdominanceofthemostimportantbehaviormodesofusers,wecomparethemostimportantbehavioralmodeandthesecondmostimportantone(i.e.,thelargestandthesecondlargestclusters)intermsoftheirsizes.InFig. 5-4 weplotthesize(i.e.numberofvectors)distributionsoftherstandthesecondbehavioralmodesunderclusteringthreshold0:2(solidlines)and0:9(dottedlines)forUSCusers.Weseethatthereisaclearseparationbetweenthesizesofthesetwobehavioralmodes.(i.e.,themostdominantbehavioralmodeismuchmoreimportantthanthesecondmost 122

PAGE 123

DistributionofassociationvectorsintherstandthesecondbehavioralmodesfortheUSCtrace.Right:therstcluster,Left:thesecondcluster. importantoneformostusers.)Dierentclusteringthresholdsdonotchangetheresultsmuch.Inotherwords,observationsofthemostdominantbehavioralmodecouldrevealusercharacteristictoagoodextentformanyusers.SimilarobservationsalsoholdfortheDartmouthusers.WeshowthedistributionofthesizeratiobetweenthelargestandthesecondlargestclusterinFig. 5-5 .HereweseeforUSCandDartmouth,respectively,36%and31%ofusershavethetwomostimportantbehaviormodeswithcomparablesizes(i.e.,withsizeratiosmallerthan2:0-Thesecondmostimportantbehavioralmodeisfollowedatleastonehalfasoftenasthemostimportantbehavioralmode).Hencelookingatthemostdominantclusterexclusivelycouldstillbesometimesmisleadingandwemightbeignoringinformationabouttheuser'sdetailedbehavior.Itisthereforedesirabletohaveasummarythattakesnotonlythedominantbehavioralmode,butalsothesubsequentonesintoaccount. 123

PAGE 124

ComplementaryCDFfortheratiooftherstbehavioralmodesizetothesecondbehavioralmodesize.NotethattheX-axisisinlogscaletomakethegraphmorevisible. inthismode).Asusersarenotalwaysonline,theaverageshouldincludeonlytheonlinedaysandignorethezerovectors.Itisdenedas 124

PAGE 125

Theaveragesignicancescoreforvarioussummariesofuserassociationvectors USC0.6460.7160.7020.764Dartmouth0.6900.7570.7470.789 clustersofdierentbehavioralmodesfromtheassociationvectorsoftheuserandidentifythedominantbehavioralmode.Inordertoquantitativelycomparethequalityofthesummarytechniques,weproposetomeasurethesignicancescoreofasummaryvectorwithrespecttoauserbysummingtheprojectionsofallassociationvectorsonthesummaryvector,normalizedbytheonlinedaysoftheuser. 5-1 .Weobservethatthecentroidoftherstclusterbetterexplainsthebehavioralpatternofagivenuserthantheaverage,sinceaveragingsometimesleadtoavectorthatfallsbetweenthebehavioralmodes.Singularvaluedecomposition:WerevisitourdenitionofthesignicancescoreinEq.( 5{5 ),andposeitasanoptimizationquestion:GiventheassociationvectorsXi's,whatisthebestpossiblesummaryvectorYtomaximizeitssignicance?Mathematically,wewantthevectorYtobe 41 ]totheassociationmatrixX.Inotherwords,ifwewantthesummaryvectorYtocapturethemaximumpossiblepowerintheassociationvectorXi's,theoptimalsolutionistoapplysingularvaluedecompositiontoextracttherstsingularvectoroftheassociationmatrixX.Weapplythistechniqueandcalculatethe 125

PAGE 126

5-1 .Itisevidentfromthenumbersthatamongallthecandidates,SVDprovidesthebestsummary.HencewefocusontheuseoftheSVD-basedsummary,anddeferthediscussionofothersummarytechniquestosection 5.10 41 ],weknowthatforanyt-by-nmatrixX,itispossibletoperformsingularvaluedecomposition,suchthat 5{7 )inadierentform: ~Xk=kXi=1uiivTi:(5{8)Hereui'sandvi'sarethecolumnvectorsofmatrixUandV.TheyareusedasthebuildingblockstoreconstructtheoriginalmatrixX.Withthisformat,SVDcanbeviewedasawaytodecomposeamatrix:ItbreaksthematrixXintocolumnvectorsui,viandrealnumbersi.Ifweretainallthesecomponents(i.e.,k=rank(X)),SVDisalosslessoperationandthematrixXcanbereconstructedaccurately.However,inpracticalapplications,SVDcanbetreatedasalossycompressionandonlytheimportantcomponentsareretainedtogivearank-kapproximationofthematrixX.Thepercentage 126

PAGE 127

5{8 )canbecalculatedby 5-4 ).HenceweexpectSVDtoachievegreatdatareductionontheassociationmatrices.Thisisindeedthecase,asweshowinFig. 5-6 :MostoftheusershaveahighpercentageofpowerintheirassociationmatricesXexplainedbyarelativelylow-rankreconstruction{Forexample,intheUSCtrace(Fig. 5-6 (a)),ifweusearank-1reconstructionmatrix,itcaptures50%ormoreofpowerintheassociationmatricesformorethan98%ofusers,andarank-3reconstructionissucienttocapturemorethan50%ofpowerintheassociationmatricesforallusers.Evenifweconsideranextremerequirement,capturing90%ofthepowerintheassociationmatrices,itisachievablefor68%ofusersusingarank-1reconstructionmatrix,andformorethan99%ofusersusingatmostarank-7reconstructionmatrix.SimilarobservationscanbemadefortheDartmouthusers(inFig. 5-6 (b)).Forbothcampuses,vecomponentsaresucienttocapture90%ormorepowerformost(i.e.,morethan90%)oftheusers.Thisindicatesalthoughusersshowmulti-modalassociationpattern,formostusersthetopbehavioralmodesarerelativelymuchmoreimportantthentheremainingones.Ifalow-rankreconstructionoftheassociationmatrixisachievable,itisnaturaltoaskfortherepresentativevectorsforthebehavioralmodesofauser.Forthispurpose,SVDcanbeviewedasaproceduretoobtainrepresentativevectorsthatcapturethemost 127

PAGE 128

(b)Figure5-6. Lowassociationmatricesdimensionality:Ahightargetpercentageofpoweriscapturedwithalowrankreconstructionmatrixformanyusers.(a)USC.(b)Dartmouth. remainingpowerinthematrix.Mathematically 5{9 ).Werefertothesevectorsastheeigen-behaviorvectorsfortheuser.Theeigen-behaviorvectors,uj's,areunit-lengthvectors.Theabsolutevaluesofentriesinaneigen-behaviorvectorquantifytherelativeimportanceofthelocationsintheuser'sj-thbehavioralmode.Forexample,supposeagivenuservisitslocationlalmostexclusively,theninhisrsteigen-behaviorvector,theentrycorrespondstolocationlwouldcarryahighvalue(i.e.closeto1),andtheweightofthersteigen-behaviorvector,21=Pri=12i,shallbehigh.Withasetofeigen-behaviorvectorsandtheircorrespondingweights,wecancaptureandquantifytherelativeimportanceofauser'sbehavioralmodes.

PAGE 130

5-6 ),insteadofgoingthrought-by-tpairsoforiginalassociationvectorsaswedidintheAMVDdistanceinsection 5.3 ,wereducethedistancecalculationto5-by-5vectorpairs.Sincewehaveatleast61daysinthetraces,thisisatleasta(61=5)2148foldsavingforallN2pairofusers.Bypayingthepre-processing(i.e.,SVDforallNusers)overheadofO(Nt2),wecanreducethedistancecalculationcomplexityfromO(N2t2)toO(cN2).Sinceusersfollowrepetitivetrendsintheassociationpatterns,itstotaleigen-behaviorvectorswouldnotgrowwiththenumberoftimeslots,t.Ifweconsiderlongertracesorassociationvectorrepresentationsinnertimescale,thereductioncanbeevenmoresignicant.Inthefollowingcomputations,weconsideronlytheeigen-behaviorvectorsthatcaptureatleast0:1%oftotalpowerintheuser'sassociationmatrix. 5-7 )ascomparedtotheresultswiththeAMVDdistance(showninFig 5-2 ),indicatingameaningfulclustering.Thisprovestheeigen-behaviordistanceisabettermetricthantheAMVDdistanceasithelpsustogroupusersintowell-separatedbehavioralgroupsbasedontheirWLANassociationpreferences,forbothcampuses.

PAGE 131

(b)Figure5-7. Cumulativedistributionfunctionofdistancesforinter-clusterandintra-clusteruserpairs(eigen-vectordistance).(a)USC.(b)Dartmouth. Wefurthervalidatewhethertheresultingclustersindeedcaptureuserswithsimilarbehavioraltrends.Forthistest,weconstructthejointassociationmatrixbyconcatenatingthedailyassociationvectorsofaclusterofmsimilarusersinalargermt-by-nmatrix,wherenisthenumberoflocationsandtisthenumberoftimeslots.WhenweperformSVDtothejointassociationmatrix,thetopeigen-behaviorvectorsrepresentthedominantbehavioralpatternswithinthegroup.Iftheusersinthegroupfollowacoherentbehavioraltrend,thepercentageofpowercapturedbythetopeigen-behaviorvectorsshouldbehigh.Ontheotherhand,ifassociationvectorsofuserswithdierentassociationtrendsareputinonejointassociationmatrix(i.e.,ifdissimilarusersareputinoneclusterbymistake),thepercentageofpowercapturedbyitstopeigen-behaviorvectorsshouldbemuchlower.Amongallclusters,wepickthosewithmorethanveusers,andcomparethecumulativepowercapturedbythetopfoureigen-behaviorvectorsoftheseclusterswithrandomclustersofthesamesize(i.e.,werandomlypickthesamenumberofusersfromthepopulationandconstructanotherjointassociationmatrix)inscattergraphs,showninFig. 5-8 .Clearly,mostthedotsarewellabovethe45-degreelineforbothcampuses.Thisindicatestheusersinthesameclusterfollowamuchstrongercoherentbehavioraltrendthanrandomlypickedusers,pointingtothesignicanceofourclusteringresults.Wewouldalsoliketoseeifeachclusterfromthepopulationshowsadistinctbehavioralpattern.Toquantifythis,weobtainthersteigen-behaviorvectorfrom 131

PAGE 132

(b)Figure5-8. Scattergraph:Cumulativepowercapturedintopfoureigen-behaviorvectorsofrandommatrices(X)andthejointassociationmatricesformedbyusersinthesamecluster(Y).Onlyclusterswith5ormoremembersareincluded.(a)USC(129clusters).(b)Dartmouth(136clusters). eachgroupandcalculateitssignicancescore,denedinEq.( 5{5 ),forallthegroups.Theresultsconrmwithourgoalofidentifyinggroupsfollowingdierentbehavioraltrend:FortheUSCtrace,thersteigen-behaviorvectorsobtainedfromthejointasso-ciationmatriceshaveanaveragesignicancescoreof0:779fortheirownclustersandanaveragescoreof0:005forotherclusters,indicatingthedominantbehavioraltrendsfromeachclusterisdistinct.ThecorrespondingnumbersfortheDartmouthtraceare0:727and0:004,respectively.Weconcludethatwehavedesignedadistancemetricthateectivelypartitionsusersintogroupsbasedonbehavioralpatterns.Inaddition,theseclustersareuniquewithrespecttotheirmajorbehavioraltrends. 5-9 .Weobservethedistributionsofgroupsizesarehighly-skewedforbothcampuses.Therearedominantbehavioralgroupsthatmanyusersfollow:thelargestgroupsinthecampusesinclude504and546members,outofthepopulationof5000forUSCand6582forDartmouth,respectively.Thetenlargestgroupscombinedaccountfor39%and33%ofthetotal 132

PAGE 133

Rankplot(groupsizerankingv.s.groupsize)inlog-logscale.Usergroupsizefollowsapower-lawdistribution. population,respectively.Ontheotherhand,therearealsomanysmallgroups,orevensingletons,forbothpopulations:outofthe200clusters,thereare68and57ofthemwithlessthanvemembers,respectively,andinbothcampusesabouthalfofthegroupshavelessthan10members.Moreinterestingly,weobservethatbesidesthesesmallclusters,thedistributionoftheclustersizeseemstofollowapower-lawdistribution.InFig. 5-9 ,weplotthestraightlinesthatillustratethebestpower-lawts.Theslopesfortheselinesare0:67forDartmouthand0:75forUSC,respectively.Thepowerlawdistributionofgroupsizesmayberelatedtotheskewedpopularityoflocationsoncampuses-ithasbeenshownthatthenumberofpatronstovariouslocationsdiersignicantly[ 13 ].However,thelinkbetweenthedistributionsofnumberofpatronsandthedistributionofgroupsizesisnotdirect.Whilethemost-visitedlocationsonbothcampuseseasilyattractthousandsofpatrons,thesepeoplearebrokenintodierentbehavioralgroupsdependingontheirassociationpreferences.Wenowstudythedetailedbehaviorsofeachclusterbyusingtheeigen-behaviorvectorsandtheirrelativeweightstounderstandthedetailedpreferencesofthegroups.Wediscoverformostofthegroups,theirtopeigen-behaviorvectorsdominate,i.e.,thecontributionofthesecond-mostimportantlocationisalmostinvisibleinthersteigen-behaviorvector.Similarrelationshipholdsbetweenthesecond-mostimportantlocationandthethird-mostimportantone,andsoon.Hencetheassociationbehaviorof 133

PAGE 134

4 ).Forsuchusers,itsmostvisitedlocationsmightbesucienttoclassifythem.Mostlargeuserclustersbelongtothefore-mentionedcase.Thelargestclustersonbothcampusesincludethelibraryvisitors,asexpected,sincelibrariesarestillthemostvisitedareaonuniversitycampuses.FortheUSCcampus,thelargestuserclustervisitsthelibrary(thersteigen-behaviorvectorhasasinglehigh-valueentrycorrespondingtothelibrary,andthiseigen-behaviorvectorcaptures83%ofthepowerinthejointassociationmatrixforthegroup),followedbyacouplelocationsaroundtheLawschool(4:45%)andtheschoolofCommunication(4:5%),botharepopularlocationsoncampus.FortheDartmouthcampus,thelargestuserclustervisitsLibBldg2(72:85%),followedbyLibBldg1(5:13%),SocBldg1(3:56%),andLibBldg3(1:93%).ItseemsthisgroupconsistsoflibrarypatronswhomainlymoveaboutthepublicareaonthecampusandaccesstheWLANfromtheselocations.WhilelibrariesarepopularWLANhotspots,wealsodiscovermanyuserclustersthatrarelyvisittheselocations.ThesecondlargestclusterforUSCconsistsofusersvisitingmostlytheLawschool(89:73%ofpower),schoolofaccounting(6:37%),andacoupleoflocationsclosetotheLawschool(0:59%).ForDartmouth,thesecondlargestclustervisitsAcadBldg18(56:38%),AcadBldg6(13:4%),ResBldg83(10:15%),AcadBldg31(3:5%),AcadBldg7(3:12%),whichseemstobeagroupofstudentsgoingtoclassesatmultipleacademicbuildings.Wehavealsoobservedvariousclustersfeatureddierentdormsandclassroomsastheirmostvisitedlocationfrombothcampuses.Ontheotherhand,wehavealsodiscoveredgroupswithmultiplehigh-valueentriesinitstopeigen-behaviorvectorsfrombothcampuses.OneprominentexamplefromtheUSCtraceconsistsof32users,whovisitbuildingsVKCandTHH,twomajorclassrooms 134

PAGE 135

5{4 ))isusedtoclassifyusers,thebehavioraltrendofvisitingmultiplelocationswithsimilartendencywillnotberevealed.Instead,amongthe32users,13areclassiedwithotherswhovisitVKCfrequently,10areclassiedwiththosewhovisitTHHfrequently,andtherestareputintovariousgroups.Asportablewirelessdevicesgainpopularity,weexpecttoseemoreusersdisplayingdiversebehavioraltrendsintermsofnetworkusage.Tofullycapturesuchbehavior,average-basedsummaryisnotsucient,andthisiswhereSVDshowsitsstrengththemost.Interestingly,wealsodiscovermanysmallclusterswithuniquebehavioralpatternsthatdeviatefromthe\mainstream"usersinbothtraces.Forexample,intheUSCtrace,thereisasmallclusterofeightuserswhovisitexclusivelyafraternityhouse.Probablythesearethepeoplewholivethere.IntheDartmouthtrace,thereisaclusterofeightuserswhovisitmostlyathleticbuildings(AthBldg5(90:9%),AthBldg10(4:62%),AthBldg2(3:14%),AthBldg3(0:8%),andResBldg26(0:54%)).Theseareprobablyeitherathletesormanagementstasoftheathleticfacility.Suchndingssubstantiateourmotivationofthestudy:asthewirelesstechnologyprevails,wecanexpectuserstodisplaydiversebehavioralpatternsthatreecttotheirpersonalpreferences,anditisimportanttocapturesuchbehavioraltrendandquantifyitssignicance.

PAGE 136

5.8 .Behavior-awareservices:Inthefuture,weexpectthewirelessdevicestobeveryportableandpersonalized.Hence,theservicesprovidedcouldbehighlypersonalized,oratleastcustomizedbasedontheinterestgroups.Ourmethodwouldfacilitatetoidentifythedominantgroups.Certainly,dierentrepresentationsofusers(e.g.,hobbies,interests)thattintothecontextmightalsobeutilizedrather,butourmethodwouldstillbeapplicable.Basedonthetargetedgroupofagivenservice,theserviceproviderscouldassignatargetbehavioralvectortodescribethepropertyoftargetusers,andtheuserdevicescouldeasilydeterminepotentialcustomersusingasignicancescore(i.e.,Eq.( 5{5 ))tocompareitseigen-behaviorvectortothetargetbehavioralvector.Werefertothisscenarioasinterest-basedgroupingandprole-casting.Wewilldesignaprotocolforthisserviceinsection 5.8 .Inadditiontofacilitatingclusteringoftheusers,theeigen-behaviorvectorscouldalsoprovideanecientmechanismforuserstoexchangetheirbehavioralfeaturesinordertomakenewfriends.Suchsocialprolescouldbeappliedinapplicationsinsocialnetworking,suchasbehaviorpatternorientedmatching.Usermodeling:ResultsfromtheclustersofuserscouldhelpustoproposemorerealisticmodelsforWLANusers,whichisachallengeandanecessityforevaluating 136

PAGE 137

58 ],therehasbeenlittleworkinrealisticmodelsbasedongroups.Ourdecompositionapproachprovidestwopiecesofimportantinformation:(1)thedistributionofgroupsizesfollowsapower-lawdistributionand(2)thedetailedeigen-behaviorvectorsofthegroups.Withsuchinformation,onecansetupagenerativemodelwiththepropergroupsizesandtheweightsforfrequentlyvisitedlocations(e.g.,itscommunitiesintheTVCmodelpresentedinchapter 4 )toevaluatetheirimpactonthenetwork.Networkmanagement:Ouranalysisprovidesadierentviewofnetworkmanagement.WLANmanagementandplanningcouldbedonebymonitoringtheactivitiesofindividualAPsinordertoidentifythebusyones.Fromourclusteringtechnique,themanagercanidentifyusergroupsandtherelativeimportanceoflocationstoeachgroup.Suchinformationcanbehelpfulintermsofloadpredictionandplanning.Forexample,ifthebusinessschoolisgoingtoexpand,bycheckingthebehavioralgroupsofbusinessschoolstudents,itmaybepossibletopredictitsimpactontheloadofdierentpartsoftheWLAN.Forbetterunderstanding,onemayalsoobservethechangeinthegroupstructurewithtimeandacrosssemesters.Aslarge-scalecity-wideWLANdeploymentsbecomecommonplace,solutionstoissuesinmanagement,servicedesign,andprotocolvalidationcouldimmenselybenetfrominsightintothebehavioralpatternsoftheusersorthesociety.Webelievethatourframeworkwillbeabletoprovidethebehavioralpatternsandhelpndsolutionstoseveralproblemsrangingfromwirelessnetworkmanagementtounderstandingbasicsocialbehaviorofusers. 137

PAGE 138

7 ,afterweunderstandtheglobalstructureofthenetworkinchapter 6 .Inthiscasestudy,weproposeasimilarity-basedprole-castprotocolthatmakesthemessageforwardingdecisionbased 138

PAGE 139

71 ](i.e.,wecaneliminatemorethanhalfofthetransmissionswithalittlereductioninthedeliverysuccessrate)andout-performrandom-walkbasedprotocolsintermsofthedeliverydelay(foratleast30%). 5-10 (a).Suchdecisionshaveimplicationsonmanyaspectsofhowecientlytheroutingstrategieswork,suchasdelay,overhead,andmessagedeliverysuccessrate.Thereexistsatradeobetweentheseperformancemetrics,andawell-designedprotocolshouldprovideamechanismforitsuserstostrikearightbalanceforthegivenenvironment.Thekeyresearchchallengeindesigningtheroutingprotocolsistomakeanintelligentdecisionwiththelocalinformationavailabletothetwoencounteringnodes,assumingnoknowledgeabouttheglobalnetworkproperties,whichisusuallyunavailableindecentralizednetworkssuchasDTNs.Forourprole-castservice,thegoalistoreachasetofnodeswithacertainsimilarproperty.TheconceptualviewoftheproblemisillustratedinFig. 5-10 (b).Weconsideravirtual,high-dimensionalprolespacewhereeachnodeisrepresentedbyapointinthespace.Thenodesthataresimilarwithrespecttothepropertyweusetoconstructtheprolespaceshouldbeclosetoeachotherinthisspace,anddissimilarnodesshouldsitfarapart.Ourspecicapplicationweconsiderherecorrespondstoascoped-oodingintheprolespace:Thegoalistoreachallsimilarnodes(withrespecttotheprolewechoosetoconstructtheprolespace)tothesender.Thenodesshouldkeepforwardingthemessagetothosewhoaresimilartothemundertheconsideredprole,butignorethosewhoaredissimilar.LinkingtheguresinFig. 5-10 ,theypointoutaneedfornodestoevaluatetheirmutualsimilarityintheconsideredprolespacewhentheyencounter,and 139

PAGE 140

(b)Figure5-10. Twodierentviewsoftheprole-castserviceintheDTN.(a)Physicalview:ForwardingdecisionsintheDTN.(b)Conceptualview:Scoped-oodingintheprolespace.Theconceptualviewofscopedoodingintheprolespacehastobeimplementedthroughmessageforwardingdecisionsatnodalencounterevents. usethispieceofinformationtoguidetheroutingdecisionsintheDTN.Weproposeasimilarity-basedprotocolforthispurposeinsub-section 5.8.2 .Weusemobilityproleasanexampletoillustratetheusefulnessoftheprole-castserviceparadigm.Wechoosethemobilityproleforthestudyforthefollowingreasons.First,ithasbeenshownintheprevioussectionsthatmobilityisoneofthedistinguishingfeaturetodierentiateusersfromalargepopulation.Groupswithdistinctbehavioralpatternscanbeidentiedwithrespecttothelong-runmobilitypatterns,andweusethesegroupsasourtargetsintheprole-castprotocol.Second,mobility-prole-casttieswithsomenewservicesintheadhocnetwork.Forexample,astudentlosesawalletandwishestosendamessagetootherfellowstudentswhovisitsimilarplacesoftenashedoestolookforit.Or,themanagerofthelibrarymaywanttosendanannouncementaboutpowershutdownonlytoitsfrequentpatrons.Theseservicesaremobilitypatternspecic,andnoneoftheexistingserviceparadigmsservestheneedofidentifyingtheintendedmessagerecipientsfromadiversepopulationwell.Third,toevaluatetheeectivenessofourproposedprotocolrealistically,weneeddetailedtracesofuserbehaviorwithrespectto 140

PAGE 141

5-1 ,todescribethelong-runmobilitytrendofamobileuser.Foreachtimeslot,eachnodegeneratesanassociationvectorthatsummarizesitsassociationwithvisitedlocationsduringthistimeslot,asdescribedinsection 5.2.1 .Theassociationmatrixrepresentationcapturestherelativeimportanceoflocationsonthecampustoeachuser(i.e.,thepreferenceintheusermobilityprocess).Basedonthisrepresentation,weclassifythewholeuserpopulationintodistinctbehavioralgroupswithclusteringmethodsdetailedintheprevioussections.Thesegroupscorrespondtouserswithuniquemobilityproles.Intheprotocolevaluationpresentedlater,wetakethesebehavioralgroupsasthetargetsformobilityprole-cast.Evaluationofusersimilaritybasedonthemobilityproles:Whennodesencounterwitheachother,theyneedtoexchangethemobilityprolefortheevaluationof 5.8.2 couldbeusedforothertypesofuserproleaswell.

PAGE 142

5.4.2 .Theeigen-behaviorvectors(denedinEq.( 5{10 ))anditscorrespondingweightsprovideaconciseyethighlyaccuraterepresentationofusermobilityproleforexchangewhentheusersencounterwitheachother.Whentwousersmeetwitheachother,theyexchangethesummarizedmobilityproles(i.e.,eigen-behaviorvectorswiththeirweights)oftheirpreviouslycollectedmobilitypatternanddecidewhethertheyaresimilaratthespot.Thesimilarityindexiscalculatedastheweightedsumofinnerproductsoftheeigen-behaviorvectors,asdenedinEq.( 5{11 ).Ifthesimilarityindexislargerthanathreshold,theyexchangethemessage.Notethisdecisionissolelylocal,involvingonlythetwoencounterednodes.Thephilosophybehindtheprotocolis,ifeachnodedeliversthemessageonlytootherswithhighsimilarityinmobilityprole,thepropagationofthemessagecopieswillbescopedwithinagroupofsimilarusers.Thethresholdthattriggersthemessagetransmissionprovidesacontrolfortheprotocolusertoadjustthetradeobetweentheperformancemetrics.Ahigh-valuedthresholdfavorslowtransmissionoverhead,whilealow-valuedthresholdleadstoshortdeliverydelayandhighdeliverysuccessrate. 5.8.3.1EvaluationsetupInthissectionwedescribetheexperimentsetuptoevaluateoursimilarity-basedprole-castprotocolpresentedintheprevioussection.WeutilizetheUSCtrace(i.e.,USC-06spring)tostudythemessagetransmissionschemesempirically.SomelogisticdetailsofthedatasetcanbefoundinTable 3-1 .Weassumethattwonodesareabletocommunicate(i.e.,encounterwitheachother)whentheyareassociatedtothesamelocationintheWLAN.NotethattheWLANinfrastructureismerelyusedtocollectuser 142

PAGE 143

71 ]inDTN,usingtheanalogythatthemessagepropagatesinthenetworklikeanepidemic.ThisisalsothemostaggressiveforwardingstrategyinDTN.Underanidealisticenvironment(i.e.,nopacketdropduetowirelesscontentionorinsucientbuersize),thisisalsothestrategythatachievestheshortestpossibledelayandbestdeliverysuccessrate.Centralized:Inthisidealscenario,weassumethatallnodesacquireperfectknowledgeofthegroupmembershipthroughanoraclewithnoadditionalcost.Inordertoreducethetransmissionoverhead,nodesonlypropagatethemessagetoothersiftheyareinthesamegroup.Thisensuresthemessagewillneverpropagatetoanunintendedreceiver,andonlymembersofeachgroupparticipateinmessagedisseminationfortheirowngroup.Random-transmission(RTx):Intherandomtransmissionprotocol,thecurrentmessageholdersendsthemessagetoanothernoderandomlywithprobabilitypwhentheyencounter

PAGE 144

Thechosenprotocolsforevaluationspanthespectrumofusergroupingknowledgeusedintheforwardingdecisionprocess. walkamongthenodes.Loopsareavoidedbynotsendingtothenodeswhohaveseenthesamemessagebefore.Thisprocesscontinuesuntilapre-sethoplimit(i.e.,TTLlimit)isreached.Wealsovarythenumberofcopiesofactivemessage(i.e.,numberofthreadsintherandomwalk)inthenetwork.Whenmrandomwalkthreadsareissued,themessageoriginatorisresponsibleforspreadingthecopiestomdierentnodesitencounterswith,andeachthreadcarriesonindependentlyasdescribedabove.Wehavechosentheaboveprotocolstospanthespectrumofthedegreeofknowledgeabouttheusergroupingintheevaluation,asillustratedinFig. 5-11 .Ononeextremeofthespectrumwehavethecentralizedprotocolwhichhasperfectknowledgeaboutusergrouping.Thisinformationprovidesanopportunityofhighlyecientoperation,butitisnotrealistictoassumeitsavailability,hencethecentralizedprotocolservesonlyasthetheoreticalupperboundoftheperformance.OntheotherextremearetheoodingandRTxprotocols,bothassumingnoknowledgeaboutusergroupingatall.Theyareextremelysimplebutnotoptimizedforthespecictaskofprole-cast.Oursimilarity-basedprotocolusesthesimilarityindexdenedinEq.( 5{11 )toestimatetheboundarywherethescopedoodingshouldbestopped.Itoperatesinthemiddleofthespectrumwithinferredgroupinginformation. 144

PAGE 145

5-12 .Foralltheperformancemetrics,wechoosetheresultsforooding(i.e.,epidemicrouting)asthebaselineandshowthenormalizedperformancemetricsoftheotherprotocolsrelativetothatoftheepidemicroutinginthegures.Intheguresweseethatoodinghasthelowestdelayandthehighestdeliveryratioasitutilizesalltheavailableencounterstopropagatethemessage.However,italsoincurssignicantoverhead.Theaveragedelay,whichisthelowestpossibleunderthegivenencounterpatterns,isintheorderofdays(3:56daysinthisparticularcase).Prole-castbasedoncentralizedgroupmembershipinformation,theidealscenario,showsagreatpromiseforthebehavior-awareprotocols,asitsignicantlyreducestheoverhead(toonly3%oftheooding)whilemaintainsalmostperfectdeliveryratio,withalittleextradelay.Thereissuchextradelayinthecentralizedprotocolbecausethemessagesarecarriedbynodesinthetargetedgrouponly.Itispossibletoevenreducethisdelaybyobtainingpredictionsoffutureencountereventsthroughanoracle,asin[ 54 ].Wechoosenottoaddressthisissueandinsteadshowwhatcanbeachievedbasedontheperfectknowledgeofusergroupingalone,focusingouranalysisonthespectrumofgroupinginformationavailability.Thecentralizedprotocoldisplaystheceilingperformanceonecanachieveintermsofoverheadreductionbyincorporatingknowledgeofusergroupingintheprole-castservice.However,notethatitisnotrealistictoassumesuchcentralizedknowledge.Foroursimilarity-basedprotocol,itsaggressivenesscanbetunedwiththeforwardingthresholdofthesimilarityindex.Weshowthesimulationresultswithvarioussimilaritythresholdsinthegures.LabelSimilarityxindicatesweusexasthethresholdfor 145

PAGE 147

5-13 ,andmarkthe\operationalregion"ofthecomparedprotocols.Ideally,onewouldwanttheprotocoltoworkatthebottom-rightcorner,withhighdeliveryrateandlowoverhead,asclosetothecentralizedprotocolaspossible.Theoodingprotocolalsoachievesgooddeliveryrate,buttheoverheadistoomuch.Oursimilarity-basedprotocolisshownbythewhiteellipse.Itsoperationalregionstretchesfrommoderatedeliveryratiowithlowoverheadtohighdeliveryratiowithmoderateoverhead.TheRTxprotocolwithinniteTTLisrepresentedbythedarkgreyellipse,takingthespaceofmoderatedeliveryratiowithmoderateoverheadtohighdeliveryratiowithhighoverhead.Asmincreases,itdegeneratestotheooding.However,withaproperlychosenstoppingthreshold,theRTx

PAGE 148

Relativeperformancemetricsofthegroup-castschemesnormalizedtotheperformanceofooding. Figure5-13. Theoperationregionsofthecomparedprotocolsinthedeliveryrate-overheadspace. protocolhasthepotentialtooperateinthehighdeliveryratio,lowoverheadarea,asindicatedbythelightgreyellipse.However,itsaveragedeliverydelayisstillmuchhigherthanthatoftheoodingorsimilarity-basedprotocols(inthebestcase,atleast30%morethanthesimilarity-basedprotocol),asRTxdoesnottakefulladvantageoftheavailableintermediatenodesintheDTNframework. 148

PAGE 149

8 ],andeventuallyleveragethisndingtodesignecientmessagedeliveryprotocolsformoregenericcasesinchapter 7 149

PAGE 150

5-7 withtheeigen-behaviordistance.However,wehavetonotethattheothersummariespresentedinsection 5.4.2 couldalsobeusedtoobtaindistancemetrics.Forthesesingle-vectorsummaries,suchastheaverageofassociationvectors(Xonavg,Eq.( 5{3 ))orcentroidoftherstcluster(Xcentroid1,Eq.( 5{4 )),wedenedistancemetricsbetweenusersbysimply 150

PAGE 151

(b)Figure5-14. Cumulativedistributionfunctionofdistancesforinter-clusterandintra-clusteruserpairs(otherdistancemetrics).(a)USC,distancebetweenaverageofassociationvectors.(b)Dartmouth,distancebetweencentroidoftherstbehavioralmode. calculatingtheManhattandistance(Eq.( 5{2 ))betweenthecorrespondingsummaryvectors.Withthesedistancemetrics,wecouldalsoarriveatmeaningfulpartitionsofuserpopulationsandhencethosearevalidmetrics,too.WeshowtwosuchexamplesinFig. 5-14 -thegeneralobservationisthatwhileXonavgleadstolesswell-separatedclustersthantheeigen-behaviordistance,Xcentroid1leadstoevenbetterresults.Weneedtofurthercomparethesedierentpartitionsoftheuserpopulationtounderstandtheirproperties.WechoosetousetheJaccardindex[ 96 ]tocomparethesimilaritybetweendierentpartitionsofthesamepopulation.TheJaccardindexbetweentwopartitionsonthesamepopulationisdenedas 96 ].Forbothtraces,welisttheJaccardindicesbetweenuserpartitionsfromvariousdistancemetrics(AverageandtheCentroidofrstbehavioralmode)andthe 151

PAGE 152

Jaccardindicesbetweenuserpartitionsbasedontheeigen-behaviordistancesandvariousdistancemetrics. DistanceAverageCentroidw/Centroidw/metricthreshold=0.5threshold=0.9 USC0.7570.7410.696Dartmouth0.8010.7060.710 partitionfromtheeigen-behaviordistanceinTable 5-2 .TheJaccardindicesaremostlyintherangeof0:7to0:8,indicatingthepartitionsareinfactsimilar.ThebetterseparationbetweenintraandinterclusterdistancedistributionswiththeXcentroid1metricispartlybecausethedistancesarecalculatedbasedonasubsetofassociationvectorswithacoherenttrend,discardingothervectors.Nonetheless,dierentdistancemetrichasitsownemphasis.Whileweargueinsection 5.6 withanexamplethattheeigen-behaviordistanceisusefulforclassifyinguserswithmultiplefrequentlyvisitedlocationswithsimilarpreferences,thisisnotalwaystheonlygoal.Dependingontheapplication,sometimesonemaywanttoconsideronlytherstbehavioralmodeandignoretheothers.InsteadofapplyingSVD,onemayproposetousethecentroidsforallbehavioralmodes(i.e.,alltheclustersformedbytheuser'sassociationvectors)ofauserasasummary.However,thebehavioralmodeforeachuserisdependentontheclusteringthreshold,anditisnotsimpletochooseonethatworkswellformanyusers,consideringthediversity.Ontheotherhand,SVDdoesnotrequireparametertuning,andisoptimalinthesenseofcapturingremainingpowerintheassociationmatrix,sowechooseitoverthemultiplecentroidsmethod,ifallbehavioralmodesofausershouldbeconsidered. 5.2.1 weproposetousethenormalizedassociationvectorinordertomitigatethedierencesofuseractivenessacrossusersandacrosstimeslotsforagivenuser.Thisiseectiveifthepreferenceoftheuserforeachtimeslotisthefocusofstudy. 152

PAGE 153

153

PAGE 154

154

PAGE 155

4 and 5 ,respectively,wetakeanevenmoremacroscopicviewinthischapter.Weconsideranimportanteventbetweenmobilenodesinwirelessnetworks{encounters.Thescopeoftheanalysisisonestepwiderthanwhatwepresentedinthelastchapter{althoughweconsiderencountereventsastheenablingeventsofnode-to-nodecommunicationintheprole-castprotocol,weutilizetheseeventsinalocalizedfashion.Inthischapter,weseektounderstandencountersinthemobilenetworkfromadierentperspective{wetakeaholisticviewonallencountereventshappeningbetweenallthenodesinthenetworkandstudytheglobalencounterpatternsinthetrace,byobservingtheencounterpatternswithagraphanalysisapproach.Suchananalysisshedslightonthefeasibilityofformingainfrastructure-lessnetworkcapableofreachingmostofthenodesthroughtime-varying,partialconnectivitytosomenodesatagiventimeinstantthroughencounters. 155

PAGE 156

3-1 )inadditiontoarealencountertracecollectedatarecentINFOCOMconference[ 27 ](i.e.,theCambridge-INFOCOM05traceinTable 3-1 ).Wecompareandcontrastourobservationsforthevarioustracestodistillandexplainthecommonalitiesanddierencesobserved.Specically,weaimtoquantifythedistributionofencountereventsaMNhas,andlookintotheencounterpatternsofallMNstounderstandtherelationshipbetweenMNsformedbyencounters.Thisisaresearchtopicthatreceivedlessattentioninthepast,butcanbeusefulandsometimesessentialforclassesoffuturemobilenetworkingprotocols.Forexample,encounterhistoriesareusedtodiscoverroutesinadhocnetworkroutingprotocols(e.g.MAID[ 46 ],EASE[ 93 ]),andencountersareuseddirectlyindelaytolerantnetworks(DTNs)topropagatepackets.WedeneanencounterbetweentwousersastheeventoftheirassociationwiththesameAPforoverlappedtimeintervals.FromalltheWLAN-basedtraceswestudied,wendthatthedistributionofencountersishighlyasymmetric,indicatingaheterogeneoususerpopulation.Surprisingly,wendthatauser,onaverage,onlyencountersbetween0:79%and6:7%ofthenetworkuserpopulationwithinamonth.WealsoestablishthatthetotalnumberofencountersforeachMNfollowsBiParetodistribution,theparametersofwhichareenvironmentspecic.WefurtherutilizeagraphanalysisapproachtounderstandtherelationshipbetweenMNsformedbyencounterevents.WeutilizetheSmallWorldmodel[ 8 ]tounderstandthecharacteristicsoftheencounter-relationshipgraphs(ERgraphs),inwhichtwonodesareconnectedbyalinkiftheyeverencounter.Wendthatalthoughdirectencountersofindividualnodeshappenonlytoasmallportionofnodepairsamongthewholepopulation,WLANusersformconnectedSmallWorldgraphsviaencounters,andthe 156

PAGE 157

4 weknowMNsinWLANtracesareinfactnotuniformlydistributed,anduserswithsimilarpreferencesshowupatthesameaccesspoint(AP)morefrequently.Welookintothisissueandtrytoidentifythecloseness(i.e.,friendship)betweennodepairs,andunderstanditsinuencesonnetworkconnectivityifwemakeconnectionsbetweennodesbasedontheirfriendship.Specically,wegiveseveralintuitivedenitionsoffriendshipbetweenMNs.ThesefriendshipindexescapturetheobservedclosenessbetweentheinvolvedMNsfromthetrace.Althoughsuchclosenessmayormaynotreectfriendshipinasocialcontext,itrevealstheclosenessbetweenwirelessdevicesasdisplayedintheirassociationpatterns.Theempiricaldistributionsofthesefriendshipindexesmostlyfollowtheexponentialdistribution,withfewnodepairsshowinghighfriendshipindex.Furthermore,weinvestigatetheissueofhowfriendshipinuencethecharacteristicsoftheencounter-relationship(ER)graphs.WendthatifonlynodeswithhighfriendshipindexesareusedinformingtheERgraph,theresultantgraphdisplayshigherclusteringcoecientandaveragepathlength.Inotherwords,itismoreinclinedtowardaregulargraph.Ontheotherhand,ifweuseonlynodeswithlowfriendshipindexintheERgraph,itdisplayslowerclusteringcoecientandaveragepathlength.Thisndingpointsout,similartosocialnetworks,closefriendsinWLANsoftenformcliquesandrandomfriendsarekeystowidely-reachedconnectivityinanetwork.Finally,weproposeinformationdiusionexperimentstounderstandhowinformationcouldbespreadamonguserswithoutthehelpofaninfrastructure.Weuseasimplemessagespreadingstrategytoinvestigatewhetheritispossibletorelyonmutualencounterstospreadmessagesacrossthenetwork.Surprisingly,giventheseeminglyverylowratioofthewholepopulationagivennodeencounterswith,theencountereventsformawide-reachingcommunicationnetwork,andthemessagesspreadtomostofthewhole 157

PAGE 158

6.2 andintroducetheSmallWorldapproachtoexplaintheencounter-relationshipgraphinsection 6.3 .WefurtherexplainthereasonfortheSmallWorldtoforminsection 6.4 .ThenwediscussthendingsinoureorttocapturefriendshipbetweenMNsinsection 6.5 .Finally,theinformationdiusionexperimentisexplainedinsection 6.6 .Weprovidesomediscussionsandconcludethechapterinsection 6.7 ,togetherwithanoutlineofourproposedfutureworkonanecientselectivebroadcastingprotocol. 3 forthedetails)fromWLANtraces.Althoughthesederivedencountertracesmaybenotcompletelyaccurate,webelievethattheencountereventsderivedfromWLANtracescaptureamajorportionofMNswithindirectcommunicationrangeundercurrentusagepattern.Thedistributionoftheseencountereventsistherststeptounderstandthestructureofinter-MNrelationshipinthetraces.Thedirectquestionstoaskabouttheencountereventsare:HowmanyotherMNsdoesausermeet?Donodesmeetwitheachotherrepeatedlyornot?Fig. 6-1 showstheCCDFoffractionofotherMNsagivenMNhasencounteredthroughthewholetraceperiod(i.e.,onemonth).Fromthegurewe 158

PAGE 159

11 ](inotherwords,theMNsinthistraceareselectedfromacorrelatedsub-groupofthewholepopulationoncampus).Inallothertraces,onaverageaMNencounterswithonly0:79%(UF)to6:70%(Dart-04)ofthewholeuserpopulationwithinthe30-daytraceperiod.Thesmallaverageencounterratioisacombinedresultofseveralreasons:(1)mostMNsarenotalwayson,and(2)mostMNsdonotvisitmanyAPs[ 66 ],hencetheycanonlymeetwiththosewhoalsovisitthissmallsetofAPs.Lowencounterpercentageasshowninthetracesisnotobservedinanyofthesimulationscenariosusedforperformanceevaluationintheliterature.Forexample,inFig. 6-2 ,weshowtheCCDFofuniqueencounterfractionobtainedfromtherandomdirectionmobilitymodel,oneofthecommonlyusedsyntheticmobilitymodel.Weobservetwoobviousdierencesfromtheempiricaltraces:(1)Theuniqueencounterfractionreaches100%forallnodeswithintwodays.Thisisbecause,intypicalsyntheticmobilityscenarios,asthosesummarizedin[ 49 ],allnodesfollowthesamemodeltomakemovementdecision,albeitwithrandomness,andeventuallyencounterwithallothernodes[ 46 ].(2)Thediversityoftheuniqueencounterfraction,givenanobservationtimeperiod,isnotveryhigh(e.g.,Withinsixhours,allnodesencounterbetween52%to75%ofthepopulation).Theencounterpatternfromrealwirelessnetworktraces,ontheotherhand,reectsthatuniversitycampusisaheterogeneousenvironmentratherthanahomogeneousoneconstructedbythesyntheticmodelsinwhichallnodesarestatisticallyi.i.d..Tobetterunderstandhowprotocolsperforminsuchheterogeneousenvironment,usinghomogeneoussyntheticmodelsisnotsucient.Thisndingaddstothemotivationofusingaexiblemobilitymodel,suchastheTVCmodelweproposeinchapter 4 ,which 159

PAGE 160

CCDFofuniqueencounterfraction,traces. iscapableofdescribingnodeswithdiverse,heterogeneousbehaviorforfutureprotocolevaluations.Ontheotherhand,fromtheCambridgetrace,mostofthe41usersmeetwiththemajorityofothersduringtheshorttraceduration(4days).Specically,thereare12MNswhomeetwithallother40MNs,and39outof41MNsmeetatleast38othernodes.ThecurveinFig. 6-1 ismostlyahorizontallineathighprobabilityuntiltheuniqueencounterfractionreaches0:95.Thishighuniqueencounterfractionmaybeduetotheenvironmentsetting(aconference,wherethepremisesisconsiderablysmallerthanauniversitycampusorcorporatebuildings,andpeoplearesupposedtomeetwitheachotherataconference)orthefactthattheselectionofparticipantsarerelated(i.e.,peoplewhoareinterestedinthestudyofmobilitypatternsandwirelessnetworksingeneral)ratherthanrandomlypickedfromtheconferenceattendees.WealsoshowtheCCDFofthetotalencountereventsaMNhasthroughoutthetraceperiodinFig. 6-3 .WeobservethetotalencountercountsforMNsineachtracespanacrossseveralordersofmagnitude.TherearebothMNswithextremelyfewormanyencounters.ThisisanevidenceofheterogeneousbehavioramongMNs.Theactualnumberoftotalencountersdependsonthesizeofpopulationinthetraces.Largetraces(i.e.,theUSCandDartmouthtraces)tendtohavemoreencountersthansmalltraces(i.e.,theUCSDandCambridgetraces).However,regardlessofthesizeof 160

PAGE 161

CCDFofuniqueencounterfraction,syntheticmodel(RandomDirectionmodel). population,thecurvesforthetotalencountercountderivedfromWLANtracesseemtofollowtheBiParetodistribution.WettheBiParetodistributioncurvestotheempiricaldistributioncurves,andusetheKolmogorov-Smirnovtest[ 98 ]toexaminethequalityoft.TheresultingD-statisticsforalltracesarebetween0:068and0:025,whichindicateswehaveareasonablygoodtbetweentheBiParetodistributioncurvesandtheempiricaldistributioncurves.ThedetailsabouttheKolmogorov-SmirnovtestandtheparametersofthettedBiParetodistributioncurvesarelistedinsection 6.8 .FortheCambridgetrace,thetotalencountercountsforMNsarenotasdiverseasthoseinWLANtraces.Thismaybeduetothefactthatmostnodesparticipatetheconferenceactivelythroughoutthewholetraceperiod(4days),butthisisunlikelyforthelonger,one-monthWLANtraces.TheBiParetodistributiondoesnotshowagoodtfortheCambridgetrace,asitstotalencounterdistributiondropssharplyata"knee"around250.AcloserinvestigationoftherelationshipbetweentheuniqueencountercountandthetotalencountercountofthesameMNrevealsthathighuniqueencountercountdoesnotalwaysimplyhightotalencountercount.Thecorrelationcoecientsbetweentheuniqueencountercountandthetotalencountercountforvarioustracesrangefrom0:732to0:195.ExceptfortheUCSDtrace,allothertraceshavecorrelationcoecientsbelow0:6.Asanillustration,weshowthescatterplotoftheuniqueencountercountversusthetotalencountercountfortheUSCtraceinFig. 6-4 .Weobservethatsome 161

PAGE 162

CCDFoftotalencountercount. Figure6-4. Uniqueencountercountversustotalencountercount,USC. nodeshavenotmanyuniqueencountercounts,buthightotalencountercounts.Thisindicatesthatsomenodepairsmayhavealotofrepetitiveencounters,suggestingthesenodepairshavecloserrelationshipthanotherpairs.Thispointwarrantsfurtherstudy,andwewillshowsomeinitialattemptsonquantifyingthefriendshipbetweenMNsinsection 6.5 6.2 ,weseethatMNshavelowpercentageofuniqueencoun-tersamongthewholepopulation.Giventhisfact,Weraiseaquestionregardingthepossibilityofestablishingcampus-widerelationshipsamongthemajorityofMNsviaencountersalone.Thatis,doencounterslinkMNsonthecampusintoonesinglecommunity,orjustmanysmallcliques? 162

PAGE 163

44 ]: 163

PAGE 164

6-5 .Thegraphsforothertracesshowverysimilartrends,andweleavetheminsection 6.9 tomaintainconcisenesshere. 164

PAGE 165

(b)Figure6-5. ChangeintheERgraphmetricswithrespecttotraceperiod.(a)Disconnectedratio.(b)Normalizedclusteringcoecientandaveragepathlength.Thegureiscutfromabovetoshowthedetailsbetween0and1onY-axis. FromFig. 6-5 (a)wenotethatgivensucientlongtracedurations,theERgraphshavelowDR(notlargerthan10%fortraceslongerthanonedayinmostcases),whichimpliesthatnodalencountersaresucienttoprovideoppor-tunitiestoconnectalmostallnodesinasinglecommunity,eventhougheachnodeencountersonlyasmallsubsetofMNsdirectly.Thisisanencouragingresultthatpointsoutthefeasibilityofbuildingalarge,widely-reachnetworkrelyingonlyondirectencounters.AlthoughtheDRstartsoutveryhighwithveryshorttraceperiods(i.e.,fortracedurationsunderoneday)sinceMNshavenotmovedaroundtocreateencountersyet,itdecreasesratherquicklyasthetraceperiodincreases.Withinoneday,DR'sreduce 165

PAGE 166

EquationsfortheCCandPLfortheregularandrandomgraphswithMnodesandaveragenodedegreed[ 8 44 ]. GraphtypeClusteringcoecientAveragepathlength Regulargraph3(d2)=4(d1)M=2dRandomgraphd=Mlog(d)=log(M) toaround10%.AlthoughthenumbersofMNsintheERgraphkeepincreasingaswelookatlongertraceperiods,inmostcasestheDRdoesnotchangesignicantlyafteroneday.Anotherinterestingndingisrevealedbytheothertwometrics,theclusteringcoecient(CC)andtheaveragepathlength(PL).TohighlightauniquepropertyoftheseERgraphs,wealsocalculatetheCCandthePLforregulargraphsandrandomgraphswiththesamecorrespondingtotalnodenumberMandaveragenodedegreed.ThesequantitiescanbecalculatedaccordingtoequationsinTable 6-1 .Intheregulargraphs,nodesarerstarrangedonacircleandeachnodeisconnectedtodclosestneighborsonthecircle.Intherandomgraphs,drandomlychosennodesareassignedasneighborsforeachnode.Typically,regulargraphshavehighCCandPLwhilerandomgraphshavelowCCandPL.Theyarethetwoextremecasesonthespectrum.InFig. 6-5 (b),weshowthenormalizedCC'sandPL'softheERgraphsforvarioustraceperiods.Thesenormalizedmetricsrepresent,onthescalefrom0(correspondingtotherandomgraph)to1(correspondingtotheregulargraph),wherethemetricsoftheERgraphsfall.Theyaredenedas:CCnorm=CCCCrand 166

PAGE 167

8 ],[ 44 ].Bylookingatvarioustraces,weindicatethattheERgraphsformedbyencountersamongnodesusingwirelessnetworkappeartobeSmallWorldgraphs.WealsoobservethatbothPLandCCconvergestoitsnalvaluesratherquicklyinaboutonedayformosttraces,althoughthesizeofERgraphskeepsincreasingasmorenodesappearinlongertraces.FortheCambridgetrace,welookintosimilarmetrics.Wendthatforevenasmallperiodoftime(e.g.,1day)the41MNsencountermostofthewholepopulation.Hence,theCCisveryhigh(above0:91evenifwetakeonlytherstdayintoconsideration),andthePLislow(lessthan1:1).Actually,the41MNspresentedinthetracealmostformafully-connectedmesh,andtheDRis0.Thismaybepartlyduethenatureoftheconferencesettingfromwhichthetracewascollected.Peoplemovearoundtomeetmoreoftenthanintheirregulardailylifeatuniversitiesorcorporations,hencetheencounterpatternataconferenceseemstobericherthaninregularenvironments.The 167

PAGE 168

5.5.1 (seeEq.( 5{11 ))andtheSmallWorldgraphstovalidatethisintuition.Wedevisethefollowingexperimenttounderstandtheeectofmutualsimilaritiesbetweenusers'associationpatternsontheglobalencounterpatterns.UsingUSCtraceasanexample,wecategorizealluserpairsintofourzones,asillustratedinFig. 6-6 .ZoneAconsistsofuserpairswhoarehighlysimilar(withthesimilaritymetricabove0:8),andzoneB,C,andDconsistofuserpairswithlesssimilarityineachzone.Theboundariesbetweenthezonesaresochosenthat,whenweconsideranaverageuser,ithasroughlysimilarnumberofencounteredusersfallingineachzone.Afterdesignatinguserpairsintozones,weredrawtheERgraphstoincludeonlylinksbetweentwonodesinthegraphifthenodepairbelongstoacertainzone.Thisisaneorttoevaluatehowlinksamongsimilarordissimilarusersplayitsrolesintheresulting 168

PAGE 169

Classicationofnodepairsintodierentcategoriesbasedontheirsimilaritymetricrange. Table6-2. ThegraphpropertiesoftheERgraphswithselectedlinks(onlylinksfallingintocertainsimilaritycategories(seeFig. 6-6 forthebins)areincluded). Averagenodedegree72.4872.1662.2762.73144.62134.43125.00206.89197.16269.62DisconnectedRatio(%)96.858.9811.357.256.364.224.262.401.490.53ClusteringCoecient0.78140.45680.17370.29680.69730.48960.35780.63390.50030.6117AveragePathLength1.5373.1022.6382.5633.0922.3972.3592.3852.2362.200 6-2 .WeseefromTable 6-2 thatwhentheERgraphsincludeonlyedgesfromonezone,undersimilaraveragenodedegreeintheERgraph(wehavechosenthecategorizationbinscarefullytoensurethis),iftheedgesareformedbetweennodeswithhighsimilarity,itresultsinhighdisconnectedratioandclusteringcoecientingeneral.ThistrendisespeciallypronouncedfortheERgraphincludingonlyedgesinzoneA,validatingourintuitionthatextremelysimilarnodes(intermsoftheirmobilitypreferences)formdisjointclusters.Thenodepairsthataredissimilartoeachother(e.g.,nodepairsinzoneD)leadtoanERgraphwithlowdisconnectedratio,lowclusteringcoecientandlowaveragepathlength 169

PAGE 170

170

PAGE 171

6-7 ,theCCDFcurvesoffriendshipindexesbasedonencountertimefollowexponentialdistributionsforallcampuses.WeagainusetheKolmogorov-Smirnovtest[ 98 ]toexaminethequalityoft.TheresultingD-statisticsforalltracesarebetween0:0356and0:0052,whichindicateswehaveareasonablygoodtbetweentheexponentialdistributioncurvesandtheempiricaldistributioncurves.Theactualparametersweuseforthettingarelistedinsection 6.8 .Theexponentialdistributionofthefriendshipindexesisanindicationthatthemajorityofnodesdonothavetightrelationshipwithoneanother.Inallthetraces,onlylessthan5%oforderednodepairs(a;b)havefriendshipindexFrdt(a;b)largerthan0:01.Thisrevealsthefactthatfornodepairsthatdoencounterwitheachother,mostofthemdonotshowstrongrelationship.Amongallnodepairswithnon-zerofriendshipindex,only4:47%ofthemhavefriendshipindexlargerthan0:7,andanother11:85%ofthemwithfriendshipindexbetween0:4to0:7.Inotherwords,wecansaythatthefriendshipbetweentheMNsisvery\sparse"(i.e.,onlyfewpairsofnodescanbecalled\friends" 171

PAGE 172

CCDFoffriendshipindexbasedontime. Table6-3. Correlationcoecientforfriendshipindexesforalltraces. TracenameFriendshipindexbasedon encountertimeencountercountAPcount MIT-rel0.4150.3270.186UCSD-0.024-0.004-0.003USC0.1580.2050.130Dart-030.3510.2780.043Dart-040.6290.2010.068UF0.1900.0910.036 basedontheabovedenitions).FriendshipindexesbasedonencounterfrequencyorencounterAPcountalsoshowsimilarexponentialdistributions.WealsolookintotheissueofwhetherthefriendshipindexforanorderednodepairFrdt(a;b)andthereversedtupleFrdt(b;a)aresymmetric.Wecalculatedthecorrelationcoecientsforallthetracesforthreedenitionsoffriendshipindexes,asshowninTable 6-3 .Theresultingcorrelationcoecientsbetweenorderednodepair(a;b)and(b;a)arelowinmostcases(rangingfrom0:415to0:024,theonlyexceptionbeing0:629forfriendshipindexbasedonencountertimeforDartmouth2004trace),implyinghighasymmetryinfriendshipindexes.AfterseeingthesparsenessandhighasymmetryofthefriendshiprelationshipbetweentheMNs,weaskthefollowingquestion:ifweconsiderfriendshipinestablishingrelationshipsbetweennodes,howwouldthatinuencethestructureoftheencounter-relationshipgraphs?Typically,aMNmaynotmaintainrelationshipswithallotherMNsitencounters 172

PAGE 173

6{2 )bythefollowing 6.3 ,butthepathsmustfollowthedirectionofedgesontheERgraph.Followingtheabovedenitions,weobtainthemetricswhenincludinggivenpercentagesofallencounterednodesfromthetop,middle,orbottomofthesorted 173

PAGE 174

(b) (c)Figure6-8. Metricsofencounter-relationshipgraphbytakingvariouspercentageoffriends.(a)Clusteringcoecient.(b)Averagepathlength.(c)Disconnectedratio. nodelistaccordingtothefriendshipindexbasedontime.TheguresareshowninFig. 6-8 .WeusetheUSCtraceasanexample,andsimilarresultsarealsoobservedinothertraces.Theguresshowacleartrendthatifneighborsrankedhighinthefriendshipindexareincluded,theresultantERgraphshowsstrongerclustering,andtheaveragepathlengthismuchhigher.Theresultstemsfromthefactthattopfriendsofagivennodearealsolikelytobetopfriendbetweenoneanother,formingsmallcliquesinthegraph.Theclusteringcoecientremainshighduetothesecliques.Thedisconnectedratioandtheaveragepathlengthsarehighduetothelackoflinksbetweendierentcliques.Ontheotherhand,whenlow-rankedfriendsareincludedinthegraph,thelinksincludedaredistributedinamorerandomfashion,reectedbythelowclusteringcoecientandlowaveragepathlength.Similarresultsarealsoobservedinasocialsciencestudyoffriendshipbetweenpupils[ 99 ].Asalargerportionoffriendsareincludedinthegraph,allthreemetricsconvergetothevalueswhenallencountersareincluded 6.3 ). 174

PAGE 175

6.6.1 .Wethenremovesomeoftheassumptionsandevaluatetheperformanceinmorerealisticsettingsinsubsequentsubsections. 6.6.2 and 6.6.3 dealwithmorerealisticscenarioswhensome 175

PAGE 176

71 ].Underperfectenvironmentwithsucientresources,itachievesthelowestdelayandthehighestdeliveryratepossible.Inallthesimulations(inthisandthesubsequentsubsections),weuseatracpatterninwhichthesourcenodehassomeinformationtosendtoallothernodes.Thesourcestartsto\diuse"theinformationwhenitisrstonline.Astimeevolves,nodesencounterwitheachotherandanincreasingportionofthewholepopulationreceivetheinformation.Westudythepercentageofnodesthathavereceivedtheinformationwithinvarioustraceperiods(i.e.,thenumberMNsthathavereceivedthemessageoverthetotalMNsthathaveappearedduringthetraceperiodunderdiscussion)andshowtheresultsinFig. 6-9 ,usingtheUSC,Dart-04,Dart-03,andMITtracesasexamples.Eachpointintheguresofthissectionisanaveragevalueofmultipleexperiments.Ineachexperimentwestarttheinformationdiusionfromadierentsourcenode.Wechoosetouse30%ofthenodesthatappeartheearliestinthecorrespondingtraceperiodasthesources.FromFig. 6-9 weobservethatevenwithinashorttraceperiod(e.g.,twodays)theinformationcanreachamoderateportionofthepopulationastheunreachableratioislessthan25%inalltraces.Asthetraceperiodincreases,reachabilityalsoimproves.InallexcepttheDart-03trace,theunreachableratiosarelessthan2%ifweallowonemonthfortheinformationdiusion.Giventhatmostnodesencounterwithonlyaverysmallportionofthewholepopulation(Fig. 6-1 ),thisresultisperhapsbeyondouroriginalexpectation.Itgivesapositiveconrmationthatitispotentiallypossible

PAGE 177

Unreachableratioofinformationdiusionusingtheepidemicrouting. Figure6-10. USCtrace:Unreachableratiowithvariousselshnodepercentageandtraceperiod. 177

PAGE 178

6-10 .Forthesakeofconciseness,weonlyshowguresfortheUSCtracehere.Theguresforothertracesdisplaysimilartrendsandtheyareshowninsection 6.9 .Theresultisverysurprising{foralltraceperiodtested,theunreachableratiodoesnotincreasesignicantlybeforeatleast20%ofnodesareselsh.Theperformanceisevenmorerobustifwetakelongerperiodoftrace.Thisimpliesthatevenasignicantportionofusersarenotwillingtopropagateinformationforothers,theunderlyingnodalencounterpatternisrichenoughfortheinformationtondanalternativewaythrough.Hencethedeliveryrateisquiterobustforuptoanintermediatepercentageofselshnodes.NotethatwemaketheMNswithmostuniqueencountersselshrst,hencetheperformanceofinformationdiusionisrobustevenifthenodeswiththemostchancestopropagatetheinformationarenotcooperative.WefurthershowhowtheaveragedelayofinformationdiusionchangeswiththeincreasingselshnodepercentageinFig. 6-11 fortheUSCtrace.Inthegure,theaveragedelayincreasesforlongertracedurationbecauseinformationthatisnotdeliverableinshortertraceperiodsbecomesdeliverable.Moreinterestingly,foralltestedtracedurations,theaveragedelaydoesnotincreasesignicantlybeforemorethan40%ofthenodesareselsh.Thisimpliestheaveragedelayisalsorobustagainstselshuserbehavioruptoanintermediatepercentage. 178

PAGE 179

USCtrace:Averagemessagedelaywithvariousselshnodepercentageandtraceperiod. notpermitpromptdiscoveryandusefulinformationexchangeinthefollowingexperiment,andre-evaluatetheperformanceofinformationdiusionwithdierentminimumdurationthresholdsforanencountereventtobeconsidereduseable.InFig. 6-12 ,weshowtherelationshipbetweentheunreachableratioversusthelowerlimitofencounterduration(i.e.,weremoveallencountereventsthathaveshorterdurationsthanthevalue),usingtherst15-daytracesfromUSCandDartmouthasexamples.Fromthegraphweobservethat,theunreachableratioincreasesalmostlinearlyasweincreasethelowerlimitofusableencounterduration.Thereisnoobviouspointatwhichtheperformancesuddenlydegradesseverely.Wecarryouttheexperimentsuptotheshortestusableencounterthresholdsetatonehour,aratherdemandingscenario.Eveninsuchcases,besidestheUFtracewhichhasaverylowencounterratio(seeFig. 6-1 ),theunreachableratioisbelow30%.Thisimpliesremovingencounterswithshortdurationsdoesnotcauseabruptdegradationintheperformanceofinformationdiusion,intermsofboththereachabilityandtheaveragedelay(seeFig. 6-13 ).Inotherwords,shortencountersarenotthekeyreasonforthesuccessofinformationdiusion.Theencountereventswithlongdurationsarealsorichenoughtobeutilizedformessagepropagationinmostcases. 179

PAGE 180

Theunreachableratioafterremovingshortencountersunderthedurationlowerlimit. Figure6-13. Thedelayafterremovingshortencountersunderthedurationlowerlimit. 180

PAGE 181

104 ].Bothobservationsareduetotherichnessintheunderlyingencounterpatternprovidingabundantchancesformessagedelivery.Theperformanceofinformationdiusionundervariousinformationdeliveryschemesandpotentialmethodstopreventmaliciousinformationfromspreadingarebothdirectionsforfuturework. 181

PAGE 182

71 ]basedonwhichweshowthepotentialofencounter-basedinformationdiusioninthischapter.Thistaskwouldbeourmainfocusinthenextchapter. 6.2 )andfriendshipindex(section 6.5 )distributions,respectively.TheBiParetodistributionisusedin[ 101 ]totthenumberofconnectionsperuserTCPsessionandmeanconnectioninter-arrivaltimeinaTCPsession.Later,BiParetodistributionisagainusedin[ 14 ]totthedistributionofassociationsessionlengthinwirelessLAN.TheCCDFofBiParetodistributionisasfollows:Prob(X>x)=(x k)(x+c k+c);x>k 98 ].ThereforewechoosetheK-Stestinourstudy. 182

PAGE 183

IllustrationoftheD-statisticsandtheK-Stest. ReferringtoFig. 6-14 ,intheK-Stestthedistancesbetweenthehypothesizeddistributionandtheempiricaldistributionaremeasuredthroughouttherangeofrandomvariablex,andthemaximumofthemeasureddistancesiscalledtheD-statistics.Moreformally,theD-statisticsisdenedas: 6-4 .FromthetableweobservethattheD-statisticsarenolargerthan0:05exceptforUCSDtrace(0:07),indicatingareasonabletoftheBiParetodistribution.WealsolisttheparametersweobtainedusingtheminimumsquarederrormethodtotexponentialdistributionstotheempiricaldistributionoffriendshipindexesbasedonencountertimeinTable 6-5 .ThecorrespondingD-statisticsarealsolisted. 183

PAGE 184

BiParetodistributionttingtothetotalencountercurvesandtheD-statisticsfortheK-Stest. TracenameBiParetoparametersD-statistics Table6-5. ExponentialdistributionttingtothefriendshipindexbasedonencountertimecurvesandtheD-statisticsfortheK-Stest TracenameD-statistics MIT-rel369.190.0167USC305.30.0356Dart-03500.40.0052Dart-04411.810.0116Dart-rel409.910.0120Dart-cons412.350.0119UF579.060.0023 6.3 ,wealsoobtainthesamemetricsforMIT,Dart-03,andUF 6-15 )havesimilartrendsasdiscussedinsection 6.3 .OneinterestingobservationhereisthatfortheMITtrace,thedisconnectedratioisveryhighuntilday3inthetrace.AfurtherinvestigationrevealsthattheMITtracecollectionwasstartedonaSaturday,andforapureworkingenvironment(i.e.,corporatebuildings)SaturdaysandSundaysaretheleastactivedays.Thedisconnectedratioisalmost100%untilday3becausetheMNsthatwereonduringtheweekendaremostlystationaryones.Weobserveajumpofnumberofnodeinthetrace,asuddendecreaseinDR,andanabruptchangeinbothCCandPLondaythree.FortheUFtrace, 112 ]formoredetailedandupdatedresults. 184

PAGE 185

(b)Figure6-15. ChangeintheERgraphmetricswithrespecttotraceperiod.(a)Disconnectedratio.(b)Normalizedclusteringcoecientandaveragepathlength. basedonthe10;000sampleduser,theDRfor30-daytraceis8:85%,thenormalizedCCis0:584andthenormalizedPLis0:099.Weperformfullanalysis(basedonall32;695usersthatappearedinthe30-daytrace)inordertounderstandtheeectofrandomsamplingontheabovemetrics.Theresultsareasfollows:DR1:94%,CC0:566,andPL0:039.ItappearstheadditionalusersinthefulltraceleadtoasignicantdecreaseintheDRandPL,duetoaddedconnectivity,buttheCCremainssimilar.Moredetailedanalysiscanbefoundat[ 112 ].InadditiontotheUSCtrace,wefurtherperformsimilarinformationdiusionexperimentsonaddingselshuserbehaviortotheDartmouth,MIT,andUFtraces.Theexperimentsetupisthesameasdescribedinsubsection 6.6.2 .TheresultsfortheaverageunreachableratioareshowninFig. 6-16 6-18 6-20 ,and 6-22 fortheDart-04, 185

PAGE 186

6.6.2 .Forlongertraceperiods(above9days),theunreachableratiodoesnotchangesignicantlyforupto20%ofselshnodes,andtherobustnessofperformanceincreasesiflongertraceperiodsareused.ThisconrmsthattherobustnessofinformationdiusionundercurrentencounterpatternsisnotanartifactofcoarselocationgranularityintheUSCtrace.IntheDart-03andUFtraces,theperformanceofinformationdiusionislessrobustthanothertraces,sincetheyhavethesmallerencounterratio(cf.Fig. 6-1 )amongallthetraces 6-17 6-19 6-21 ,and 6-23 fortheDart-04,MIT,Dart-03,andUFtraces,respectively.TheresultsaresimilartoFig. 6-11 insubsection 6.6.2 .Onenoticeabledierenceisthat,insomecasestheaveragedelayrstincreasesastheselshnodepercentageincreases,butlateritdecreases.Thisisduetothelowreachability(i.e.,highunreachableratio){inthissituation,onlyMNsthatareeasytoreachwillbeabletoreceivethemessage,leadingtoadecreaseintheaveragedelay(calculatedfromthesmallsubgroupofstillreachableMNs). 112 ]. 186

PAGE 187

Dart-04trace:Unreachableratiowithvariousselshnodepercentageandtraceperiod. Figure6-17. Dart-04trace:Averagemessagedelaywithvariousselshnodepercentageandtraceperiod. Figure6-18. MITtrace:Unreachableratiowithvariousselshnodepercentageandtraceperiod. 187

PAGE 188

MITtrace:Averagemessagedelaywithvariousselshnodepercentageandtraceperiod. Figure6-20. Dart-03trace:Unreachableratiowithvariousselshnodepercentageandtraceperiod. Figure6-21. Dart-03trace:Averagemessagedelaywithvariousselshnodepercentageandtraceperiod. 188

PAGE 189

UFtrace:Unreachableratiowithvariousselshnodepercentageandtraceperiod. Figure6-23. UFtrace:Averagemessagedelaywithvariousselshnodepercentageandtraceperiod. 189

PAGE 190

5.8 .Insuchaparadigm,messagesaresenttoinferredbehavioralproles,insteadofexplicitIDs.Usingbehavioralprolespacegradientsandsmallworldstructures,weprovidefullydistributedandmoregenericmessagedisseminationprotocols,namedCSI,relyingontheImplicityetStablerelationshipdiscoveredbetweenmobileusers.ThechoiceofmessagetargetinCSIismoregeneric.Onecanchooseatargetbehavioralproleeitherwiththesamerepresentationastheusereigen-behavior(i.e.,themobilitypreferences)orinatotallyorthogonalcontext. 190

PAGE 191

191

PAGE 193

Illustrationoftheassociationmatrixtodescribeagivenuser'slocationvisitingpreference. (1)Weintroducethenotionofmulti-dimensionalbehavioralspace,anddevisearepresentationofuserbehavioralprolestomapusersintothebehavioralspace.Ourstudyisthersttoestablishconditionsforstabilityoftherelationshipbetweencampususersinthisspace.(2)WeproposeCSI,anewcommunicationparadigmdeliveringmessagebasedonuserproles.ThetargetproleinCSIcanevenbeindependentofthecontextofbehavioralproleweusetoconstructthebehavioralspace.(3)WedesignanecientdisseminationprotocolutilizingthestabilityofbehavioralprolesandSmallWorldinmobilesocieties,thenempiricallyevaluateandvalidatetheecacyofourproposalusinglarge-scaletracesfromuniversitycampuses.Theoutlineofthechapterisasfollows.Wesummarizetheimportantbackgroundfrompreviouschaptersinsection 7.2 .Thisisfollowedbyananalysistounderstandtheuserbehavioralpatterninsection 7.3 .Wefurtherdiscussthepotentialusagesofthisunderstandinginsection 7.4 anddesignourCSIschemesinsection 7.5 asanexample.WeusesimulationstoevaluatetheperformanceofCSIschemesinsection 7.6 .Finally,wediscusssomenerpointsinsection 7.7 andconcludeinsection 7.8 7.2.1Mobility-BasedUserBehaviorRepresentationWerepresentmobileuserbehaviorofagivenuserusingtheassociationmatrixasdenedinchapter 5 andillustratedinFig. 7-1 .Inthematrix,eachrowvectordescribesthepercentageoftimetheuserspendsateachlocationonaday,reectingtheimportance 193

PAGE 194

5 ithasbeenshownthatthelocationvisitingpreferencescanbeleveragedtoclassifyusersofwirelessnetworksonuniversitycampuses.Foragivenuser,thesingularvaluedecomposition(SVD)[ 41 ]isappliedtoitsassociationmatrixM,suchthat 5{11 )butreproducedhereforclarity),vectorsai'sandbj'sandthecorrespondingweights,as 3-1 .Theinformationavailablefromtheseanonymizedtracescontainsmanyaspectsofthenetworkusage(e.g.,time-locationinformationoftheusersbytrackingtheassociationanddisassociationeventswiththeaccesspoints,amountoftracsent/received,etc.). 194

PAGE 195

Illustration:considerthetrailingddaysofbehavioralproleattimepointsthatareTdaysapart. Therichnessinuserbehavioraldataposesachallengeinrepresentingtheuserbehaviorinameaningfulway,suchthattherepresentationnotonlyrevealsanintrinsic,stablebehavioralproleofauser,buttheidentiedbehavioralprolealsoleadstopracticalapplications.Weshowherethatthelocationvisitingpreferences(whichisonlyasubsetoftheuserbehavioraldata)isastableattributeforbothindividualusersandtherelationshipbetweenusers.Thispropertywillprovequitevaluabletothedesignofecientmessagedisseminationschemes,whichweempiricallyvalidateusingtheabovetraces. 38 73 111 ]areusefulandinlinewithourgoal,thestabilityofsuchclassicationovertimehasnotbeenstudiedsystematically.Inparticular,theshort-termbehaviorofausermaydeviatesignicantlyfromthenorm,andthestabilityofuserbehavioralprolesisadecisivefactorforwhetheritcanbeleveragedtorepresenttheuser'sfuturebehavior.Inthissectionweinvestigatethefollowingquestions:(1)Howlongofbehavioralhistorydoweneedtoclassifyauser?and(2)Howmuchdoesthebehaviorofagivenuseranditsrelationshipwithotheruserschangewithrespecttotime?Weconsidertheeectoftheamountofpasthistory(ofuserbehavior)onitsbe-havioralproles.Eachuserusesthelocationvisitingpreferencevectorsinthepastddaystosummarizethebehaviorinthemostrecenthistory{theuserretainsdlocationvisitingpreferencevectorsforthesedays,organizetheminamatrix,andusesingular 195

PAGE 196

SimilaritymetricsforthesameuserattimegapTapart. Figure7-4. CorrelationcoecientofthesimilaritymetricsbetweenthesameuserpairattimegapTapart. valuedecompositiontoobtainthebehavioralprole,asdescribedinsection 7.2.1 .Weseektounderstandhowdinuencestherepresentationandsimilaritycalculations.Morespecically,welookintotwoimportantaspects:(1)Whethertherepresentationofagivenuserisstableacrosstime,and(2)whethertherelationshipsbetweenuserpairsremainstableastimeevolves.Werstconsiderthestabilityoftherepresentationofagivenuser.ConsideringtwopointsintimethatareTdaysapart,weobtainthebehavioralprolesforthesameuseratbothendpoints,usingthelogsofthetrailingddaysendingatthoseendpoints,asillustratedinFig. 7-2 .ThenweusethesimilaritymetricdenedinEq.( 7{2 )tocomparehowstableauser'sbehavioralproleistoone'sformerselfafterTdayshaselapsed.Theaverageresultswithvariousvaluesofthetimegap,T,andconsideredbehavioralhistory 196

PAGE 197

7-3 .Wenoticethat,evenifwecollectashorthistoryofuserbehavior(sayd=3),therepresentationissimilartothebehavioroftheuserforalongtimeintothefuture.WhenweconsiderT=35daysapart,thebehavioralprolesfromthesameuserstillshowhighsimilarity,atabout0:6.TheamountofhistoryuseddoesnotinuencetheresulttoomuchwhentheconsideredTislargeenoughtoavoidoverlapsintheusedbehavioralhistory(i.e.,whenT>d).Weconcludethatonuniversitycampuses,thebehavioralproleforagivenuserisstable,i.e.,itremainshighlysimilarforthesameuseracrosstime.Oneinterestingnoteisthat,whenthebehavioralproleincludesonlypartofaweek(d<7),thesimilarityoftheusertoitsformerselfshowsaweeklypattern(i.e.,whenTisanintegermultipleofseven,thesimilaritypeaks),especiallyinUSC.Second,wetrytoquantifyhowthebehavioralsimilaritybetweenthesamepairofusersvarieswithtime.Forthispart,weuseEq.( 7{2 )tocalculatethesimilaritybetweentwousers,AandB,attwopointsintime,SimT1(A;B)andSimT2(A;B),whereT1andT2areTdaysapart.Weperformthiscalculationtoalluserpairs,andthencalculatethecorrelationcoecientofthesimilaritymetricsobtainedafteraT-dayinterval,as 7-4 .Weobservethatthesimilaritymetricsbetweenuserpairscorrelatereasonablywelliftheconsideredtimeperiodsarenotfarapart.ForTsmallerthanoneweek,thecorrelationcoecientisabove0:62.Thisindicates,oncethesimilaritybetweenapairofuserisobtained,itremainsareasonablepredictorfortheirmutualrelationshipforsometimeperiodintothefuture.Althoughthereliabilityofthestalesimilaritydatadecreaseswithrespecttotime,thecurrentsimilarityofauserpairremainsmoderatelycorrelatedto 197

PAGE 198

61 ],inwhicheachuser'sbehaviorisquantiedasahighdimensionalpoint 108 ]architecturewherenodeswithhighmobility 198

PAGE 199

7.2.1 .Thisisanindividualeortbyeachnodeinvolvingnointer-nodeinteractions.Thiscanbedonebythenodesover-hearingthebeaconsignalsfromthexedaccesspointsintheenvironmenttondoutitscurrentlocation.Notethat,theuseofthesebeaconsignalsisonlyforthenodetoproleitsownbehavior{theyarenotusedtohelpthecommunicationinourprotocols(wewillre-visitdetailedpointsofthisassumptioninsection 7.7 ).Also,fortheeaseofunderstanding,weassumeinthis 199

PAGE 200

7.7 .ThegoalofourCSIschemeistoreachagroupofnodesmatchingwiththetargetprolespeciedbythesender,underthefollowingperformancerequirements:(1)Theprotocolshouldbescalable,inparticularnotbeingdependentonacentralizeddirectorytomaptargetprolestouseridentities.(2)Itshouldworkinanecientmannerandavoidtransmissionandstorageoverheadwhenpossible.Also,itshouldavoidcontrolmessageexchangesintheabsenceofdatatrac.(3)Thesyntaxofthetargetproleshouldbeexible,allowingthetargetproletobenotinthesamecontextasthebehavioralprolesweusetorepresenttheusers.Alsotheoperationoftheprotocolshouldbeexibletoallowtradeobetweenvariousperformancemetrics.Andnally,(4)thedesignshouldberobustandhelpinprotectinguserprivacy.WedesigntwomodesofoperationfortheCSIschemeundertheaboverequirements.Whenthetargetproleisinthesamecontextasthebehavioralprole(inourexample,sincethebehavioralproleisasummaryofusermobility,thiscorrespondstothescenariowhenthetargetproledescribesusersthatmoveinaparticularway),theCSI:Targetmode(CSI:T)shouldbeused.Whenthetargetproleisirrelevanttothebehavioralprole(e.g.,whenIwanttosendtoeveryoneinterestedinmoviesoncampus),theCSI:Dmodeshouldbeusedinstead.AlthoughitseemsthattheapplicabilityofCSI:Tislimited,wenotethatthebehavioralprole(intermsmobility)cansometimesbeusedtoinferothersocialaspectsoftheusers,suchasaliationsoreveninterests(e.g.,peoplewhovisitthegymoftenshouldlikesportsingeneral).SuchinferencesexpandthescenariosinwhichCSI:Tcanbeused.Whenthisisnotpossible,CSI:Disseminationmode(CSI:D)providesamoregenericoption.Themajorchallengeinvolvedinthedesignprocessisthateachnodeisonlyawareofthebehavioralproleofitself.Furthermore,werequirenopersistentcontrolmessageexchangesforthenodesto\learn"thestructureofthenetworkproactivelywhenthey 200

PAGE 201

(b) (c)Figure7-5. Relationshipbetweenthesimilarityinbehavioralpatternandotherquantities.(a)Totalencounterduration.(b)Encounterprobability.(c)Similarityofencounterednodesets. havenomessagetosend.Nodesonlycomparetheirbehavioralproleswhentheyareinvolvedinmessagedissemination.Basedonthisverylimitedknowledgeaboutthebehavioralspace,anodemustpredicthowusefulagivenencounteropportunityisintermsofachievingthefore-mentionedrequirements.Sinceencountereventsmayoccursporadicallyinsparse,opportunisticnetworks,thenodesmustmakethisdecisionforeachencountereventindependentofotherencounterevents(thatmayoccurlongbeforeorafterthecurrentoneunderconsideration).Suchaheuristicmustrelyontheunderstandingoftherelationshipbetweennodalbehavioralprolesandencounters,whichwediscussthenext. 201

PAGE 202

7{2 )),andobtainvariouscharacteristicsofencountereventsasafunctionofthepair-wisebehavioralsimilarity.InFig. 7-5 (a),weshowtheaggregateencountertimedurationbetweenanaveragepairofnodesgiventhebehavioralsimilarity.InFig. 7-5 (b),weshowtheprobabilityforagivennodepairtoencounterwitheachother,giventheirsimilarity.Combiningthesetwographs,weseethatiftwousersaresimilarinbehavioralproles,theyaremuchmorelikelytoencounter,andthetotaltimetheyencounterwitheachotherismuchlonger{anindicationthatnodeswithsimilarbehavioralprolesindeedaremorelikelytohavebetteropportunitiestocommunicatedirectly.Whentwousersaresimilarenough(withbehavioralsimilaritylargerthan0:3),theyarealmostguaranteedtoencounteratsomepoint(withprobabilityabove0:9).However,wenotethatsome\random"encountereventshappenbetweendissimilarusers.Foruserswithverylow(almostzero)similarity,theprobabilityforthemtoencounterisnotzero,althoughsuchencountereventsaremuchlessreliable(i.e.,theyoccurwithmuchshorterdurations,seeFig. 7-5 (a)).InFig. 7-5 (c)wefurthercomparethebehavioralsimilarityofnodeAandBversusthesetsofnodesAandBencounter.WedenotethesetofnodesAencounterswithasE(A).ThesimilarityofthetwosetsofnodesisquantiedbyjE(A)\E(B)j=jE(A)[E(B)j,wherejjisthecardinalityoftheset.Thisgraphshows,astwonodesareincreasinglysimilar,thereislargerintersectionofnodestheyencounter.Whenanunlikelyencountereventbetweendissimilarnodesoccurs,ithelpsbothnodestogainaccesstoaverydierentsetofnodes,whichtheyareunlikelytoencounterdirectly.

PAGE 203

60 ]wediscussinthepreviouschapter.ThekeyfeaturesofSmallWorldnetworks[ 8 ]arehighclusteringcoecientandlowaveragepathlength.Inthehumannetworksweanalyzeinthissection,peoplewithsimilarbehaviorform\cliques".The\random"encountereventsbetweendissimilarnodesbuildshort-cutsbetweenthesecliquestoshortenthedistancesbetweenanytwonodes.Weleveragethesepropertiesintheprotocoldesign. 7-6 asanillustration.Aspersection 7.5.2 ,todelivermessagestoreceiversdenedbyagivenTP,onewayistograduallymovethemessagetowardsnodeswithincreasingsimilaritytotheTPviaencounters,inthehopethatsuchtransmissionswillimprovetheprobabilityofencounteringtheintendedreceivers.Finally,whenthemessagereachesanodeclosetotheTP(inthebehavioralspace),mostnodesencounterfrequentlywiththisnodearealso 203

PAGE 204

1 .Therearetwophasesintheoperation,thegradientascendphaseandthegroupspreadphase.(1)Startingfromthesender,ifnodeAcurrentlyholdingthemessageisnotanintendedreceiver(i.e.,Sim(BP(A);TP)
PAGE 205

IllustrationoftheCSI:Tschemeinthehighdimensionbehavioralspace.Onecopyofthemessagefollowsincreasingsimilaritygradienttoreachtheneighborhoodofthetargetprole,thentriggersgroupspread. ifSim(BP(A);TP)>thsimthen spread();else ascend(); ascend()fwhilethemessageisnotsentdo foreachnodeEencountereddo spread()fforeachnodeEencountereddo 205

PAGE 206

DesignphilosophyoftheCSI:Dscheme.Leftchart:Thegoalistosendamessagetoagroupofnodeswithasimilarcharacteristicintheinterestspace(whitenodesinthecircle).Rightchart:However,theymaynotbesimilartoeachotherinthebehavioralspace(nodeswiththesamelegendrepresentsimilarnodesinthebehavioralspace). thereislittleinsightprovidedbythesimilaritiesbetweenthenodalbehavioralprolestoguidemessagepropagation,astheintendedreceiversinthiscasemaybescatteredinthebehavioralspace,andtherelationshipbetweenthetargetproleandthebehavioralprolecannotbequantied.Althoughitisalwayspossibletoreachmostusersthroughepidemicrouting,aswehaveshownitsrobustnessinthepreviouschapter(seesection 6.6 ),thisleadstohighoverhead,andrequiresallnodesinthenetworktokeepacopyofthemessage.TheobjectiveofCSI:Dmodeistoreducethenumbersofmessagecopiestransmittedandstoredinthenetwork,yetmakeitpossibleformostnodestogetacopyquickly,iftheybelongtotheintendedreceivers.WeagainrstdiscusstheintuitionbehindthedesignoftheCSI:Dmodeinthisparagraph,usingFig. 7-8 asanillustration.Fromsection 7.5.2 ,sincethenodeswithhighsimilarityintheirbehavioralprolesarealmostguaranteedtoencounter,thereisreallynoneedforeachofthemtokeepacopyanddisseminatethemessage.Electingafewmessageholderswithinasinglegroupofsimilarnodeswouldsuce.ThisintuitionleadstotheconstructionofourmessagedisseminationstrategyfortheCSI:D.Weaimtohaveonlyonemessageholderamongthenodeswhoaresimilarintheirbehavioralproles(orequivalently,pickonlyonemessageholderwithinaneighborhoodinthebehavioralspace.InFig. 7-7 ,thiscorrespondstohavingonlyone 206

PAGE 207

7-5 (c),weshouldselectnodesthatareverydissimilarintheirbehavioralprolestoachievelowoverlaps.Recallthatdissimilarnodepairsstillencounterwithnon-zeroprobability,ourdesignphilosophyistoleveragethese\random"encountereventsasshort-cutstonavigatethroughthebehavioralspaceeciently,hoppingacrossthespacetoreachdissimilarnodeswithrelativelyfewmessagetransmissions.SuchadesignphilosophyisalsorelatedtotheSmallWorldhumannetworkstructure{amessagewillbereceivedbyanintendedreceivershortlyonceithasreachedsomeoneinthereceiver's\clique".Considerthepseudo-codeinAlgorithm 1 .(1)Thesenderitselfstartsastherstmessageholderinthenetwork.(2)Eachmessageholdertriestostrategicallyaddadditionalmessageholdersinthenetwork.Whenitencounterswithothernodes,itasksforthebehavioralproleoftheothernodetobeconsideredasapotentialadditionalmessageholder.Eachmessageholderkeepsalistofthebehavioralprolesofallknownmessageholders 207

PAGE 208

IllustrationoftheCSI:Dscheme.Theideaistoselectthemessageholdersinanon-overlappingfashiontocovertheentirebehavioralspace. knownholderslist,togainabetterviewofthesituationofmessagespreading.(5)Iftwosimilarholders(i.e.,whentheirsimilaritymetricisabovethethresholdthnbr)encounter,oneofthemshouldceasetobeaholdertoreduceduplicatedeorts.Eachmessageholderisresponsiblefordisseminatingtheactualmessagetotheintendedreceivers.ThemessageholderssendstheTPspeciedbythesenderinthemessagetotheencounterednodes.Iftheencounterednodeisanintendedreceiver,thefullmessagewillbetransferred. 71 ]andvariantsofrandomwalk 61 74 76 107 ])inDTNasmostofthemhaveadierentroutingobjective:reachingaparticularnetworkID. 208

PAGE 209

/* /* in group(A):IfAknowsthereisamessageholderinitsneighborhood*/ifnodeAisamessageholderthen foreachnodeEencountereddo ifSim(BP(E);BP(Hi(A)))thnbrthen in group(A)=truethen foreachnodeEencountereddo in group(E)=true; intotwohalves,usethersthalftoobtainthebehavioralprolesforallusers,andthenusethesecondhalfofthetracetoevaluatethesuccessofourproposedschemes. 7.6.1.1SimulationsetupInthescenarioofCSI:Tmode,thesenderspeciestheTPandathresholdofsimilaritythsim.IfanodeshowsasimilaritymetrichigherthanthsimtotheTP,itisan 209

PAGE 210

71 ]isamessagedisseminationschemewithsimplisticdecisionrules:allnodesinthenetworksendcopiesofmessagestoalltheencounterednodeswhohavenotreceivedthemessageyet.Therandomwalk(RW)protocolgeneratesseveralcopiesofthemessagefromthesender,andeachcopyistransferredamongthenodesinarandomfashion,untilthehopcountreachesapre-setTTLvalue.Groupspreadonlyisasimpliedversionofourprotocol.Itusesonlythegroupspreadphase,i.e.,theoriginalsenderholdsontothemessageuntilitencounterswithsomeonewhoismoresimilarthanthsimtotheTPandstartsthegroupspreadphasedirectlyfromthere.Wealsoconsiderthreeprotocolsthatrequireglobalknowledgeofthefuture.Thedelay-optimalprotocolsendscopiesofthemessageonlytothenodeswhichleadtothefastestdeliverytothetargetedreceivers,andnooneelse.Thisistheoracle-basedoptimalprotocolachievableifonehasperfectknowledgeofthefuture,andservesastheupperboundforperformance.Theoverhead-optimalprotocol,ontheotherhand,optimizes(i.e.,minimizes)thenumberoftransmissioncountsusingtheknowledgeoffutureencounterevents.Thisprotocoldeliversmessagestoallreachablereceiversundertheminimumpossibletransmissioncount.Thepseudo-codeweuseforthesetwooptimalprotocolsbasedoncompleteknowledgeofallencountereventsissummarizedinAlgorithm 3 .Notice 210

PAGE 211

7.7.1 211

PAGE 212

/* /* /* time[i]:thetimenodeireceivesthemessage*/ /* /* time[i]=inf:; setdone[s]=true;setmetric[s]=0;setreach time[i]=sendtime;setcandidate=s;whilecandidate6=nulldo foreachnodekthatdone[k]=falsedo foreachEncountereventafterreach time[candidate]betweencandidateandkdo ifMessagedeliveryfromcandidatetokimproves(reduces)metric[k]then time[k]=Encounter event time;setfrom[k]=candidate; 212

PAGE 213

7-9 .Weobservethatepidemicroutingleadstothehighestoverheadwhileitsaggressivenessalsoresultsinthehighestpossibledeliveryratioandthelowestpossibledelay.TherandomwalksdonotworkwellregardlessthenumberofcopiesandthevalueofTTL,astheyusenoinformationtoguidethepropagationofthemessagetowardstherightdirection.OurCSI:Tprotocolleadstoasuccessrateclosetotheepidemicrouting(0:96forUSC,0:94forDartmouth)withverysmalloverhead(0:02forUSC,0:018forDartmouth).Forthesimpliedversion,groupspreadonly,thedelayislongerandthesuccessrateislowerthanourprotocol.Wewillfurtherinvestigatethisphenomenonlater.WhencomparingCSI:Twiththeoptimalprotocolswithfutureknowledge,weseethatthereisreallynotmuchroomforimprovementintermsofthesuccessrateandtheoverhead.OurgradientascendapproachinCSI:Tissimilartowhatisachievableevenonehastheknowledgeofthefutureinthesetwoaspects.Specically,CSI:Thasmorethan94%ofdeliveryrateanduseslessthan84%overheadofthedelay-optimalstrategy.Whencomparingwiththeoverhead-optimalprotocol,weobservethattheoverheadCSI:Tincursisaboutthesame(withlessthan5%dierence)totheoverhead-optimalprotocol,andthedelayislessintheUSCcase(by20%)butslightlymoreintheDartmouthcase(by11%).WecanthereforeconcludethatourCSI:Tprotocoldoeswellintermsofoverheadanddeliveryrate,evencomparedtotheoptimalprotocolswithperfectinformationoftheintendedreceiversandfutureencounterevents.Thedelay,ontheotherhand,hassomeroomforimprovement.Thekeyreasonofthisdierence(intermsofdelay)isthatourgradientascendphasegeneratesonlyonecopyofmessagefromthesenderanditmovestowardstheTPfollowingstrictlyascendingsimilarity.Comparingwiththebest(fastest)pathtotheTPusedintheoptimalsingle-forwarding-path,ourCSI:Thas1:40and1:47 213

PAGE 214

(b)Figure7-9. PerformancecomparisonofCSI:Ttootherprotocols.(a)USC.(b)Dartmouth. timesmoredelay,forUSCandDartmouth,respectively.Ifwecomparewiththedelay-optimalstrategy,wheremultiplecopiesaregeneratedwheneverithelpstoimprovethedelay,thedierenceisevenlarger.Thiscallsforafurtherinvestigationofselectinggoodpath(s)fromthesendertotheTP,whichweleaveoutforfuturework.Wetakeacloserlookattheperformancemetricsbysplittingthesimulationcasesintocategories,dependingontheoriginalsimilaritymetricbetweenthesender'sbehavioralproleandtheTP,Sim(BP(S);TP).BythesplitstatisticsshowninFig. 7-10 ,weseewhythegradientascendphaseisneededtoimprovethesuccessrateandreducethedelay.Whenweuseonlythegroupspreadphase,andthesenderisdissimilarfromtheTP,ittakesalongertimebeforeanyencountereventhappensdirectlybetweenthesenderand 214

PAGE 215

(b)Figure7-10. Splitperformancemetricsbythesimilaritybetweenthesenderandthetargetprole(USC).(a)Deliveryratio.(b)Averagedelay. anyoneintheneighborhoodoftheTP,ifithappensatall{hencethedelayislonger,andthesuccessrateislower.Comparingthedierencesbetweentwoversionsofrandomwalks,fewlongthreadsandmanyshortthreads,revealsaninterestingdierence.TheconceptthatleadstothedierenceisillustratedinFig. 7-11 .ManyshortthreadsarebetterifthesenderisclosetotheTP,intermsofbothdeliveryratioanddelay,asthesendergeneratesalotofthreadsto\occupy"theneighborhood{sincethethreadsareshort,andsimilarusersencountermorefrequently,theyarelikelytostayintheneighborhood.Contrarily,ifthesenderisfarawayfromtheTP,longrandomwalkthreadsprovidealegitimatechanceofmovingclosetotheTP,whileshortthreadsprovidelesshope. 215

PAGE 216

Illustrationsforthecomparisonbetweenonelongrandomwalkandmanyshortrandomwalks. 7.6.2.1SimulationsetupInthescenarioofCSI:Dmode,thetargetprolespeciedbythesendercannothelptodeterminetowherethemessageshouldbesentinthebehavioralspace.Hence,thestrategyseekstokeeponecopyineveryneighborhoodinthebehavioralspace.Inourevaluation,westartfrom1000randomlyselectedusersasthesenders.Sincethetargetproleoftheintendedreceiverscanbeorthogonaltothebehavioralprole,wecreatethescenarioforevaluationbyrandomlyselecting500nodesastheintendedreceiversforeachsender,andconsidertheaverageperformances.Wevarythetwothresholds,thfwdandthnbrinourCSI:Dmodeschemeproposedin 7.5.4 ,toadjusttheaggressivenessoftheforwardingscheme.Settinglowvaluesforboththresholdsleadstolessaggressiveoperationsandinferiorperformances.Atthesametimeisalsoleadstoloweroverheads,asthemessagesarecopiedtofewermessageholders,andtheexistenceofamessageholderpreventsnodesinalargerneighborhoodfrombecominganothermessageholder.WecomparevariousparametersettingsofourCSI:Dmodewithtwobaselineprotocols,theepidemicroutingandtherandomwalk.Theepidemicroutingworksthesamewayasbefore,servingasthebaselineforcomparison.Intherandomwalks,thevisitednodesalongthewalksbecomemessageholdersandtheywilllaterdisseminatethemessagesfurtherwhenencounteringwiththeintendedreceivers.Thedelay-optimalprotocolagainassumesglobalviewofthenetworkandtheknowledgeofthefuture.Everynodeinthenetworkknowswhotheintendedreceiversare,andsendsthemessagesto 216

PAGE 217

7-12 weshowtheaverageresultofthe1000simulationcaseswiththe95%condenceinterval.WeusethelegendCSI:D-thfwd-thnbrforourCSI:Dscheme.Comparingwiththeepidemicrouting,ourprotocolsavesalotoftransmissionandstorageoverhead.Itispossibletouseonlyabout7:2%strategicallychosennodesasthemessageholderandreachtheintendedreceiverswithlittleextradelay(about32%more),when 217

PAGE 218

(b)Figure7-12. PerformancecomparisonofCSI:Dtootherprotocols.(a)USC.(b)Dartmouth. 218

PAGE 219

219

PAGE 220

73 ].Thisisasmallconstantoverheadwepayforeachencounterwhenoneofthenodeshassomemessagetosend.Ifthemessagesizeismuchlargerthantheoverhead,whichisusuallythecaseasmessagesaretransferredinabiggerunit(i.e.,a\bundle")inDTNs,itisworthwhiletopaythisoverheadtogainthereductionoftransmissioncountsasweseeinsection 7.6 .Furthermore,withCSI,ifthereisnomessagetosend,thereisnoneedtoexchangethebehavioralprole.Thus,comparingwiththeprotocolsthatrequireproactive,persistentexchangesofcontrolmessageswhennodesencounter(e.g.,ProPHET[ 74 ]requirestheexchangeofencounterprobabilityvectors),qualitatively,theCSIschemeshaveloweroverhead,especiallywhenthevolumeoftracislowinthenetwork. 220

PAGE 221

7.3 .Inaddition,ifthebeaconsignalsfromlocationsarenotavailable,itispossibletousethemutualencountervectorsasthebehavioraldescriptorsforthenodes{nodeswhomovesimilarlyshouldhavesimilarencountersets.Inthissense,wecouldreplacetherepresentationtobetotallyindependentoftheinfrastructure. 7.5 inherentlypossessesaprivacy-preservingfeature:weonlyuseasmallsubsetofuserbehavior 221

PAGE 222

222

PAGE 223

6 7.5.1 withrespecttoeciency,exibilityandprivacypreservingproperties.TheCSIschemesperformcloselytothedelay-optimalprotocols(with94%ormoresuccessrate,lessthan83%ofoverhead,andthedelayisinferiorby40%orless).Inaddition,wealsoobservethathumanbehaviorasobservedinthelarge-scaleempiricaltracesisquiterobustandonlyafewdays'worthofdataisadequatetosummarizeandleverageformessagedissemination,whichisquitesurprising.WeareworkingtowardanimplementationoftheCSIschemesbasedonmobiledevicesandconsiderareal-worldevaluation.Onekeyissueistoadaptouralgorithminamoreprivacy-preservingfashionwhichisalsoresistanttospam(e.g.,includeareputationsystem).Wearealsoconsideringdierentapplicationsofbehavioralproles,includingtargetedadvertisingviaourCSIschemes. 223

PAGE 224

224

PAGE 225

225

PAGE 226

226

PAGE 227

Inthisappendixwebrieydiscusstheeortwemadetocollectmobilityinformationfromstudentsonauniversitycampus.Thesurveywehandedoutisshowninsection A.1 .Webuildamobilitymodel,calledtheWeightedWaypointmodel[ 21 ],baseonthesurveyresults.DescriptionsabouttheWWPmodelcanbefoundinsection A.2 .Notethatthisworkisaprimitive,bare-boneversionoftheTVCmodelwepresentinchapter 4 {manyimportantideascanalsobefoundhere.Usingthismodel,weshowthattherecanbehighuserconcentrationatpopularspotsoncampusduetothehighprobabilitiesofvisitingthepopularlocationsandthehighstaydurationsattheselocations,leadingtopotentialcongestionsattheAPsintheselocations.Weproposeacongestionalleviationprotocolbyroutingsometractonear-byAPswithlowloadthroughmulti-hoproutingacrossthemobileusers. A-1 (b)FigureA-1. Thesurveyform.(a).Mobilitypattern.(b).Networkusage. 227

PAGE 228

A-1 (b),theprobabilityofusingwirelessnetworks)cannotbedirectlydeducedfromthenetworktraces.

PAGE 229

Thevirtualcampus. A-2 .Wereferthistopologyasthe\virtualcampus"henceforth.Inthisscenarioweidentify7noncontiguouslocations:3classrooms(CL),2libraries(L),and2cafeterias(Ca).InordertondadequateparametersforourWWPmodelexamplefortheUSCcampus,weconductedamobilitysurveytargetedatrandomlyselectedstudentsoncampus.DuringtheperiodbetweenMarch22nd2004andApril16th2004,wecollected268surveyresponsesontheUSCcampus.Thedetailedquestionsweaskinthesurveyscanbefoundinsection A.1 .Thelocationgranularityofourmobilitysurveyisper-building.Ineachsurvey,thestudentisaskedtollinhis/hercurrentlocation(building),thepreviousbuildingvisited,thenextbuildingtovisit,andthepausedurationateachofthese3buildings.TosetuptheWWPmodelforacampusenvironment,wecategorizebuildingsoncampusintothreedierentlocationtypes:classrooms,li-braries,andcafeterias.Thebuildingsandareathatdoesnotbelongtothese3categoriesarecollectivelyreferredtoasotherarea.Wealsomodelthemobilenodesmovingtoo-campusareawithcertainprobabilities.MNchoosesitsnextdestinationfromone 229

PAGE 230

Markovmodeloflocationtransitionofmobilenodes. ofthese5locationtypesaccordingtoaMarkovmodel,asshowninFig. A-3 .Wesetupthetransitionprobabilitiestodierentlocationtypesaccordingtoits\weights"orpopularity.Fromthesurveywecapturestatisticsaboutthefollowingparameters:(a)Thepausetimedistributionsatclassrooms,libraries,cafeterias,andotherarea.(b)Thetime-varyingtransitionprobabilitygiventhecurrentlocationtypeandtimesection(morning:9AM-1PMorafternoon:1PM-5PM)oftheday.(c)Inadditiontothesemobility-relatedparameters,wealsosurveyforthewirelessnetworkusage{theprobabilityanddurationarespondentuseswirelessnetworksatdierenttypesoflocations.Wediscussthemainndingsofourmobilitysurveybelow.PauseTimeDuration ThepausetimedurationisasshowninFig. A-4 .(a)Thedistributionofpausetimeatclassroomislikeabell-shapednormaldistributionwiththepeakaroundthe60-120minutesinterval,whichistheregularclassduration(about90minutes)atUSC.(b)Alsowecanseethatpeoplearemorelikelytostayinthelibraryforintervalsgreaterthan240minutesthaninanyotherlocations.Forotherareaoncampus,thedurationtendstobeexponentiallydistributed.TransitionProbability The\transitionprobabilitymatrix"fromthesurveydataisshowninTable A-1 .(a)Peopletendtogotoacafeteriamoreinthemorninginterval(lunchtime)thanintheafternoon.Insteadofvisitingtheothercategory,mosttransitions(morethan50%)arebetweenclassroomsandlibraries.(b)Alsomosttransitionsinvolvingo-campus 230

PAGE 231

Pausetimedistributionforlocations. ClassroomLibraryCafeOthersOCurrentlocation/time Campus Classroom9-13 0.150.440.000.220.1913-17 0.200.500.000.300.00 Others9-13 0.090.120.250.300.2413-17 0.200.430.090.140.14 O9-13 Transitionprobabilitymatrix. locationareofthetype\ocampus-class-ocampus"or\ocampus-library-ocampus"whichwebelievereectsthegeneralstudentbehavior.Thisimpliesthefactthato-campusstudentscometocampusmostlytoattendclassesortouselibraries.WealsotrytoobtainthetransitionprobabilitymatrixfromtheUSCwirelessnetworktraces[ 80 ],withbuilding-levelgranularity.Therearethreeinitialndingsonthis:(a)Startingfromagivenbuilding,thetransitionprobabilitiestowardtheothersarenotequallydistributed.Thissupportsourassumptionthatsomelocationsaremorepopularthanothersinacampusenvironment.(b)Fromthetraceweobservesimilartrendstothesurvey{Cafeteriasaremorepopularinthemorninginterval,andtherearealottransitionsbetweenlibrariesandclassrooms.(c)Fromagivenbuilding,thetransitionprobabilitiestowardclose-bybuildingsarehigherthanbuildingsthatarefaraway.Thismaysuggestthatpedestrianmobilityoncampusexhibitslocality.WirelessNetworkowduration ThehistogramofowdurationdistributionsatdierenttypesoflocationsisshowninFig. A-5 .Theowdurationdistributionshowsaheaviertailforthelibrary,probablyduetopeopleworkinginthelibrarieswiththeirlaptopconnectedtothewirelessnetwork.Wefurthercomparethendingsofthispartwith 231

PAGE 232

Flowdurationdistributionforlocations. thedistributionsofuseronlinetimeintheDartmouthWLANtraces[ 13 ].FromtheDartmouthtrace[ 81 ]wendthatformostbuildingstheonlinetimedistributionishighlyskewedtowardshortdurations,regardlessofthebuildingtype.Theobservationbasedonoursurveysandtracesaresimilarexceptforthelibraries. A.2.3.1PropertiesofWWPmodelWeusesimulationstoshowthecharacteristicsoftheWWPmodel,incomparisonwiththeRWPmodel.First,WWPmodelshowsunevenspatialdistributionofMNs.TheMNstendtoclusterwithinthepopularlocations,asshowninFig. A-6 .Howeverthenodedensityisquitelowforotherareaandocampuslocations.Thisisacombinedeectofpopularlocationsbeingchosenasdestinationswithhigherprobabilitiesandpausetimesatthoselocationsbeinglongwithhigherprobability.Second,althoughforagivenxedtransitionprobabilitymatrixthereshouldbesometheoreticalsteadystateoftheMNdistribution,thetransitionprobabilitymatrixistime-dependentandchangesfromtimetotimethroughouttheday,hencetheMNdistributioninthesimulationareaneverreachesasteadystateinFig. A-6 .Thissuggestsconvergingtoasteady-statedistributionisnotnecessarilyarequirementofrealisticmobilitymodels.Third,weusethemove-stopratio(thetotalmovetimedividedbythetotalstationarytime)asonemetricofamobilitymodelandndthattheWWPmodelbasedonourmobilitysurveydatahasalowermove-stopratio0.12ascompareto0.99fromtheRWPmodelwithcommonparameter 232

PAGE 233

Mobilenodedensityversustime. WWPwithbothtransitionmatrix0.12RWPwithpausetime=[0,100](s),0.99speed=[2,50](m/s)|typicalparametersetting Move-stopratio. settings,asshowninTable A-2 .Thisindicates,forauniversitycampusscenario,peoplearelessmobilethantypicalscenariosgeneratedbytheRWPmodel. 233

PAGE 234

103 ]astheroutingprotocolundertwodierentMNlocationrelationships,MNpairsinthesamelocationandMNpairsindierentlocations.IftheWWPmodelisused,weshowthattheroutediscoverysuccessratesare88.61%and28.53%forMNsinthesamelocationandindierentlocations,respectively.ThereasonforthelowroutediscoverysuccessrateforMNsindierentlocationsisthatthenumberofnodespresentbetweentheselocationsisverysmallduetothepreferenceofchoosingpopularlocationsasdestinations.HencefewnodesareabletoserveastheintermediatenodestoestablisharoutebetweenMNsindierentlocations.Thereforeitislikelythatthenetworkwillbepartitionedintosmallsubsetsclusteredatthepopularlocations,anditisdiculttondaroutebetweenthesesubsets. A-2 )asafunctionoftimeinFig. A-7 whenthereare200MNsinthesimulation.WhiletheAPatlibrary1hasalargenumberofows,APsatclassroom2andcafeteria2arequiteunderutilized.Thisunevendistributionofowssuggeststhepossibilityofusingadhocnetworktechniquestore-routesomeowstotheunderutilizedneighboringAPsinordertoalleviatelocalcongestion.ItisfeasibletoimprovetheQoSoftheowsatthecongestedAP,ifwecanndamulti-hopadhocroutetoredirectittounderutilizedneighboringAPs(NAPs).WeproposethefollowingMN-initiatedow-switchingmechanismtoachievethisgoal. 234

PAGE 235

UnevenowdistributionacrossAPs. FigureA-8. Thecontrolowchartoftheproposedow-switchingmechanism. A-8 .TheMNnotiestheLAPofitsrequesttobeswitchedtootherAPsbysendinga\re-routerequest"totheLAP.Uponreceivingthismessage,theLAPrequestshelpfromitsneighborsbysendinga\help"messagetooneofthem.ThechoiceoftheneighborisbasedontheAPsgeographicalknowledgeoftheAP-deploymenttopology.TheLAP 235

PAGE 236

103 ]astheadhocroutingprotocol.)foreachothersimultaneously.Thisisachievablebecausethewirednetworkbetweentheaccesspointsprovidesa\tunnel"toexchangeinformationbetweentheMNandtheNAPbeforetheyactuallyestablishanadhocroutetoeachother.Theintermediatenodesatwhichthebi-directionalrouterequestmessagesmeetwillconcatenatethepartialroutesfrombothendsandsendbackroutereplymessagestotheMNandtheNAP.Such\meetinthehalfway"behaviorispossiblebecauseDSRcachesthepartialroutearoute-requestpackettraversedbeforereachingthenode,thereforeanintermediatenodeisabletoestablishtheend-to-endpathifitisvisitedbyroute-requestpacketsfrombothendsoneaftertheother.Thebi-directionalsearchfortheadhocroutecanpotentiallyreducetheroutediscoverytime.InourworkweassumethatMNsuseadedicatedwirelesschanneltocommunicatewithotherMNs,sothattheadhocnetworkdoesnotinterferewithcongestedlocalwirelesschannelusedbytheLAPandotherMNs.Thiscanbeachievedbyreservingadedicatedchannelfortheadhocnetworkcommunication.AllAPsandMNsinthesystemmustagreeonusingthisreservedchannelonlyfortheadhocnetworkcommunication.ThechannelisnotusedlocallybyanyAP.IftheLAPassignsaMNtobeswitchedtooneofitsneighbor,butthereisnoavailablemulti-hoproutefromtheMNtotheNAP,theswitchingisconsideredafailureandtheMNwillreestablishitsconnectiontotheLAPafteraxedperiodoftime.IftheMNisabletoestablisharoutetothedesignatedNAP,buttheroutebreakslaterduetomovementofintermediatenodes,theMNwillalsoreestablishtheconnectiontotheLAP.Suchfall-back-to-LAPbehaviorisnecessarytoavoidaMNwaitingindenitelyfor 236

PAGE 237

88 ]networksimulatortosimulateourproposedowswitchingmechanism.WevarythetotalnumberofMNsinthesimulationareafrom100to200tocreatedierentdegreeofcongestion.ThemobilitymodelusedbytheMNsistheproposedWWPmodelintroducedinsection A.2 .Inthesimulation,weassumethateachAPoperatesat2Mb/sbitrate.EachMNowrequires200Kb/sthroughput.Tosimplifythesimulation,theMNsidentifytheLAPcongestionbycountingcurrentnumberofowsconnectedtotheLAP.ThelocalAPbecomescongestedandthethroughputforlocalowsstarttodropif7ormoresimultaneousowsareconnectedtotheLAP(Thisnumberwasobtainedviadetailedsimulations.Thewirelesschannelcannotreach100%utilizationbecauseofcontentionsinthewirelesschannel.)Wesimulateboththescenarioswithandwithouttheowswitchingmechanism.Theeectoftheow-switchingalgorithmisprimarilytore-distributetheloadoftracacrosstheAPs.IfsomeAPbecomescongested,theMNssensethecongestionby 237

PAGE 238

Flowsre-distributedacrossAPs,relievingcongestionatlibrary1. observingdegradationinthethroughputoftheon-goingowandtrytoswitchtheowtooneoftheNAPs.IfsomeoftheMNssucceedinowswitching,theexcessiveowsattheLAPwillshifttoitsneighbors,andboththeowsthatareswitchedandtheowsthatstayattheLAPcanenjoyuncongestedwirelesschannelandbetterthroughput.Theconsequenceoftheow-switchingisillustratedbycomparingFig. A-7 toFig. A-9 ,whereweillustratethenumberofowsatthesame3APslocatedintheupper-rightcornerofthevirtualcampus(Fig. A-2 ),withtheow-switchingmechanism.Weseethatsomeowsatlibrary1areswitchedtoclassroom2andcafeteria2,sothecongestionatlibrary1isnotasbadasinthecasewithoutow-switchingshowninFig. A-7 .Tobetterunderstandtheeectoftheow-switchingmechanismontheoverallimprovementsofthesystem,weproposetousethemetrics\APcongestedtimeratio"and\owqualitytimeratio".TheformerisdenedasthetimeratioanAPhasatleast7owsconnectedtoit.ThisisthetimeratiothattheAPcannotprovideadequateQoStotheconnectedows.ThelatterisdenedasthetimeratioofaowconnectedtoanyAPwithlessthan7owsconnectedsimultaneously.Thisistheproportionoftimetheowreceivesadequatethroughput.NotethatbetweenthetimeaMNdecidestoswitchaowtoNAPuntilthetimeitndsaroutetothedesignatedNAP,theowisnotconnectedtoanyAPhencethistimeperiodwillnotbeincludedthequalitytime 238

PAGE 239

AllAPs:AverageAPcongestedtimeratio. ratio.Theresultsshownbelowaretheaveragesof6independentsimulationruns,usingadierent,randomlygeneratedmobilityscenarioforeachrun.Fig. A-10 showstheaverageofAPcongestedtimeratioofallAPs.Fig. A-11 showstheaverageofAPcongestedtimeratioofthemostcongestedAPineachsimulationrun.WecanseethatduetotheunevenMNdistributionresultingfromtheWWPmodel,theoverallcongestedtimeratioislowforthewholesystem.However,themostcongestedAPisquiteoverloaded.ThisisexactlythesituationwhenswitchingsomeowstotheNAPsshouldbehelpful.FromtheguresweseethatthecongestedtimeratioofthemostcongestedAPisreducedbymorethan50%inallexceptforthe100MNcase.Thisimpliesow-switchinghelpstoreducethelocalcongestionofwirelessLANsmorethanhalfofthetimewhencongestionexists.Theowqualitytimeratioisthemetrictoobservefortheimprovementwegetbyemployingow-switchingfromuser'sperspective.InFig. A-12 weshowtheow-switchingmechanismimprovesthequalitytimeratioforallcases.WeobserveinthecaseofsmallerMNnumbers(100or125MNs)theeectofow-switchingisnotsopronounced.Thisisbecausewhenthenetworkissparse,thereislesschancetondaroutetothedesignatedNAPfortheswitchingows.Hencetheeectivenessofow-switchingislimited.ThesuccessrateforaswitchingowtondaroutetothechosenNAPisabout0:27whenthereare100MNs,andthesuccessrateincreasesto0:43whenthereare200MNs. 239

PAGE 240

ThemostcongestedAP:AverageAPcongestedtimeratio. FigureA-12. Averagequalitytimeratioofallows. 240

PAGE 241

[1] MobiLib:Community-wideLibraryofMobilityandWirelessNetworksMeasurements.http://nile.usc.edu/MobiLib. [2] CRAWDAD:ACommunityResourceforArchivingWirelessDataAtDartmouth.http://crawdad.cs.dartmouth.edu/index.php. [3] C.Perkins,"AdHocNetworking,"Addison-Wesley,publishedDec.2000. [4] Delaytolerantnetworkingresearchgroup.http://www.dtnrg.org. [5] R.Aldunate,S.Ochoa,F.Pena-Mora,andM.Nuaabaum,"RobustMobileAdHocSpaceforCollaborationtoSupportDisasterReliefEortsInvolvingCriticalPhysicalInfrastructure,"InJournalofComp.inCivilEngineering,vol.20,issue1,pp.13-27,2006. [6] S.Jain,K.Fall,andR.Patra,"Routinginadelaytolerantnetwork,"InProceedingsofACMSIGCOMM,Aug.2004. [7] F.Bai,T.Elbatt,G.Hollan,H.Krishnan,V.Sadekar,"TowardsCharacterizingandClassifyingCommunication-basedAutomotiveApplicationsfromaWirelessNetworkingPerspective,"inProceedingsofthe1stIEEEWorkshoponAutomotiveNetworkingandApplications(AutoNet2006),Nov.2006. [8] D.J.WattsandS.H.Strogatz."CollectiveDynamicsof'Small-World'Networks,"Nature,vol.393,pp.440-442,1998. [9] Simulationcodesusedinthisworkanditsdetaileddescriptionareavailableathttp://nile.cise.u.edu/~weijenhs/TVC model [10] M.BalazinskaandP.Castro,"CharacterizingMobilityandNetworkUsageinaCorporateWirelessLocal-AreaNetwork,"InProceedingsofMobiSys2003,pp.303-316,May2003. [11] M.McNettandG.Voelker,"AccessandmobilityofwirelessPDAusers,"ACMSIGMOBILEMobileComputingandCommunicationsReview,v.7n.4,October2003. [12] D.KotzandK.Essien,"AnalysisofaCampus-wideWirelessNetwork,"InProceedingsofACMMobiCom,September,2002. [13] T.Henderson,D.KotzandI.Abyzov,"TheChangingUsageofaMatureCampus-wideWirelessNetwork,"inProceedingsofACMMobiCom2004,September2004. [14] M.Papadopouli,H.Shen,andM.Spanakis,"CharacterizingtheDurationandAssociationPatternsofWirelessAccessinaCampus,"11thEuropeanWirelessConference2005,Nicosia,Cyprus,April10-13,2005. 241

PAGE 242

W.HsuandA.Helmy,"OnImportantAspectsofModelingUserAssociationsinWirelessLANTraces,"theSecondInternationalWorkshopOnWirelessNetworkMeasurement(WiNMee2006),April2006. [16] N.EagleandA.Pentland,"Realitymining:sensingcomplexsocialsystems,"inJournalofPersonalandUbiquitousComputing,vol.10,no.4,May2006. [17] CabSpotting,aprojectthattrackstaximobilityintheSanFranciscoBayArea.Traceavailableathttp://cabspotting.org/api. [18] W.Leland,M.Taqqu,W.Willinger,andD.Wilson,"OntheSelf-similarNatureofEthernetTrac(extendedversion),"inIEEE/ACMTransactionsonNetworking,Vol.2,Issue1,pp.1{15,Feb.1994. [19] C.Faloutsos,P.Faloutsos,andM.Faloutsos,"OnPower-LawRelationshipsoftheInternetTopology,"InProceedingsoftheACMSIGCOMMConferenceonNetworkArchitecturesandProtocols,Sep.1999. [20] F.Bai,N.Sadagopan,andA.Helmy,"TheIMPORTANTFrameworkforAnalyzingtheImpactofMobilityonPerformanceofRoutingforAdHocNetworks",AdHocNetworksJournal-Elsevier,Vol.1,Issue4,pp.383{403,Nov.2003. [21] W.Hsu,K.Merchant,H.Shu,C.Hsu,andA.Helmy,"WeightedWaypointMobilityModelanditsImpactonAdHocNetworks-Mobicom2004PosterAbstract,"MobileComputingandCommunicationReview,Jan2005. [22] D.Batacharjee,A.Rao,C.Shah,M.Shah,andA.Helmy"EmpiricalModelingofCampus-widePedestrianMobility:ObservationontheUSCCampus",inProceedingsofIEEEVehicularTechnologyConference(VTC),Sep.2004. [23] D.TangandM.Baker,"AnalysisofaLocal-areaWirelessNetwork,"InProceedingsofthe6thannualinternationalconferenceonMobilecomputingandnetworking(MobiCom2000),Aug.2000. [24] D.TangandM.Baker,"AnalysisofaMetropolitan-AreaWirelessNetwork,"WirelessNetworks,vol.8,no.2-3,pp.107{120,Nov.2004 [25] H.ZangandJ.Bolot,"MiningCallDatatoIncreasetheRobustnessofCellularNetworkstoSignalingDoSAttacks,"inProceedingsofACMMobiCom2007,September2007. [26] TheHaggleProject.http://www.cl.cam.ac.uk/research/srg/netos/haggle/index.html [27] A.Chaintreau,P.Hui,J.Crowcroft,C.Diot,R.Gass,andJ.Scott,"ImpactofHumanMobilityontheDesignofOpportunisticForwardingAlgorithms,"inProceedingsofINFOCOM2006,Apr.2006. 242

PAGE 243

J.Leguay,A.Lindgren,J.Scott,T.Friedman,andJ.Crowcroft,"OpportunisticContentDistributioninanUrbanSetting,"inProceedingsofWorkshoponChallengedNetworks(CHANTS-06),co-locatedwithACMSIGCOMM2006,Sep.2006. [29] J.Su,A.Chin,A.Popivanova,A.Goel,andE.deLara,Usermobilityforopportunisticad-hocnetworking,inProceedingsofthe6thIEEEWorkshoponMobileComputingSystemsandApplications(WMCSA04),2004. [30] J.Burgess,B.Gallagher,D.Jensen,andB.Levine,"MaxProp:RoutingforVehicle-BasedDisruption-TolerantNetworks,"InProceedingsofIEEEINFOCOM,Apr.2006. [31] N.Banerjee,M.Corner,andB.Levine,"AnEnergy-EcientArchitectureforDTNThrowboxes,"inProceedingsofINFOCOM2007,May2007. [32] P.Juang,H.Oki,Y.Wang,M.Martonosi,L-S.Peh,andD.Rubenstein,"Energy-EcientComputingforWildlifeTracking:DesignTradeosandEarlyExperiencewithZebraNet,"InASPLOS-Xconference,Oct.2002. [33] T.SmallandZ.Haas,"TheSharedWirelessInfostationModel:aNewAdHocNetworkingParadigm(orwherethereisawhale,thereisaway),"InProceedingsofthe4thACMMOBIHOC,June2003. [34] M.KimandD.Kotz,"ModelingUsers'MobilityamongWiFiAccessPoints,"inProceedingsoftheInternationalWorkshoponWirelessTracMeasurementsandModeling(WiTMeMo'05),Seattle,Washington,June2005. [35] M.Papadopouli,H.Shen,M.Spanakis,"ModelingClientArrivalsatAccessPointsinWirelessCampus-wideNetworks,"14thIEEEWorkshoponLocalandMetropolitanAreaNetworks,Chania,Crete,Greece,September18-21,2005. [36] G.Chen,H.Huang,andM.Kim,"MiningFrequentandPeriodicAssociationPatterns,"DartmouthCollegeComputerScienceTechnicalReportTR2005-550,July2005. [37] X.Meng,S.Wong,Y.Yuan,andS.Lu,"CharacterizingFlowsinLargeWirelessDataNetworks,"inProceedingsofACMMobiCom,September2004. [38] M.KimandD.Kotz,"Periodicpropertiesofusermobilityandaccess-pointpopularity,"JournalofPersonalandUbiquitousComputing,11(6),August,2007. [39] J.Kang,W.Welbourne,B.Stewart,andG.Borriello,"Extractingplacesfromtracesoflocations,"inSIGMOBILEMobileComputingandCommunicationReview,vol.9,no.3,July2005. [40] I.T.Jollie,PrincipalComponentAnalysis,seconded.,Springerseriesinstatistics,published2002. 243

PAGE 244

R.HornandC.Johnson,MatrixAnalysis,CambridgeUniversityPress,published1990. [42] A.Lakhina,K.Papagiannaki,M.Crovella,C.Diot,E.D.Kolaczyk,andN.Taft,"StructuralAnalysisofNetworkTracFlows,"ACMSIGMETRICS,NewYork,June2004. [43] C.J.F.terBraak,"PrincipalComponentsBiplotsandAlphaandBetaDiversity,"Ecology,vol.64,pp.454-462,1983. [44] R.AlbertandA.Barabasi,"Statisticalmechanicsofcomplexnetworks,"Reviewofmodernphysics,vol.74,no.1,pp.47-97,Jan.2002. [45] A.Helmy,"SmallWorldsinWirelessNetworks,"IEEECommunicationsLetters,pp.490-492,Vol.7,No.10,October2003. [46] F.BaiandA.Helmy,"ImpactofMobilityonLastEncounterRoutingProtocols,"inProceedingsoftheFourthAnnualIEEECommunicationsSocietyConferenceonSensor,MeshandAdHocCommunicationsandNetworks(SECON2007),June2007. [47] V.Srinivasan,M.Motani,andW.T.Ooi,"AnalysisandImplicationsofStudentContactPatternsDerivedfromCampusSchedules,"inProceedingsofMOBICOM2006,Sep.2006. [48] M.GrossglauserandD.Tse,"Mobilityincreasesthecapacityofadhocwirelessnetworks,"IEEE/ACMTransactionsonNetworking,10(4),2002. [49] T.Camp,J.Boleng,andV.Davies,"ASurveyofMobilityModelsforAdHocNetworkResearch,"WirelessCommunication&MobileComputing(WCMC):SpecialissueonMobileAdHocNetworking:Research,TrendsandApplications,vol.2,no.5,pp.483-502,2002. [50] F.Bai,A.Helmy,"ASurveyofMobilityModelingandAnalysisinWirelessAdhocNetworks",BookChapterin"WirelessAdHocandSensorNetworks",Springer,October2006,ISBN:978-0-387-25483-8. [51] C.Bettstetter,G.Resta,P.Santi,"TheNodeDistributionoftherandomwaypointmobilitymodelforwirelessadhocnetworks,"inIEEETrans.onMobileComputing,vol.2,issue3,pp.257-269,Jul.2003. [52] T.SpyropoulosandK.Psounis,"PerformanceAnalysisofMobility-assistedRouting,"InProceedingsofACMMOBIHOC,May2006. [53] A.JindalandK.Psounis,"FundamentalMobilityPropertiesforRealisticPerformanceAnalysisofIntermittentlyConnectedMobileNetworks,"InProceedingsofIEEEPerComWorkshoponIntermittentlyConnectedMobileAdHocNetworks(ICMAN),2007. 244

PAGE 245

S.Jain,K.Fall,andR.Patra,"Routinginadelaytolerantnetwork,"InProceedingsofACMSIGCOMM,Aug.2004. [55] R.C.Shah,S.Roy,S.Jain,andW.Brunette,"Datamules:Modelingandanalysisofathree-tierarchitectureforsparsesensornetworks,"ElsevierAdHocNetworksJournal,2003. [56] T.Spyropoulos,K.Psounis,andC.S.Raghavendra,"Single-copyroutinginintermittentlyconnectedmobilenetworks,"InProceedingsofIEEESECON,Oct.2004. [57] T.Spyropoulos,K.Psounis,andC.S.Raghavendra,"Sprayandwait:Ecientroutinginintermittentlyconnectedmobilenetworks,"InProceedingsofACMSIGCOMMworkshoponDelayTolerantNetworking(WDTN),Aug.2005. [58] X.Hong,M.Gerla,G.Pei,C.andChiang,"Agroupmobilitymodelforadhocwirelessnetworks,"inProceedingsofthe2ndACMinternationalworkshoponmodeling,analysisandsimulationofwirelessandmobilesystems,August,1999. [59] A.Jardosh,E.Belfding-Royer,K.Almeroth,S.Suri."TowardsRealisticMobilityModelsforMobileAdhocNetworks."inProceedingsofACMMobiCom,Sep.2003. [60] W.HsuandA.Helmy,"OnNodalEncounterPatternsinWirelessLANTraces,"theSecondInternationalWorkshopOnWirelessNetworkMeasurement(WiNMee2006),April2006. [61] J.Leguay,T.Friedman,andV.Conan,"EvaluatingMobilityPatternSpaceRoutingforDTNs,"inProceedingsofIEEEINFOCOM,April,2006. [62] C.TuduceandT.Gross,"AMobilityModelBasedonWLANTracesanditsValidation,"InProceedingsofIEEEINFOCOM,Mar.2005. [63] R.Jain,D.Lelescu,M.Balakrishnan,"ModelT:AnEmpiricalModelforUserRegistrationPatternsinaCampusWirelessLAN,"InProceedingsofACMMOBICOM,Aug.2005. [64] D.Lelescu,U.C.Kozat,R.Jain,andM.Balakrishnan,"ModelT++:AnEmpiricalJointSpace-TimeRegistrationModel,"InProceedingsofACMMOBIHOC,May2006. [65] M.Kim,D.Kotz,andS.Kim,"Extractingamobilitymodelfromrealusertraces,"InProceedingsofIEEEINFOCOM,Apr.2006. [66] W.HsuandA.Helmy,"OnImportantAspectsofModelingUserAssociationsinWirelessLANTraces,"InProceedingsoftheSecondInternationalWorkshopOnWirelessNetworkMeasurement,Apr.2006. 245

PAGE 246

W.Hsu,T.Spyropoulos,K.Psounis,A.Helmy,"ModelingTime-variantUserMobilityinWirelessMobileNetworks,"InProceedingsofIEEEINFOCOM,May2007. [68] M.MusolesiandC.Mascolo,"ACommunityBasedMobilityModelforAdHocNetworkResearch,"InProceedingsoftheSecondInternationalWorkshoponMulti-hopAdHocNetworks(REALMAN),May2006. [69] S.Merugu,M.Ammar,andE.Zegura,"RoutinginSpaceandTimeinNetworkswithPredictableMobility,"GeorgiaInstituteofTechnologyCCTechnicalReport,GIT-CC-04-07,2004. [70] W.Zhao,M.Ammar,andE.Zegura,"MulticastinginDelayTolerantNetworks:SemanticModelsandRoutingAlgorithms,"InProceedingofthe2005ACMSIGCOMMworkshoponDelay-tolerantnetworking,Aug.2005. [71] A.VahdatandD.Becker,"EpidemicRoutingforPartiallyConnectedAdHocNetworks,"TechnicalReportCS-200006,DukeUniversity,April2000. [72] Y.KoandN.Vaidya,"Flooding-BasedGeocastingProtocolsforMobileAdHocNetworks,"MobileNetworksandApplications,vol.7,Issue6,pp.471-480,Dec.2002. [73] W.Hsu,D.Dutta,andA.Helmy,"ExtendedAbstract:MiningBehavioralGroupsinLargeWirelessLANs"InProceedingsofACMMOBICOM,Sep.2007.Longertechnicalreportavailableathttp://arxiv.org/abs/cs/0606002 [74] A.Lindgren,A.Doria,andO.Scheln,"Probabilisticroutinginintermittentlyconnectednetworks,"LectureNotesinComputerScience,vol.3126,pp.239-V254,Sep.2004. [75] A.Helmy,S.Garg,P.Pamu,N.Nahata,"ContactBasedArchitectureforResourceDiscovery(CARD)inLargeScaleMANets",ThirdIEEE/ACMInternationalWorkshoponWireless,MobileandAdHocNetworks(WMAN),Apr.2003. [76] E.DalyandM.Haahr,"SocialNetworkAnalysisforRoutinginDisconnectedDelay-TolerantMANETs,"InProceedingsofACMMOBIHOC,Sep.2007. [77] W.Hsu,K.Merchant,H.Shu,C.Hsu,andA.Helmy,"Preference-basedMobilityModelingandtheCaseforCongestionReliefinWLANsusingAdhocNetworks,"IEEEVehicularTechnologyConference(VTC),LosAngelesCA,Sep.2004. [78] P.Hui,E.Yoneki,S.Chan,andJ.Crowcroft,"DistributedCommunityDetectioninDelayTolerantNetworks,"InProceedingsofSigcommWorkshopMobiArch'07,August,Kyoto,Japan. [79] A.B.McDonaldandT.F.Znati,"AMobility-basedFrameworkforAdaptiveClusteringinWirelessAdHocNetworks,"InIEEEJournalonSelectedAreasinCommunications,vol.17,issue8,pp.1466-1487,Aug.1999. 246

PAGE 247

W.HsuandA.Helmy,MobiLibUSCWLANtracedataset.Downloadedfromhttp://nile.cise.u.edu/MobiLib/USC trace/ [81] D.Kotz,T.HendersonandI.Abyzov,CRAWDADdatasetdartmouth/campus/movement/01 04(v.2005-03-08).Downloadedfromhttp://crawdad.cs.dartmouth.edu/dartmouth/campus/movement/01 04 [82] M.BalazinskaandP.Castro,CRAWDADdatasetibm/watson(v.2003-02-19).Downloadfromhttp://crawdad.cs.dartmouth.edu/ibm/watson [83] M.McNettandG.M.Voelker,WirelessTopologyDiscoveryprojectdataset.Downloadfromhttp://sysnet.ucsd.edu/wtd/ [84] J.Scott,R.Gass,J.Crowcroft,P.Hui,C.Diot,andA.Chaintreau,CRAWDADtracecambridge/haggle/imote/infocom(v.2006-01-31).Downloadfromhttp://crawdad.cs.dartmouth.edu/cambridge/haggle/imote/infocom [85] D.Kotz,T.Henderson,andI.Abyzov,CRAWDADtracesetdartmouth/campus/tcpdump(v.2004-11-09).Weusethelistofdevicetypesinthefall03tcpdumpdata.Downloadfromhttp://crawdad.cs.dartmouth.edu/dartmouth/campus/tcpdump [86] P.Krishna,N.H.Vaidya,M.Chatterjee,andD.K.Pradhan,"ACluster-basedApproachforRoutinginDynamicNetworks,"InACMSIGCOMMComputerCommunicationReview,vol.27,issue2,pp.49-64,Apr.1997. [87] K.Seada,M.Zuniga,A.Helmy,B.Krishnamachari,"Energy-EcientForwardingStrategiesforGeographicRoutinginLossyWirelessSensorNetworks,"InProceedingsofACMSensys,Nov,2004. [88] TheNetworkSimulator-NS-2.http://www.isi.edu/nsnam/ns/ [89] B.KarpandH.Kung,"GPSR:greedyperimeterstatelessroutingforwirelessnetworks,"InProceedingsofACMMobiCom,Aug.2000. [90] S.TanachaiwiwatandA.Helmy,"Encounter-basedWorms:AnalysisandDefense",IEEEConferenceonSensorandAdHocCommunicationsandNetworks(SECON)2006Poster/DemoSession,September2006. [91] X.Zhang,G.Neglia,J.Kurose,andD.Towsley,"PerformanceModelingofEpidemicRouting,"inProceedingsofIFIPNetworking2006. [92] R.Groenevelt,P.Nain,andG.Koole,"TheMessageDelayinMobileAdHocNetworks,"InProceedingsofPERFORMANCE,Oct.2005. [93] M.GrossglauserandM.Vetterli,"LocatingNodeswithEASE:MobilityDiusionofLastEncountersinAdHocNetworks,"InProceedingsofIEEEINFOCOM,April2003. 247

PAGE 248

H.Dubois-Ferriere,M.Grossglauser,andM.Vetterli."AgeMatters:EcientRouteDiscoveryinMobileAdHocNetworksusingEncounterAges,InProceedingsofACMMobiHoc,June2003. [95] A.Jain,M.Murty,andP.Flynn,"DataClustering:AReview,"ACMComputingSurveys,vol.31,no.3,September,1999. [96] L.Denoeud,H.Garreta,A.Guenoche,"Comparisonofdistanceindicesbetweenpartitions,"InternationalSymposiumonAppliedStochasticModelsandDataAnalysis,May2005. [97] J.FaruqueandA.Helmy,"RUGGED:RoUtingonnGerprintGradientsinsEnsorNetworks,"InProceedingsoftheTheIEEE/ACSInternationalConferenceonPervasiveServices(ICPS'04),Jul.2004. [98] R.HoggandE.Tanis,"ProbabilityandStatisticalInference,"sixthedition,PrenticeHall,2001. [99] A.RapoportandW.Horvath,"AStudyofaLargeSociogram,"BehavioralScience6,279-291,1961. [100] C.Gkantsidis,G.Goel,M.Mihail,andA.Saberi,"TowardsTopologyAwareNetworks,"intheProceedingsofIEEEINFOCOMmini-symposium,Anchorage,Alaska,May2007. [101] C.Nuzman,I.Saniee,W.Sweldens,andA.Weiss,"ACompoundModelforTCPConnectionArrivalsforLANandWANApplications,"ComputerNetworks,40:319V337,October2002. [102] NOMADS:NetworkofMobileAdhocDevicesandSensors,researchgroupleadbyDr.AhmedHelmy.Grouphomepagehttp://nile.cise.u.edu/ [103] D.B.JohnsonandD.A.Maltz,DynamicSourceRoutinginAd-HocWirelessNetworks,MobileComputing,pp.153-181,1996. [104] S.TanachaiwiwatandA.Helmy,"OnthePerformanceEvaluationofEncounter-basedWormInteractionsBasedonNodeCharacteristics"ACMMobicom2007WorkshoponChallengedNetworks(CHANTS2007),Montreal,Quebec,Canada,Sep.2007. [105] P.Costa,C.Mascolo,M.Musolesi,andG.Picco,"Socially-awareRoutingforPublish-SubscribeinDelay-tolerantMobileAdHocNetworks,"toappearinIEEEJournalonSelectedAreaofCommunications. [106] A.Miklas,K.Gollu,K.Chan,S.Saroiu,K.Gummadi,andE.Lara,"ExploitingSocialInteractionsinMobileSystems,"inProceedingsof9thInternationalConferenceonUbiquitousComputing,Sep.2007. 248

PAGE 249

M.Thomas,A.Gupta,andS.Keshav,"GroupBasedRoutinginDisconnectedAdHocNetworks",inProceedingsof13thAnnualIEEEInternationalConferenceonHighPerformanceComputing,Dec.2006. [108] W.Zhao,M.Ammar,andE.Zegura,"AMessageFerryingApproachforDataDeliveryinSparseMobileAdHocNetworks,"inProceedingsofACMMobihoc2004,May2004. [109] W.Hsu,D.Dutta,andA.Helmy,"Prole-Cast:Behavior-AwareMobileNetworking,"inProceedingsofIEEEWCNC,LasVegas,NV,Mar.2008. [110] M.Motani,V.Srinivasan,andP.Nuggehalli,"PeopleNet:EngineeringAWirelessVirtualSocialNetwork."inProceedingsofMOBICOM2005,Sep.2005. [111] J.Ghosh,M.J.Beal,H.Q.Ngo,andC.Qiao,"OnProlingMobilityandPredictingLocationsofWirelessUsers,"inProceedingsofACMREALMAN,May2006. [112] W.HsuandA.Helmy,"AnalysisofNodalEncounterPatternsinWirelessLANTraces,"manuscriptinpreparation.Latestversionavailableathttp://nile.cise.u.edu/~weijenhs/SmallWorld.pdf 249

PAGE 250

Wei-JenHsuwasborninTaipei,Taiwan,inMarch1977.HereceivedtheB.S.degreeinelectricalengineeringandtheM.S.degreeincommunicationengineeringfromNationalTaiwanUniversity,respectively,inJune1999andJune2001.Wei-JenstartedhisstudytowardsthePh.D.degreein2003atUniversityofSouthernCalifornia.HereceivedtheEngineerdegreeinelectricalengineeringfromUniversityofSouthernCalifornia,inAugust2006,andtransferredtoUniversityofFloridatocontinuehisstudy.Wei-Jen'sresearchinterestsincludeanalysisofuserdataandbehavior-awareprotocoldesign.Wei-JenisastudentmemberofIEEEandACM. 250