PAGE 1
1
PAGE 2
2
PAGE 3
3
PAGE 4
Firstofall,Iwouldliketothankmyadvisor,Prof.ShigangChen,forhisconstantguidance,support,andinsightfuladvicethroughoutmygraduatestudy.Heisaterricadvisor,apassionateresearcherandacriticalthinker.Withoutthenumerousdiscussionswithhim,theworkpresentedinthisdissertationwouldneverhaveexisted.IamgratefultoProf.SartajSahni,Prof.RandyChow,Prof.JonathanLiu,Prof.TanWong,andProf.LiuqingYang,fortheirinstructivecommentsandsupportduringmystudy.IwouldalsoliketothankallmycolleaguesinProf.Chen'sresearchgroup,includingZhanZhang,MyungKeunYoon,YingJian,MingZhangandTaoLi,forprovidingvaluablefeedbackandhighlevelofresearchsupport.IwouldalsoliketotakethischancetoexpressmyendlesslovetomywifeXiaojieSun,myparents,andmybrother.Withouttheirlove,understanding,encouragementandsupport,noneofthesewouldhavebeenpossible. 4
PAGE 5
page ACKNOWLEDGMENTS ................................. 4 LISTOFTABLES ..................................... 8 LISTOFFIGURES .................................... 9 ABSTRACT ........................................ 11 CHAPTER 1INTRODUCTION .................................. 14 1.1End-to-EndFlowRateFairness ........................ 14 1.2LifetimeFairnessinSensorNetworks ..................... 17 1.3MaximizingLifetimeVectorandMaximizingRateVectorinSensorNetworks 19 1.4RelatedWork .................................. 19 1.4.1FlowRateFairness ........................... 19 1.4.2LifetimeFairnessinSensorNetworks ................. 21 2CROSS-LAYERDESIGNFORACHIEVINGEND-TO-ENDMAXMIN .... 24 2.1NetworkModelandMaxminModel ...................... 24 2.1.1NetworkModel ............................. 24 2.1.2MaxminModel ............................. 25 2.2AGeneralizedMaxminModel ......................... 26 2.2.1ResourcesinWMNs ........................... 26 2.2.2GeneralizedMaxminModel ...................... 27 2.3PacketSchedulingAlgorithm .......................... 29 2.3.1Overview ................................. 29 2.3.2Inter-NodeScheduling ......................... 30 2.4PerformanceEvaluation ............................ 32 2.5Summary .................................... 34 3FULLYDISTRIBUTEDSOLUTIONFORACHIEVINGGLOBALEND-TO-ENDMAXMIN ....................................... 37 3.1Preliminaries .................................. 37 3.1.1NetworkModelandProblemStatement ................ 37 3.1.2CongestionAvoidanceandBuer-BasedBackpressure ........ 39 3.2LinkClassication ............................... 40 3.2.1SaturatedBuer ............................. 40 3.2.2ThreeLinkTypes ............................ 41 3.2.3SaturatedClique ............................ 42 3.3LocalConditionsforGlobalMaxmin:Single-DestinationCase ....... 43 5
PAGE 6
................................ 43 3.3.2NormalizedRate ............................. 44 3.3.3LocalConditionsforGlobalMaxmin ................. 45 3.3.4CorrectnessProof ............................ 47 3.4LocalConditionsforGlobalMaxmin:Multiple-DestinationsCase ..... 52 3.4.1Per-DestinationPacketQueueing ................... 52 3.4.2VirtualNodes,VirtualLinks,andVirtualNetworks ......... 53 3.4.3LocalizedRequirementsforGlobalMaxmin .............. 54 3.5DistributedGlobalMaxminProtocol(GMP) ................. 56 3.5.1Overview ................................. 56 3.5.2MeasurementPeriod .......................... 57 3.5.3AdjustmentPeriod ........................... 59 3.6Simulation .................................... 62 3.6.1EectivenessofGMP .......................... 62 3.6.2PerformanceComparison ........................ 63 3.7Summary .................................... 65 4DISTRIBUTEDPROGRESSIVEALGORITHMFORMAXIMIZINGLIFETIMEVECTORINWIRELESSSENSORNETWORKS ................. 70 4.1NetworkModelandProblemDenition .................... 70 4.1.1SensorNetworkModel ......................... 70 4.1.2VolumeSchedule ............................ 71 4.1.3MaximumLifetimeVectorProblem .................. 72 4.1.4RoutingGraph ............................. 73 4.2NecessaryandSucientConditionsforMaximizingLifetimeVector .... 74 4.3DistributedProgressiveAlgorithm ....................... 77 4.3.1RateSchedule,Volume-BoundDistribution,VolumeSchedule .... 77 4.3.2InitializationPhase ........................... 79 4.3.3IterativePhase|Step1:FromRatestoVolumeBounds ...... 80 4.3.4IterativePhase|Step2:FromVolumeBoundstoVolumesandRates ................................... 82 4.3.5Property ................................. 85 4.3.6TerminationConditions ......................... 89 4.3.7Overhead ................................. 89 4.3.8NetworkDynamics ........................... 90 4.4Simulation .................................... 91 4.4.1ASimpleIllustrativeTestCase .................... 91 4.4.2ConvergenceSpeedofDPA ....................... 92 4.4.3ScalabilityofDPA ............................ 93 4.4.4ComparisonwithHou'sCentralizedAlgorithm ............ 93 4.4.5ComparisonwithOtherCentralizedandDistributedSolutions ... 94 4.5Summary .................................... 95 5CONCLUSION .................................... 100 6
PAGE 7
....................................... 101 BIOGRAPHICALSKETCH ................................ 106 7
PAGE 8
Table page 1-1DualityrelationshipbetweenthetwoproblemsprovedbyHouetal.in[ 1 ] ... 23 2-1SimulationresultsonthetopologyinFig. 2-2 ................... 35 2-2Simulationresultsofthecomplexscenario ..................... 35 3-1SimulationresultsonthetopologyinFig. 3-5 .................... 66 3-2SimulationresultsofweightedmaxmininFig. 3-5 ................. 66 3-3SimulationresultsonthetopologyinFig. 3-6 ................... 66 3-4SimulationresultsonthetopologyinFig. 3-7 .................... 66 3-5SimulationresultsonthetopologyinFig. 3-8 ................... 66 4-1Datasourcelifetimes(indays) ............................ 96 4-2Datasourcevolumes(inthousandsofpackets) ................... 96 4-3SomedatapointsusedtoproduceFig. 4-5 ..................... 96 8
PAGE 9
Figure page 1-1Two-hopowsarestarved. ............................. 23 2-1Asimpleexampleofthegeneralizedmaxminmodel ................ 35 2-2Anexampleofwireless-linkcontentiongraphandcliques ............. 36 2-3Flowsdescribedbythegeneralizedmaxminmodel ................. 36 2-4Schedulingamongcontendingnodes ......................... 36 3-1Amoresophisticatedexampletoillustratethepurposeofthefourlocalconditions 67 3-2Anexampleofrate-limitcondition ......................... 67 3-3Thepathofaowwithanunsaturatedbueratthesource ........... 67 3-4Per-destinationpacketqueueingisnecessarywhentheowspassinganodearedestinedfordierentdestinations. .......................... 68 3-5Networktopologyofasimplescenario ....................... 68 3-6Athree-linkstopology ................................ 68 3-7Networktopology ................................... 68 3-8Networktopology ................................... 69 3-9RatesoftheowsonthetopologyinFig. 3-8 ................... 69 4-1ThereisnoexhaustednodeonP1orP2;nodessandwareunrestrictedfeedingsourcesofi.ThereisanexhaustednodexonP3;nodeuisarestrictedfeedingsourceofi.Thereisnoforwardingpathfromztoi;nodezisapotentialsourceofi. .......................................... 96 4-2IterationsofDPA ................................... 97 4-3Thereisnoexhaustednodefromstoi;nodesisanunrestrictedfeedingsourcesofi.Thereisanexhaustednodexfromutoi;nodeuisarestrictedfeedingsourceofi.Theupstreambottleneckxmaypreventsourceufromfullyutilizingthevolumeboundsetbyionlink(k;i). ...................... 97 4-4Asimpleillustrativetestcase. ............................ 97 4-5MaxdeviationandavgdeviationoflifetimevectorwithrespecttothenumberofiterationsthatDPAhasperformed ........................ 98 4-6DPAscaleswell.Itsoverheadgrowsslowlywiththenetworksize. ........ 98 9
PAGE 10
................ 98 4-8Leftplot:comparisonofnodaloverheaddistributionbetweenLPandDPA.Rightplot:comparisonofmaximumnodaloverheadbetweenLPandDPA ....... 99 4-9NetworklifetimesofDPA,SLPandMPR ..................... 99 4-10AvgandmaxdeviationsofSLPandMPR ..................... 99 10
PAGE 11
Thisstudyfocusesonfairnessinwirelessnetworks.Twofairnessproblemsareaddressed:end-to-endowratefairnessinmultihopwirelessnetworksandlifetimefairnessinwirelesssensornetworks. Inrecentyears,theadventofmultihopwirelessnetworkshasgreatlyacceleratedtheresearchonbandwidthmanagementinsuchnetworkstosupportnewapplications.WhilemuchresearchconcentratesontheMAClayer,theusersperceptiononthesenetworksishoweverdeterminedmainlybasedonthenetworksend-to-endeectiveness.Itisimportantforustodevelopexibletoolsfortracengineeringinmultihopwirelessnetworks.Inthisstudy,twosolutionsareproposedtoachieveend-to-endmaxminowratefairnessinsuchnetworks. Across-layerdesignisrstlyproposedforachievingend-to-endmaxminfairnessinwirelessmeshnetworks.Inthisapproach,ageneralizedmaxminmodelisrstproposedformultihopwirelessnetworks.Atthenetworklayer,ourdesignallocatesnetworkcapacitytoend-to-endowsformaxminbandwidthallocation.AttheMAClayer,ourdesignachievestheallocatedbandwidthsharesforowsthroughatwo-levelweightedfairqueuingalgorithm.Theproposeddesignisabletoequalizetheend-to-endbandwidthallocationtocompetingowsthatsharecommonbottlenecks,whilefullyutilizingthenetworkcapacity.Resultsofsimulationsarepresentedtodemonstratetheeectivenessoftheproposedsolutioninenhancingend-to-endfairness. 11
PAGE 12
Wirelesssensornetworkshaveawiderangeofapplicationsinhabitatobservation,seismicmonitoring,battleeldsensing,etc.Asanothertypeofmultihopwirelessnetwork,asensornetworkconsistsofbattery-poweredsensornodesthatarelimitedinenergysupply.Animportantproblemofwirelesssensornetworksismaximizingtheoperationallifetimeofasensornetwork.Thelifetimeofasensornetworkisdenedasthelifetimesofallsensorsthatproduceusefuldata.Acentralizedsolutionproposedbypreviousworkrequiressolvingasequenceoflinearprogrammingproblems.Thecomputationoverheadcanbeprohibitivelyhighforlargesensornetworks.Collectingthecompleteinformationaboutthenetworkanduploadingthecompleteforwardingpoliciestoallnodesrequiresignicantamountoftransmissions,particularlyfornodesaroundthesink.Weproposeafullydistributedprogressivealgorithmwhichiterativelyproducesaseriesoflifetimevectors,eachbetterthanthepreviousone.Insteadofgivingtheoptimalresultinoneshotafterlengthycomputation,theproposeddistributedalgorithmhasaresultatanytime,andthemoretimespentgivesthebetterresult.Weshowthatwhenthealgorithm 12
PAGE 13
13
PAGE 14
Thetechnologyofwirelessnetworkinghasbeenwidelyadoptedduetoitsadvantagesonaccessibilityandportability.Inamultihopwirelessnetwork,eachnodeoperatesbothasanendhostandasarouter,forwardingpacketsforothernodesthatcannotcommunicatedirectly.Multihopwirelessnetworksprovidemoreexibilityastheyoperateinadecentralizedandself-organizingmanneranddonotrelyonxednetworkinfrastructure.Inrecentyears,theadventofvariousmultihopwirelessnetworks,includingwirelessmeshnetworksandwirelesssensornetworks,hasgreatlyintensiedresearchonsuchnetworkstoimprovetheirapplicabilityinpractice. 2 3 ].\Fair"isdeneddierentlyinthosealgorithms.Onecommonfeatureofthosealgorithmsisthatatleastcertainamountofbandwidthisguaranteedforeverysingle-hopowinthenetwork. Theuser'sperceptiononmultihopnetworksishoweverdeterminedmainlybasedonthenetworks'end-to-endeectiveness.Forexample,fornewuserstoparticipateinawirelessmeshnetwork,theywanttobesurethattheirend-to-endtracistreatedfairlyaseveryoneelse.Moreover,ifausercontributesmoretothenetwork,shemaydemandthathertracisgivenmoreweightthanothers'trac.Inordertomeetdiverseuserrequirements,itisimportantforustodevelopexibletoolsfortracengineeringinmultihopwirelessnetworks. However,thesolutionsforsingle-hopowfairnesscannotbeextendedtoachieveend-to-endowfairnessbecausetheyignoretherelationshipamongthesubowsfromthesamemultihopow.Ifanupstreamsubowisallocatedmorebandwidththanits 14
PAGE 15
Previouswork[ 4 ]haspointedouttheaboveproblemexistinginmultihopowsifsimplyapplyingtheabovesingle-hopowalgorithmstomultihopowsbybreakingeachmultihopowintomultiplesingle-hopows.In[ 4 ],Lipointedoutarelationshipamongallsubowsofamultihopow,whichisthatallsubowsfromthesamemultihopowareexpectedtohavethesamerateandtoreceiveequalamountofbandwidth.However,thebasicfairnessmodelproposedin[ 4 ]hasseriouslimitationthathindersitsapplicabilityinWMNs.Themodelensuresabasicshareofbandwidthforeachend-to-endowinacontendingowgroupandthentriestomaximizetheoverallnetworkthroughput.Thebasicshareiscalculatedasthechannelcapacitydividedbythetotaleectivelengthoftheroutingpathsoftheowsinthegroup.Theeectivelengthofapathisthesmalleroneofthepathlengthand3.Twoows,f1andfn,belongtothesamecontendingowgroupifthereexistsasequenceofows,f2throughfn1,suchthatficontendswithfi+1,1i
PAGE 16
Westudyafundamentalproblem,howtosupportweightedbandwidthallocationamongallend-to-endowsinamultihopwirelessnetworkbasedonIEEE802.11DCF.Amoreprecisebutlessintuitivedenitionoftheproblemishowtoadapttheowratestoachievetheglobalmaxminobjective[ 5 ]:therateofanyowinthenetworkcannotbeincreasedwithoutdecreasingtherateofanotherowwhichhasanequalorsmallernormalizedrate,wherethenormalizedrateisdenedastheowratedividedbytheowweight.Twosolutionsareproposedtoachievetheglobalmaxminobjective. Therstsolutionisacross-layerdesign.Ageneralizedmaxminmodelisproposedformultihopwirelessnetworks.Atthenetworklayer,itallocatesnetworkcapacitytoend-to-endowsformaxminbandwidthallocation.AttheMAClayer,ourdesignachievestheallocatedbandwidthsharesfortheowsthroughatwo-levelweightedfairqueuingalgorithm.Theproposeddesignisabletoequalizetheend-to-endbandwidthallocationtocompetingowsthatsharecommonbottlenecks,whilefullyutilizingthenetworkcapacity. ThesecondsolutionproposedisafullydistributedsolutionthatiscompatiblewithIEEE802.11DCF.Wetransformtheglobalmaxminobjectivetofourlocalconditionsandprovethat,ifthefourlocalconditionsaresatisedinthewholenetwork,thentheglobalmaxminobjectivemustbeachieved.Wethendesignadistributedrateadaptationprotocolbasedonthefourconditions.Wheneveralocalconditionistestedfalseatanode,thenodeinformsthesourcesofcertainselectedowstoadapttheirratessuchthattheconditioncanbesatised.Comparingwith[ 4 ],whichwebelieveisthemostrelatedwork,ourprotocolhasanumberofadvantages.First,itdoesnotmodifythebackoschemeofIEEE802.11.Second,itreplacesper-owqueueingwithper-destinationqueueing.Packetsfromallowstothesamedestinationisqueuedtogether.Thirdandmostimportant,ourprotocolachievesfarbetterfairness(orweightedfairness)amongend-to-endowsthanthebasicfairschemein[ 4 ]. 16
PAGE 17
Thenetworkcannotcarryoutitstaskafterthenodes'energyisexhausted.Hence,maximizingtheoperationallifetimeofasensornetworkisacriticalproblem. Whatisexactlythelifetimeofasensornetwork?Manypriorworks[ 6 { 15 ]denethenetwork'slifetimeasthetimebeforetherstsensorinthenetworkrunsoutofenergy,orbeforetherstlossofcoverage[ 16 ].ThisdenitionsimpliestheproblemofmaximizinglifetimetoalinearprogrammingproblemoranNP-hardnon-polynomialprogrammingproblemifthesinkisallowedtomove[ 15 ].However,inreality,theoperationallifetimeofthenetworkisnotlimitedtothesmallestlifetimeofallnodes.Whenonesensordies,therestofthenetworkcanstillwork,aslongasusefuldatageneratedbyothersensorscanstillreachthesink.Itisnottruethat,sincesensorsaroundthesinkforwardothers'data,theywillalwaysexhausttheirenergyrstandpreventtherestofthenetworkfromreachingthesink.Onecandeploymoresensorsaroundthesink,uselargerbatteriestoboosttheenergylevelthere,orperformin-networkdataaggregation. Anappropriatedenitionforthelifetimeofasensornetworkshouldincludethelifetimesofallsensorsthatproduceusefuldata.Asensor'slifetimeisthedurationfromthetimewhenitbeginstogeneratetherstdatapackettothetimewhenitgeneratesthelastpacketthatisdeliverabletothesink.Thenetwork'slifetimecanbedenedasthevectorofallsensors'lifetimessortedinascendingorder,whichiscalledthelifetimevector.Thevalueofthelifetimevectorisdeterminedbythenodes'packetforwardingpoliciesthatspecifyhowpacketsareforwardedfromthesensorsthroughthenetworktothesink.Morespecically,foreverynode,itsforwardingpolicyspeciestheproportionofpacketsthatshouldbeforwardedoneachoutgoinglinktowardsthesink. 17
PAGE 18
1 17 ]denetheproblemofmaximizingasensornetwork'slifetimeastondthepacketforwardingpoliciesforallnodesthatcollectivelyproducethelexicographicallylargestlifetimevector,calledthemaximumlifetimevector.Inlesspreciseterms,itrstmaximizesthesmallestlifetimeofallnodes,thenmaximizesthesecondsmallestlifetimeofallnodes,andsoon.Houetal.showthatthisproblemcanbemodeledasaseriesoflinearprogramming(LP)problems.AftersolvingtheLPproblems,thesinkuploadstheoptimalpacketforwardingpoliciestothesensors.Basedonitsforwardingpolicy,eachsensorforwarditspackets.Suchasolutionishoweveracentralizedone.ItrequiressolvingO(jNj)LPproblemsofsizeO(jEj),wherejNjisthenumberofsensorsinthenetwork,jEjisthenumberoflinks,andLPhashigh-orderpolynomialcomplexity.Thecomputationoverheadcanbeprohibitivelyhighforlargesensornetworksthatneedtobeoperationalsoonafterdeployment.Collectingthecompleteinformationaboutthenetworkanduploadingthecompleteforwardingpoliciestoallnodesrequiresignicantamountoftransmissionsinthenetwork,particularlyfornodesaroundthesink.Toavoidtheseproblems,adistributedalgorithmthatspreadstheoverheadevenlyonallnodesbecomesimportant. Weproposetherstdistributedsolutionfortheproblemofmaximizingthelifetimevectorofasensornetwork.Ourstrategyistodesignadistributedprogressivealgorithmthatworksinaseriesofiterations,eachproducingaresult(inourcase,alifetimevectoranditscorrespondingforwardingpolicies)thatisbetterthanthepreviousone.Thesequenceofresultsapproachestotheoptimalsolution.Adistributedprogressivealgorithmispracticallyattractivebecausearesultisavailableatanytimeandisgettingbetterasmoretimeisspent.Weshowthatwhenthealgorithmstabilizes,itsresultproducesthemaximumlifetimevector.Wehaveperformedthousandsofsimulationrunsonrandomnetworksofvarioussizes,andcomparedwithHou'scentralizedalgorithmaswellasotherrelatedalgorithms.Theresultsdemonstratethatouralgorithmrapidlyconvergestothemaximumlifetimevectoranditsoverheadissmall.Fornetworksofthousandsofnodes,it 18
PAGE 19
1 ]thatthereexistsanunderlyingdualityrelationshipbetweentheproblemofmaximizingthelifetimevectorofasensornetworkandtheproblemofmaximizingtheratevectorinasensornetwork.ThedualityrelationshipissummarizedinTable 1-1 ,inwhich,giisthelocaldatarateofnodeiandtiisthelifetimeofnodei. 1.4.1FlowRateFairness 18 { 20 ]notonlyrequireper-owqueueingbutalsoassumeaxedbandwidthcapacityforeachlink,whichmakesthemnotapplicableinrandom-accesswirelessnetworks. ItiswellknownthatTCPdoesnotperformwellinwirelessnetworks[ 21 22 ].MuchresearchhasbeendonetoimproveTCP'sperformance,andarecentsurveycanbefoundin[ 23 ].Mostexistingsolutionsemployheuristicmechanismsforbettercongestion 19
PAGE 20
Utility-basedsolutionsonwirednetworks[ 24 { 27 ]alsorequireeachlinktohaveaxedcapacity.Eortshavebeenmadetoadaptutility-basedsolutionsinwirelessnetworksbyconsideringonlysingle-hopows[ 28 29 ],eliminatingcontentionamongneighboringnodesbyusingseparateCDMA/FDMAchannelsforwirelesslinks[ 30 ],modelingresourcesasmaximalcontentioncliquesinsteadofwirelesslinks[ 31 ],relyingoncross-layerdesigntointegrateend-to-endrateadaptationwithMAC-layerpacketscheduling[ 32 ],assumingallwirelesslinkssharethesamechannel[ 33 ],orassumingaxedbandwidthcapacityforeachwirelessnode[ 34 ].Assigningoneseparatechannelforeachcontendingwirelesslink[ 30 ]requiresalargenumberchannelsforadensenetwork,causinglowcapacityforeachchannel.Thisapproachdoesnotworkwellwithwidely-deployedIEEE802.11b/gthathasonlythreenon-overlappingchannels.Themaximalcliqueapproach[ 31 ]requiresthateachclique'seectivecapacityisknown,butitisnotclearhowtoaccuratelymeasuresuchcapacity,whichisacomplexfunctionofnearbycontentionandenvironmentalnoise.Thecross-layerapproach[ 32 ]requiresthenodestodynamicallyestablishgloballycoordinated(orlocallyapproximated[ 35 ])time-slottedtransmissionschedulesattheMAClayer,whichdoesnottwellwithIEEE802.11'srandomaccessmodel.Moreimportantly,theutilityfunctionthatapproximatesmaxminfairnesscontainsanexponentapproachingtoinnity[ 36 ],whichmakesthesystemhardtostabilize.Insummary,existingutility-basedapproachesdonotprovideamaxminsolutionforIEEE802.11DCF. Thereareotherworksthatarenotutility-based.MostofthemaredesignedtoachieveMAC-layerfairness[ 3 37 38 ]ormaxminfairness[ 2 39 ]amongone-hopows.Whilesomestudymultihopows,eachhasitslimitation.Basicend-to-endfairnessinwirelessad-hocnetworksisachievedin[ 4 ].However,thebasicfairshareguaranteedforeachowishighlyconservative;itcanbefarbelowthemaxminrate.End-to-endmaxminisinvestigatedin[ 40 ],whichassumesaseparateCDMA/FDMAchannelforeach 20
PAGE 21
41 ],whichhoweverdoesnotprovideanalgorithmthatcomputesthetemporally-fairrates.Adistributedalgorithmthatachievesaggregatefairnessinsensornetworksisproposedin[ 42 ],assumingthatallowsaredestinedtothesamebasestation.Tothebestofourknowledge,nodistributedalgorithmhasbeenproposedtoprovideweightedmaxminbandwidthallocationinamultihopwirelessnetworkbasedonIEEE802.11DCF. IEEE802.11e[ 43 ]hasbeenunderdevelopmenttosupportQoS,primarilyforWLAN.ItsEDCAprovidesprioritizedchannelaccessonlyforfouraccesscategories(background,besteort,video,andvoice).Itdoesnotprovidene-levelcontrolforweightedbandwidthallocationamongend-to-endows. 6 { 13 44 ]denethenetwork'slifetimeasthetimebeforetherstsensorinthenetworkrunsoutofenergy,orbeforetherstlossofcoverage[ 16 ]. Houetal.showin[ 1 17 ]thattheproblemofmaximizingthelifetimevectorofasensornetworkcanbemodeledasaseriesofcentralizedlinearprogrammingproblems.Houetal.alsoprovein[ 1 ]thatthereexistsanunderlyingdualityrelationshipbetweentheproblemofmaximizingthelifetimevectorofasensornetworkandtheproblemofmaximizingtheratevectorofasensornetworkwithaglobalnodelifetimerequirement. Someresearchersalsodesignenergy-ecientroutingalgorithmstoachievethegoalofminimizingenergyconsumption[ 45 { 49 ].Thetypicalapproach[ 45 46 ]istouseashortestpathalgorithminwhichtheedgecostisthepowerconsumedtotransmitapacketalongthisedge.Thougheectivelyreducingtheenergyconsumptionrate,thisapproachcancauseunbalancedconsumptiondistribution.Thenodesontheminimum-energypatharequicklydrainedofenergy,causingnetworkpartition. 21
PAGE 22
2 andChapter 3 proposetwosolutionstoachievetheglobalend-to-endowratemaxminobjectiveinmultihopwirelessnetworks:across-layersolutionandafullydistributedsolution.Chapter 4 proposesadistributedprogressivealgorithmformaximizinglifetimevectorinwirelesssensornetworks.Chapter 5 concludesourstudy. 22
PAGE 23
Maximizinglifetimevector Totaldatavolumeatnodei:giT=tiR DualityrelationshipbetweenthetwoproblemsprovedbyHouetal.in[ 1 ] Figure1-1. Two-hopowsarestarved. 23
PAGE 24
Inthischapter,across-layerdesignisproposedforachievingend-to-endmaxmininwirelessmeshnetworks(WMNs).Ageneralizedmaxminmodelisrstproposedformultihopwirelessnetworks.Atthenetworklayer,itallocatesnetworkcapacitytoend-to-endowsformaxminbandwidthallocation.AttheMAClayer,ourdesignachievestheallocatedbandwidthsharesfortheowsthroughatwo-levelweightedfairqueuingalgorithm.Theproposeddesignisabletoequalizetheend-to-endbandwidthallocationtocompetingowsthatsharecommonbottlenecks,whilefullyutilizingthenetworkcapacity. Thischapterisorganizedasfollows.Section 2.1 describesthenetworkmodelandourobjective.Section 2.2 presentsageneralizedmaxminmodel,basedonwhichwedesignthemaxminbandwidthallocationalgorithmsforWMNs.Section 2.3 presentsthetwo-levelweightedfairqueuingschedulingalgorithm.Section 2.4 evaluatestheperformanceofoursolution.Section 2.5 summarizesthechapter. 2.1.1NetworkModel 24
PAGE 25
Westudyend-to-endows,whicharereferredtosimplyasows.Aowconsistsofoneormoresingle-hopows,whicharecalledsubows.Twosubowscontendiftheyarecarriedbythesamelinkortwocontendinglinks.Twoowscontendifanyoftheirsubowscontend. Eachowisentitledtoafairshareofnetworkbandwidthinproportiontoitsweight.Itiswellknownthatmaintainingfairnessandmaximizingnetworkthroughputarecontradictivegoals[ 3 ].Stricterfairnesscanbeachievedoftenattheexpenseoflowernetworkthroughput.Somepreviousstudiesfocusedmoreonthroughputoptimizationundercertainbasic,relaxedfairnesscriteria[ 4 37 ].Weputmorefocusonfairness.Specically,wewanttoachievetheclassicalmaxminfairnessamongend-to-endowsin 25
PAGE 26
Theclassicmaxminmodelforwirednetworksisdescribedasfollows.GivenasetQofresources(i.e.,links),acapacitybqforeachresourceq2Q(i.e.,bandwidth),asetFofows,anominalweightwfforeachowf2F,andaroutingpathpfforeachowf,theproblemistoassignaraterfforeachowfsuchthat 1. 2. foranyowf2F,itsraterfcannotbeincreasedwithoutdecreasingtheraterf0ofanotherowf0,forwhichrf0=wf0rf=wf. ThesetofratesR=frfjf2Fgthatsatisfytheaboveconditionsarecalledthemaxminrates. Theabovemodelassumesthateachresourcehasaxedcapacityandthataresourcecanappearinaow'sroutingpathatmostonce.InordertoapplythismodeltoWMNs,wehavetoidentifywhattheresourcesare.Wirelesslinkscannotbeusedastheresourcesbecausetheydonothaveindividuallyxedcapacities.FollowingHuangandBensaou'swork[ 2 ]whichconsidersonlyone-hopows,weshalluse\cliques"fromthecontentiongraphastheresources,whichwillbeexplainedindetailinSection 2.2 .However,inordertoaccommodatethe\cliqueresources"inthecontextofend-to-endows,wemustgeneralizethemaxminmodelrstinthefollowingsectiontoallowaresourcetoappearinaow'sroutingpathformultipletimes. 2.3 26
PAGE 27
Awireless-linkcontentiongraphcanbeemployedtodescribethespatialcontentionrelationshipamongcontendinglinks.Verticesinawireless-linkcontentiongraphrepresentwirelesslinksinthecorrespondingnetworktopology.Twoverticesareconnectedifthecorrespondinglinkscontendwitheachother.Awirelesslinkisidleifthereisnoowpassingit.Asimpliedwireless-linkcontentiongraphcanbeconstructedfromanetworktopologywithallidlelinksremoved.Anexampleofsimpliedwireless-linkcontentiongraphisgivenbyFig. 2-2 (b). Acliqueisacompletesubgraphwithalinkbetweeneverypairofnodes.Amaximumcliqueisacliquethatisnotcontainedinanotherclique.Inthischapter,werefertomaximumcliquesascliqueshenceforth.Acliqueinawireless-linkcontentiongraphrepresentsagroupofmutuallycontendingwirelesslinksinwhichonlyonelinkcanbeintransmissionatanytime.Thechannelbandwidthissharedbyallwirelesslinksofaclique.Thecliquesfromthewireless-linkcontentiongraphcanbeusedasresources. Followingtheroutingpathofaow,wecanobtainasequenceofcliquesthattheowpasses.Whenaowpassesmultiplelinksofaclique,weconsidertheowpassesthecliquemultipletimes.Forawirelesslinkbelongingtomultiplecliques,ifaowpassesthislink,weconsidertheowpassesthosecliquesinturn. Inthegeneralizedmodel,aresourceisallowedtoappearinaow'sroutingpathformultipletimesindierentpositions.Thenumberofappearancesofresourceq2Qinowf'spathpfisdenotedbynqf.Wehavethefollowingfeasibilityconstraintforasetofow 27
PAGE 28
Asetofratesthatsatisestheaboveconstraintissaidtobefeasible.Itismaxminfairifitisfeasibleand,foreachf2F,rfcannotbeincreasedwhilemaintainingfeasibilitywithoutdecreasingrf0foranotherowf0,forwhichrf0=wf0
PAGE 29
28 ].Tosolvethisproblem,weusetheeectivechannelcapacityofacliqueqasbqinourmodel.Anodemeasurestheeectivebitratesofitsincidentlinks.Theconceptofeectivebitrateissimilartotheonein[ 50 ],whichincorporateslinklayerdetails.Aclique'seectivechannelcapacitycanbeobtainedbysumminguptheeectivebitratesofalllinksofthatclique. IninfrastructureWMNs,meshroutershaverelativelystrongcomputingcapabilityandstablepositionswhichmakethecentralizedimplementationofthealgorithmfeasible.Theimplementationcanalsobedistributed.Nodesonlyworkonlocallinkcontentiongraphwhichismuchsmallerthantheglobalone.Theworkofcliquedecompositionisreducedremarkably.Somedistributedmaxminalgorithmsforwirelinenetworks(e.g.,[ 51 ])couldbecustomizedtocalculateowmaxminfairshares. 29
PAGE 30
52 ])canbeadoptedtoachievethis. Therestofthissectiondescribestheinter-nodeschedulingalgorithmthatisbasedonthe802.11DCFwithRTS-CTS-DATA-ACKhandshake. Fig. 2-4 showsallpossiblecontendingpackettransmissionswithintwohopsawayfromrouterxandroutery.InFig. 2-4 ,circlesrepresentrouters.Alinebetweentworoutersmeanstheyarewithinthetransmissionrangeofeachother.Anarrowfromrouterxtoymeansthenext-to-sendpacketofxneedtobetransmittedtoy.GiventheexampleinFig. 2-4 ,thetransmissionfromxtoyconictswithallothertransmissionsindicatedbythearrowsinFig. 2-4 Ifanodexhasapackettotransmit,thecontendingnodesetofx,denotedbyx,isdenedasthegroupofnodesthatarecompetingforthemediaaccesswithx.InFig. 2-4 ,x=fi;j;m;y;n;v;wg.Let+x=x[fxg.Whenxhasapackettotransmitanditsbackotimerbecomeszero,itshouldcompareitstagwiththoseofthenodesinx.Ideally,thepacketfromxshouldbetransmittedimmediatelyifitstagisthesmallest.Otherwise,x'stransmissionshouldbewithhelduntilallpacketsfromxwithsmallertags 30
PAGE 31
Whenxhasapackettobetransmittedtoy,ifTsx>TnxorTsx>Tny,xcanknowitdoesnothavethesmallesttagin+xandthetransmissionshouldbewithheld.Inordertoobtainmostup-to-dateTnxandTny,RTS,CTS,DATAandACKpacketscanpiggybacknecessarytags.Eachroutermaintainsatabletokeeptrackofitsneighbors'tags. However,itisdiculttoenforcetheabovestrictconditionsforeachtransmission.Thereasonisthatsenderxcannotalwayshavethefreshtagsofitsneighbors,especiallyTnyfromreceivery,whicharebasedonthetagsofnodesasfarasthreehopsawayfromx.Staletagsmaycausedeadlocks.Toavoidpotentialdeadlocks,aheuristicmethodisusedbyxtoestimateTny.ThebasicideaistoestimatetheincrementrateofTny,denotedbyrny.Foreachi2Nx,besidesTni,xalsorecordsrniitestimatesandthetimetiwhenTnigetsupdated.Whenxneedstotransmitapackettoy,xuses^TnyinsteadofTnytocheckthesecondcondition,where ^Tny=Tny+rny(tty)(2{2)tisthecurrenttime.OnceTnygetsupdatedandbecomeslarger,thenewrnyiscomputedas: whereisaparametertocontroltheinuenceofTny'snewincrementrateonrny.Ifaboveapproachisemployed,^Tnywilleventuallybeincreasedlargeenoughsuchthatthesecondwithholdingconditionwillbecomefalse. 31
PAGE 32
1. 2. WecomparetheperformanceofMMFwith(1)802.11DCF(abbreviatedas802.11);and(2)thetwo-phaseprotocol(abbreviatedas2PP)proposedin[ 4 ].Wecomparethealgorithmsfromtwoaspects:end-to-endowfairnessandspatialreuseofspectrum. Toevaluatetheend-to-endfairness,weadoptthemaxminfairnessindex[ 5 ](denotedbyImm)andtheequalityfairnessindex[ 53 ](denotedbyIeq).Imm=minf2Ffrfg
PAGE 33
Tomeasurethespatialreuseofspectrum,weemploytheeectivenetworkthroughputU,whichisdenedasPf2Frflf,wherelfisthenumberofhopsontheroutingpathofowf.Thepacketsdroppedbytheintermediatenodesdonotcounttowardstheeectivenetworkthroughputastheydonotcontributetoend-to-endthroughput.Theeectivenetworkthroughputgivesusameasurementfornetworkbandwidthutilizationandtheeciencyofaprotocol. Wepresentsimulationresultsintwonetworkscenarios:asimplenetworktopologyshowninFig. 2-2 andacomplexnetworktopologythatwillbedescribedlater.Allowsinbothscenarioshavetheequalnominalweights.Intherestofthissection,theunitoftheowornetworkthroughputispacketspersecond(PPS). ThesimulationresultsoftheexampleinFig. 2-2 areshownbyTable 2-1 .Thelengthofaowisthenumberofitssubows.MMFshowsgoodend-to-endfairnessandcomparablebandwidthutilization.In2PP,theobjectiveofthebasicfairnessmodelistomaximizethetotalend-to-endthroughput.Thussinglehopowh9;8ihasmuchhigherratethanotherows. ThecomplexscenariosimulatesthetracinthebackboneofaWMN.27nodesareplacedina900900region,inwhich25arenon-gatewaynodesand2aregatewaynodes.Gatewaynodesareevenlyplacedinthehorizontalmidlineoftheregion.Theregionisdividedinto25grids.Eachnon-gatewaynodeisplacedintoagrid.Thelocationofanon-gatewaynodeinitsgridisrandomlychosen.Anon-gatewaynodeconnectstotheInternetthroughthenearestgatewaynode.Everynon-gatewaynodehasadownloadowfromitsgatewaynode.5non-gatewaynodesarerandomlypickedtohave5uploadowstotheirgatewaynodes.Wealsorandomlycreate5internalowsamongnon-gatewaynodes.ThesimulationresultsareshowninTable 2-2 33
PAGE 34
34
PAGE 35
2PP MMF ow length thro. ee.weight thro. ee.weight thro. 114.44 1.00 115.60 1.00 173.74 198.82 2.50 288.65 1.00 173.74 272.21 1.00 114.36 1.67 291.49 414.32 6.00 682.03 1.67 292.24 ee.networkthro. 1814.13 1950.46 1917.66 0.168 0.595 0.627 0.940 Table2-1. SimulationresultsonthetopologyinFig. 2-2 802.11 2PP MMF ee.networkthro. 1550.15 998.86 1528.45 0.026 0.500 0.453 0.895 Table2-2. Simulationresultsofthecomplexscenario Figure2-1. Asimpleexampleofthegeneralizedmaxminmodel 35
PAGE 36
Anexampleofwireless-linkcontentiongraphandcliques Figure2-3. Flowsdescribedbythegeneralizedmaxminmodel Figure2-4. Schedulingamongcontendingnodes 36
PAGE 37
Inthischapter,wepresentafullydistributedapproachtosupportweightedbandwidthallocationamongallend-to-endowsinamultihopwirelessnetwork.Ourgoalistoenablethenetworktoadapttheowratessuchthatglobalmaxmincanbeachieved.InordertodesignafullydistributedsolutionthatiscompatiblewithIEEE802.11DCF,wetransformtheglobalmaxminobjectivetofourlocalconditionsandprovethat,ifthefourlocalconditionsaresatisedinthewholenetwork,thentheglobalmaxminobjectivemustbeachieved.Wethendesignadistributedrateadaptationprotocolbasedonthefourconditions.Wheneveralocalconditionistestedfalseatanode,thenodeinformsthesourcesofcertainselectedowstoadapttheirratessuchthattheconditioncanbesatised.Comparingwith[ 4 ],whichwebelieveisthemostrelatedwork,ourprotocolhasanumberofadvantages.First,itdoesnotmodifythebackoschemeofIEEE802.11.Second,itreplacesper-owqueueingwithper-destinationqueueing.Packetsfromallowstothesamedestinationisqueuedtogether.Thirdandmostimportant,ourprotocolachievesfarbetterfairness(orweightedfairness)amongend-to-endowsthanthebasicfairschemein[ 4 ]. Therestofthechapterisorganizedasfollows.Section 3.1 denesthenetworkmodel.Section 3.2 classieswirelesslinksintothreecategories.Section 3.3 presentsthelocalconditionsforglobalmaxmininwirelessnetworkswithasingledestination.Section 3.4 presentsthelocalconditionsfornetworkswithmultipledestinations.Section 3.5 designsadistributedglobalmaxminprotocolbasedthelocalconditions.Section 3.6 evaluatestheprotocolbysimulations.Section 3.7 summarizesthechapter. 3.1.1NetworkModelandProblemStatement 37
PAGE 38
54 ].Mobilead-hocnetworksarebeyondthescopeofthisstudy.TwonodesareneighborsofeachotherifthereareabletoperformRTS/CTS/DATA/ACKexchange.Twonodesthatarenotneighborscommunicateviaamultihopwirelesspath.Timeisnotslotted.Radiointerferenceisresolvedbyrandombacko.Twowirelesslinkscontendiftheycannottransmitsimultaneously.BasedonthemostpopularMACprotocol,thismodelexcludesthemajorityofrelatedworks[ 30 32 34 35 40 41 ]. LetFbeasetofend-to-endowsinthenetwork.Eachowfhasadesirablerated(f)andaweightw(f).Buttheowsourcewillgeneratenewpacketsatasmallerrateifthenetworkcannotdeliveritsdesirablerate.Theactualrateofowfisdenotedasr(f)(d(f)).Thenormalizedrateofowfisdenedas Inthischapter,whenwereferto\owrate"or\normalizedrateofaow",wemean\end-to-endrate".Theglobalmaxminobjectiveisdenedasfollows:Thenormalizedrate(f)ofanyowfcannotbeincreasedwithoutdecreasingthenormalizedrate(f0)ofanotherowf0,forwhich(f0)(f). Inamoreintuitivebutlessprecisedescription,ourgoalistoequalizethenormalizedratesofallowsasmuchaspossible,particularly,raisingthesmallestones.Directlycompetingowstendtoreceivebandwidthinproportionaltotheirweights.Achievingglobalmaxminisafundamentalfunctionofend-to-endtracengineeringinmultihopwirelessnetworks.Itaddsanewentryintheexistingtoolbox(whichincludesprice-basedandothersolutions)fortracdierentiationamongapplications.Forexample,wemayestablishseveralserviceclassesinthenetworkandassignlargerweightstoapplicationsbelongingtohigherclasses.Howtoenforceacertainweightassignmentschemethroughservicecontractorothermeansisbeyondthescopeofthisstudy. Weassumethereexistsaroutingprotocolthatestablishesaroutingtableateachnode.Theroutingtablemaybeimplicitundergeographicrouting[ 55 56 ],orexplicitly 38
PAGE 39
57 ]orlink-stateroutingprotocol.Consideraspecicdestination.Anodemayreceivepacketsfrommultipleupstreamneighborsandforwardthemtoadownstreamneighbortowardsthedestination.Thelinksfromtheupstreamneighborstoanodearecalledupstreamlinksofthenode,andthelinkfromanodetoitsdownstreamneighboriscalledthedownstreamlink. 3.4.1 .Notethatotherworks[ 4 40 ]requireper-owfairqueueing,whichisamorestringentrequirement.Nowconsiderthepacketstoasinglearbitrarydestination.Anodebuerspacketsreceivedfromupstreamlinksbeforeforwardingthemthedownstreamlink.Thebuerspaceforthequeueislimited.Toavoidpacketdropsduetobueroverow,weadoptthecongestionavoidanceschemein[ 54 ],whichallowsanodeitosenditsdownstreamneighborjapacketonlywhenjhasenoughfreebuerspacetoholdthepacket.Supposethebuerspaceisslottedwitheachslotstoringonepacket.Tokeeptheneighborsupdatedwithj'sbuerstate,wheneverjtransmitsapacket(RTS/CTS/DATA/ACK),itpiggybacksitscurrentbuerstate,forexample,usingonebittoindicatewhetherthereisatleastonefreebuerslot.Whenanupstreamneighborioverhearsapacketfromj,itcachesthebuerstateofj.Ifj'sbuerisnotfull,itransmitsitspacket.Ifj'sbuerisfull,iwillholditspacketandwaituntiloverhearingnewbuerstatefromj.Notethattheresidualbueratnodejchangesonlywhenjreceivesorsendsadatapacket.Wheneverthishappens,jwillsendeitherCTS/ACKorRTS/DATA,immediatelyinformingtheneighborsofitsnewbuerstatethroughpiggybacking.Nocyclicwaitingispossibleifroutingisacyclic.Tohandlefailedoverhearing,iwillstopwaitingandattempttransmittingifitdoesnotoverhearj'sbuerstateforcertaintime.Readersarereferredto[ 54 ]fordiscussiononotherissues. 39
PAGE 40
54 ],whenthebueratjbecomesfull,itforcestheupstreamneighborstoslowdowntoacombinedratethatmatchestherateonthedownstreamlink.1Wheneverjsendsoutapacket,itfreessomebuerspacesuchthattheupstreamneighborscancompetefortransmission.Wheneverjreceivesapacket,itsbuermaybecomefullagainandtheupstreamneighborsmayhavetowaitforthenextreleaseofbueratj.Abuerissaturatedifitcontinuouslyswitchesbetweenfullandunfull,whichslowsdowntheratesofupstreamlinksastheupstreamneighborshavetospenttimewaitingforbuerrelease.Abuerisunsaturatedifitstaysunfull(formostofthetime). 40
PAGE 41
Basedontheabovethreecases,wirelesslinksareclassiedintothreetypes:bandwidth-saturatedlinks,buer-saturatedlinks,andunsaturatedlinks. 41
PAGE 42
Abottlenecklinkmustbebandwidth-saturatedasthereissucientdatatouseupallbandwidthavailabletothelink.Anon-bottlenecklinkiseitherbuer-saturatedlinkoranunsaturatedlink.Theavailablebandwidthisnotfullyutilizedbecauseofadownstreambottleneckintheformercaseorshortageofdatasupplyfromupstreaminthelattercase. 2 4 31 ].Apropercliqueisacliquethatisnotcontainedbyalargerclique.Inthefollowing,whenwerefertoacontentionclique,wealreadymeanaproperclique.Alinkmaybelongtomultiplecliques,consistingofnearbycontendinglinks.Packettransmissionsonthelinksofacliquemustbemadeserially.Therefore,thecombinedrateonalllinksofacliqueisboundedbythechannelcapacity.Acliqueissaturatedifthelinkshaveutilizedallavailablebandwidthsuchthatincreasingtherateononelinkwillalwaysleadtodecreasingtherateonanotherlinkintheclique.Becauseabandwidth-saturatedlinkuses 42
PAGE 43
3.5 .Fornow,weassumethatallowsgotothesamedestination.Theassumptionwillberemovedinthenextsection. 43
PAGE 44
Fortheaboveexample,onemayarguethat,althoughIEEE802.11DCFdoesnotprovidefairness,manyMACprotocols[ 2 3 37 { 39 ]havebeenproposedtoachievethat.IEEE802.11ecanalsoprovidecoarse-levelratecontrol.Buttheexampleisasingleonewithonlyone-hopows.TheseMACsolutionscannotprovideend-to-endfairness,letaloneprovableweightedmaxmin. wherep(f)istheroutingpathofowf.Thereisaneasywayforeachlinktoknowitsnormalizedrate.Whenthesourceofaowproducesnewpackets,itletsthepacketscarrytheow'snormalizedrate.Thenodesofalinkinspectthepassingpacketsandtakethelargestnormalizedratecarriedinthepacketsasthelink'snormalizedrate. Thesetofowsthatpass(i;j)consistsofallowspassingtheupstreamlinksandallowsthatbeginfromi.Bythedenitionofnormalizedrate,wehavethelemmabelow.
PAGE 45
Checkingtheaboveconditionsdoesnotrequireglobalstateoftheentirenetwork.AswewillseeinSection 3.5 wherewedesigntheprotocol,thersttwoconditionscanbetestedbyeachnodeindividuallyandthethirdconditiononlyrequiresinformationexchangeamongnearbynodes,whichcanbeecientlydone.Thefourthconditionrequirestheratelimitataowsourcetobeadditivelyincreaseduntilasource,buer-saturatedorbandwidth-saturatedconditionisviolatedinthenetwork.Whenthishappens,thesourcewillbesignaledtotightenitsratelimit.Forexample,ifthebandwidth-saturatedconditionisviolated,alinklthathasthehighestnormalizedrateinthesaturatedcliquewillbeaskedtoreduceitsrateinordertogiveupsomebandwidthforthebandwidth-saturatedlink.Linklwillidentifythepacketscarryingthelargestnormalizedrateandinformthesourcesofthosepacketstoreducetheirrates.Inresponse,thesourceswillself-imposetighterratelimits. Weillustratethepurposeofthefourlocalconditionsbyacoupleofexamples.First,examinethesimplecaseinSection 3.3.1 ,wherethenetworkhasonlytwowirelesslinks,(i;t)and(j;t).Therearethreeows,onefromitotandtwofromjtot.Assumebothiandjhavesaturatedbuer.Satisfyingthesourceconditionensuresthatthetwoowson(j;t)havethesamenormalizedrate.Satisfyingthebandwidth-saturatedcondition 45
PAGE 46
3.5 .) Figure 3-1 givesamoresophisticatedexample.Whitecirclesrepresentowsources.Greycirclesrepresentothernodes.Thickarrowsrepresentbandwidth-saturatedlinks.Thinarrowsrepresentunsaturatedlinks.Thindashedarrowsrepresentbuer-saturatedlinks.InFigure 3-1 (a),aportionofthenetworkisshownwitheacharrowpointingfromanupstreamnodetoitsdownstreamneighbor.InFigure 3-1 (b),therearesixows,f1throughf6,whoseweightsareshownbesidetheirsources.TheactualdataratesofthelinksareshowninFigure 3-1 (c).(i;j)isabandwidth-saturatedlink,whichsendsbuer-basedbackpressureupstream,creatingbuer-saturatedlinksallthewaytotheowsourcesandslowingtheowrates.InFigure 3-1 (d),thenormalizedratesoftheowsareshownbesidethesources.ThenormalizedratesonthelinksareshowninFigure 3-1 (e). Satisfyingthesourceconditionensuresthatthenormalizedrateofowf4isashighasthatofanyotherupstreamow.Thebuer-saturatedconditionrequiresthatowf1hasthesamenormalizedrateasf2,f3andf4.Becausef1'sweightis2,itsactualrateshouldbetwicethatoff2,f3orf4.Tosatisfythiscondition,ratelimitsmustbeappliedatv,wandxtogivemorebandwidthtou.Satisfyingthebandwidth-saturatedrequirementensuresthatthenormalizedratesofows(f1throughf5)passingthebandwidth-saturatedlink(i;j)areaslargeasanycontendingows(f6).Thismayrequire 46
PAGE 47
Itisnotedthat,therateincrementofaowsourcemayleadtotheviolationofoneormorelocalconditionsbuttheowsourceisnotrequiredbyothernodestoreduceitsrate.Inthatcase,therateincrementattheowsourcedoesnotviolatethelocalconditions.AnexampleisgiveninFig. 3-2 .Thethreelinksareinasaturatedclique.Link(k;t)isbandwidth-saturated.InFig. 3-2 (c),thenormalizedrateoff1islowerthanthoseoftheothertwoows.Byrate-limitcondition,iincreasesthenormalizedrateoff1from1to3.Duetothelimitedbandwidthinthesaturatedclique,thenormalizedrateof(k;t)dropsto4,asshowninFig. 3-2 (d).Bandwidth-saturatedconditionisviolatedas(j;t)haslargernormalizedratethan(k;t).Thusjisrequiredtoreducetherateoff2.Inthisexample,althoughtheincrementoff1'srateresultsintheviolationofthebandwidth-saturatedcondition,onlythesourceoff2,whichisj,violatesthelocalcondition.Fig 3-2 (e)showsthenalrateallocationofthethreeowsthatsatisesallfourlocalconditions.Theratelimitoff1orf2cannotbefurtherincreasedwithoutviolatingthebandwidth-saturatedcondition. Theportionofaow'sroutingpathfromtherstnodewhosebuerissaturatedtotherstbandwidth-saturatedlinkiscalledtheprimarysaturatedsubpathoftheow.Itiseasytoseethattheprimarysaturatedsubpathofaowconsistsofachainofbuer-saturatedlinksandabandwidth-saturatedlinkattheend.Thechainofbuer-saturatedlinksintheprimarysubpathistheresultofbuer-basedbackpressureoriginatedfromthebandwidth-saturatedlink,whichisdemonstratedinFigure 3-1 (c),wherethebottlenecklink(i;j)causestheupstreamlinksbuer-saturated.Itispossiblethattheprimarysaturatedsubpathofaowdoesnothaveabuer-saturatedlink.For 47
PAGE 48
Foraowfwithr(f)
PAGE 49
3-3 .InFig. 3-3 ,whitecirclerepresentstheowsource.Greycirclesrepresentothernodes.Thickarrowsrepresentbandwidth-saturatedlinks.Thinarrowsrepresentunsaturatedlinks.Thindashedarrowsrepresentbuer-saturatedlinks.\"ontopofanodeindicatesanunsaturatedbueratthatnode.\+"indicatesasaturatedbuer.Itispossiblethataowdoesnothaveaprimarysaturatedsubpath.Inthatcase,theprimaryunsaturatedsubpathoftheowformstheentireroutingpath. Proof:Bytherate-limitcondition,whenr(f)isincreased,oneormorelocalconditionsareviolatedandthesourceoffisrequiredtoreducetherateoff.Supposetheamountoff'srateincrementisverysmallandthebueratthesourceoffisstillunsaturated.Theviolationwillnotappearatthesourcebecausethesourceconditionandthebuer-saturatedconditionarenotapplicableatanodewithanunsaturatedbuer.Therefore,theviolationmustappearonatleastonelinkontheroutingpathoff.Becauseanysmallamountofrateincrementonfcanintroduceaviolationandtheratereductionrequestwillalwaysbesenttothesourceoff,thenormalizedrateofthelinkwheretheviolationappearsmustbeequalto(f)beforetherateincrementonf.2
PAGE 50
Proof:Theremustbearatelimitatthesourceoffbecauser(f)
PAGE 51
,wehave(i1;i2)=(f).Bythebuer-saturatedconditionandLemma 1 ,allotherlinksontheprimarysaturatedsubpathhavethesamenormalizedrateas(i1;i2).Then(ik;ik+1)=(f).Therefore,(ik;ik+1)istheprimarybandwidth-saturatedlinkoff. Iftheviolationhappensonalinkafter(i0;i1)ontheroutingpath,thenormalizedrateof(i1;i2)mustbeequalto(f)beforer(f)isincreased.Thiscanbeprovedbycontradiction.Assume(i1;i2)>(f)beforer(f)isincreased.ByLemma 1 ,alllinksontheroutingpathafter(i1;i2)alsohavenormalizedrateslargerthan(f).Amongalllinksontheroutingpathfrom(i1;i2),thereisalinkonwhichtheviolationofowfoccurs.Thenormalizedrateofthatlinkislargerthan(f)beforer(f)isincreased,whichcontradictswithLemma 3 .Bythebuer-saturatedconditionandLemma 1 ,allotherlinksontheprimarysaturatedsubpathhavethesamenormalizedrateas(i1;i2).Then(ik;ik+1)=(f).Therefore,(ik;ik+1)istheprimarybandwidth-saturatedlinkoff. Bythebandwidth-saturatedcondition,(ik;ik+1)hasthelargestnormalizedrateinatleastonesaturatedclique.2 2 andLemma 4 ,wecangetthelemmabelow. Proof:Supposethelocalrequirementsareachieved.Foranarbitraryowfwithr(f)
PAGE 52
ByLemma 5 ,owfhasaprimarylink(i;j)and(i;j)=(f).Itmeansthatthenormalizedratesofallotherowspassing(i;j)arenotgreaterthan(f).Whenweincrease(f)byincreasingtherateoff,basedontheassumption,thenormalizedratesofallotherowspassing(i;j)willnotbedecreased,whichmeansthat(i;j)'sratewillgoup.ByLemma 5 ,(i;j)hasthelargestnormalizedrateinasaturatedclique.When(i;j)'srategoesup,therateofanotherlink(i0;j0)inthesaturatedcliquewillhavetogodown.Amongallowspassing(i0;j0),atleastoneowf0hastodecreaseitsrate(andthus(f0)).Since(i;j)hasthelargestnormalizedrateintheclique,wehave(f0)(i0;j0)(i;j)=(f),whichcontradictswiththepreviousassumption. 3-4 .Whitecirclesrepresentowsources.Blackcirclesrepresentdestinations.Thickarrowsrepresentbandwidth-saturatedlinks.Thinarrowsrepresentunsaturatedlinks.Thindashedarrowsrepresentbuer-saturatedlinks.Figure 3-4 (a)showsaportionofthenetworkwithtwoowswhoseweightsarebothoneanddesirableratesareboth5.InFigure 3-4 (b),eachnodehasonequeueforalldestinations.First,weshowthatonequeuepernodewillunnecessarilyreducetherateoff2inFigure 3-4 (b),where(z;t)isabandwidth-saturatedlink,causingbuer-basedbackpressuretosaturatethebuersatj,i,xandy.Supposetherateoff1is1duetothebottleneck(z;t).Becausethesourcenodes,xandy,competefairlyfortransmissiontoiwheneveri'sbuer 52
PAGE 53
Anodeissaidtoserveadestinationifitisontheroutingpathofaowwiththatdestination.Anodeshouldmaintainaseparatequeueforeachserveddestination,notforeachpassingow.Itshouldbenotedthat,inameshnetwork,manyowsmaydestineforthesamedestination,i.e.,thegatewaytotheInternet.InFigure 3-4 (c),eachnodehasonequeueperserveddestination.Wheniandjkeepseparatequeuesfordestinationstandv,f2willbeabletosendatitsdesirablerateof5. Separatequeuesachieve\isolation"betweenpacketsfordierentdestinations,whichallowsustomodelthephysicalwirelessnetworkasasetofoverlappingvirtualnetworks,eachforonedestination.Figure 3-4 (d)showsthatf1andf2aredeliveredintwovirtualnetworkswithseparatepacketqueuesbutsharingthesamechannel. 53
PAGE 54
3-4 (d),wherethewirelessnetworkismodeledastwovirtualnetworks,and(i;j)astwovirtuallinks. Eachvirtualnetworkcarriesasubsetofows,whichisdisjointfromthesubsetscarriedbyothervirtualnetworks.Buer-basedbackpressure(Section 3.1.2 )isperformedindependentlywithineachvirtualnetwork.Thenormalizedrateofavirtuallinkisdenedasthelargestnormalizedrateofanyowpassingthelink.Withinavirtualnetwork,weclassifyvirtuallinksasbandwidth-saturated,buer-saturated,orunsaturatedinthesamewayaswedidinSection 3.2.2 .Otherconceptscanalsobetriviallyextendedtovirtualnetworks. 3.3.3 tosuitforawirelessnetworkwhoseowshavedierentdestinations. Lemma 1 -Lemma 5 canbeeasilyextendedtovirtualnetworks.
PAGE 55
Proof:Supposethelocalrequirementsareachieved.Foranarbitraryowfwithr(f)
PAGE 56
3.4.3 ,whichisequivalenttomeetingtheglobalmaxminobjectiveinwirelessnetworkswithmultipledestinations. Evenaftertheconditionsaresatised,thenetwork/tracdynamicsmaycausethemtobeviolatedagain.Theprotocolwillcontinuouslychangetheowratestorestoretheconditionsandachieveglobalmaxmininthecurrentnetwork/tracenvironment. Intheprotocoldescription,werefertoaphysicalnodesimplyas\node",denotedas\i",incontrasttoa\virtualnode",denotedas\it"fordestinationt.Werefertoalinkbetweentwophysicalnodesas\wirelesslink",denotedas\(i;j)",whichmaycontainmultiple\virtuallinks",denotedas\(it;jt)".Werefertotheoriginalnetworkas\wireless 56
PAGE 57
Flowfisalocalowatnodeiifiisthesourceoff.Flowfisalocalowofvirtualnodeitiffisalocalowofianditsdestinationist.Theprimaryowofa(virtual)linkistheowthathasthelargestnormalizedrateamongallowspassingthat(virtual)link.Whenmultipleowshavethelargestnormalizedrate,theyareallprimaryows. Belowweexplaintheoperationsperformedinthemeasurementandadjustmentperiods.Notethattheoperationsbyavirtualnodeitareactuallyperformedbythephysicalnodei. 57
PAGE 58
Attheendofameasurementperiod,foreachvirtuallink(it;jt),theendnodesexchangetheirbuerstate,whichcanbepiggybackedinRTS/CTS/DATA/ACKpacketswithoneextrabit(saturatedornot).Basedontheirbuerstate,bothitandjtcandeterminethetypeof(it;jt),whichisbuer-saturated,bandwidth-saturated,orunsaturated. 58
PAGE 59
3.4.3 ,anodemustalsoknowthenormalizedratesandthechanneloccupanciesofallwirelesslinksthatcontendwithanyofitsadjacentlinks. Werefertothenormalizedrateandthechanneloccupancyofawirelesslinkasthestateofthelink.Wemustdisseminatethestateofeachwirelesslink(i;j)toallnodesthathavealinkcontendingwith(i;j).ForIEEE802.11DCF,itincludesallnodesthatarewithintwohopsfromeitheriandj.Thedisseminationprotocolisdescribedasfollows.Recallthatweonlyconsiderstaticwirelessnetworks.Afterdeployment,weassumeeachnodeidiscoversthewirelesstopologyinitstwo-hopneighborhood,andidentiesaminimumsubsetofone-hopneighbors,calledi'sdominatingset,whoseadjacentlinksreachalltwo-hopneighbors.Nodeiinformsthenodesinitsdominatingsetoftheirmembershipintheset.Attheendofeachmeasurementperiod,ifthestateof(i;j)changesfromthepreviousperiod,bothiandjbroadcastthenewstatetotheirone-hopneighbors.Whenanodeintheirdominatingsetsoverhearsthisinformation,thenodere-broadcaststheinformationtoitsneighbors. Thestateofalinkisverysmall.Insteadofmakingaseparatetransmission,suchinformationcanbedisseminatedbypiggybackinginRTS/CTS/DATA/ACKpackets,whichareoverheardbyallnodesinone-hopneighborhood.Inthisdesign,ipiggybacksthestateof(i;j)initsnormaltransmission,andafteroverhearingtheinformation,anodeini'sdominatingsetdoesthesamething.Toovercomefailedoverhearing,thesameinformationshouldbebepiggybackedinanumberoftransmissions.Westressthatthepiggybackdesigncanbeappliedtodisseminateotherinformationintherestoftheprotocolaswell. 59
PAGE 60
Theoperationsperformedbythenodesinthisperiodareexplainedbelow.RemovingUnnecessaryRateLimits 1. If(jt;it)isequaltoL1,thenaratereductionrequestisissuedfortheprimaryowsonvirtuallink(jt;it).IfL1>3S1,itrequeststheprimaryowstohalvetheirrates;otherwise,itrequeststheprimaryowstoreducetheirratesbypercentage.(Themotivationfortheaboveratereductionschemeisstraightforward.Whilereducingbypercentageisthenorm,anoptimizationisadded|whenthegapbetweenL1andS1istoobig,reducingbyhalfhelpstoclosethegapquickly.Thenumber3isarticiallyset.) 2. If(jt;it)isabuer-saturatedlinkand(jt;it)isequaltoS1,thenarateincreaserequestisissuedfortheprimaryowsonvirtuallink(jt;it).IfL1>3S1,itrequeststheprimaryowstodoubletheirrates;otherwise,itrequeststheprimaryowstoincreasetheirratesbypercentage. 60
PAGE 61
Atthebeginningofeachadjustmentperiod,icomputesthechanneloccupancyofeachclique,whichisequaltothesumofthechanneloccupanciesofthewirelesslinksintheclique.Forawirelesslink(i;j)thathasatleastonebandwidth-saturatedvirtuallink,wedothefollowing:First,amongitsbandwidth-saturatedvirtuallinks,weidentifytheone(it;jt)withthesmallestnormalizedrate.Amongallcliquesthat(i;j)belongsto,wetreatthosethathavethelargestchanneloccupancyasbeingsaturated.Second,wecheckwhether(it;jt)satisesthebandwidth-saturatedcondition.If(it;jt)isnotthelargestnormalizedrateinanyofitssaturatedcliques,wemustincrease(it;jt)byissuingrateadjustmentrequests.LetL2bethelargestnormalizedrateonwirelesslinksinallsaturatedcliquesthat(i;j)belongsto.NodeidisseminatesL2,(it;jt),andtheidentiersofsaturatedcliquesviaitsdominatingsettoallnodesintwo-hopneighborhood.Whenanodekreceivesthisinformation,ifawirelesslink(k;m)belongstooneofthosesaturatedcliques,kcallsAdjust(kv;mv;L2;(it;jt))foreachofitsvirtuallinks(kv;mv),with 61
PAGE 62
3-5 .First,weassignallowsthesameweight1,sothataow'snormalizedrateisthesameastheow 62
PAGE 63
3-1 areconsistentwiththeaboveanalysis.Inthesimulation,aftertheowratesarestabilized,(0;1)and(1;2)arebandwidth-saturatedlinks,while(3;4)and(4;5)areunsaturatedlinks.Thebandwidth-saturatedconditionensuresthat,inthesaturatedclique1,thenormalizedrateoff2isnolessthanthoseoff3andf4.Therate-limitconditionensuresthatf1willsendatthehighest-possiblerateaslongasitdoesnotdrivetherateoff2toolowthatviolatesthebandwidth-saturatedconditionoff2inclique1. Nextwetestweightedmaxminonthesamenetworktopologybyassigningdierentweightstoows.ThesimulationresultsaregiveninTable 3-2 .Theratesofthethreeowsinclique1areapproximatelyproportionaltotheirpre-assignedweights.Flowf1hasahigherratethanowf2eventhoughitsweightissmaller.Thatisbecauseitopportunisticallyutilizesallremainingbandwidthinclique0thatcannotbeusedbyf2. 4 ].Thesethreeprotocolsusedierentbuermanagementstrategiestoaccommodatetheirpacketqueuingalgorithms.In802.11,allowspassinganodesharethesamebuerspace.Whenapacketarrivesatanodewhosebuerisfull,itwilloverwritethepacketatthetailofthequeue.In2PP,eachowisallocatedaseparatedqueuethatcanhold10packets.InGMP,allowstothesamedestinationshareacommonqueuethatcanhold10packets. 63
PAGE 64
AsinSection 2.4 ,weadoptthemaxminfairnessindexImmandtheequalityfairnessindexIeqtoevaluatetheend-to-endfairnessandeectivenetworkthroughputUtomeasurethespatialreuseofspectrum. FirstwesimulatethescenarioinFig. 3-6 .ThesimulationresultsareshowninTable 3-3 .GMPismuchfairerthan2PP,whichisinturnmuchfairerthan802.11.Duetothehiddenterminalproblemunder802.11,asevereunfairnessinmediaaccessexistsbetweenlink(0;1)and(2;3)[ 2 ].Node0hasmuchlesschancetograbthechannelwhenithaspacketstobetransmittedtonode1.Thisexplainswhytheowfromnode0tonode3,whichpasses(0;1),hasthelowestrateunder802.11.Theeectivenetworkthroughputsof2PPandGMParecomparable,andtheyarehigherthanthatof802.11,whichdropsmorepacketsduetobueroverow. Thedesignof2PPistoensureabasicfairshareofbandwidthforallowsandthenfavorshortowsinallocatingtheremainingbandwidth.Thebasicfairsharecanbeverysmall,andtherearecasesinwhichitisoutperformedby802.11.WeperformsimulationsonthetopologyinFig. 3-7 ,andtheresultsareshowninTable 3-4 .Withthistopology,thebasicfairsharecalculatedbasedontheformulain[ 4 ]issmall,andtheremainingbandwidthisdistributedheavilybiasedtowardsf2andf8basedonthelinearprogrammingapproachinthesamepaper.Under802.11,theowsinthemiddle(f3,f4,f5andf6)havelowerratesthantheowsonthesides(f1,f2,f7andf8).Thereasonisthataowinthemiddleneedcompeteforbandwidthwithmoreowsthanaowontheside.WithGMP,allowshaveapproximatelyequalratesregardlessoftheirlocationsandlengths.Theowsinthemiddlehaveslightlylowerratesfortwopossiblereasons.First,underGMP,twoowratesareconsideredtobe\equal"iftheirdierenceisbelow,whichis10%inoursimulations.Second,themaximumcombinedrateofthefourlinks 64
PAGE 65
Finally,weperformsimulationsonamorecomplexnetworktopologyshowninFig. 3-8 .Thenetworkconsistsof25nodesthataredeployedina900900m2region.Wecreate25multihopowsinthenetwork,wherethesourceandthedestinationofeachowarerandomlychosen.ThedestinationsoftheowsstartingfromanodearelistedinsquarebracketsafterthesourcenodeID.Thewirelesslinksareshownasedgesinthegraph.Asolidlinemeansthatthelinkisontheroutingpathofatleastoneow.Thetotalnumberofowspassingalinksisshowninparenthesesbesidethatlink.Adottedlinerepresentsanunusedlink.ThesimulationresultsareshowninFig. 3-9 andTable 3-5 .InFig. 3-9 ,theowratesthatareunder100pps(packetspersecond)usethenumbersontheleftverticalaxis;theowratesabove100ppsusethenumbersontherightverticalaxis. Under802.11,halfofallowshaveratesunder10pps.Severalows(e.g.f7andf13)arealmoststarved.Under2PP,threeone-hopows,f0(fromnode6tonode5inFig. 3-8 ),f3(fromnode6tonode1),andf5(fromnode12tonode7),haveveryhighratesandcontributemorethan50%ofthetotalend-to-endthroughput.Thethreeowswhoseratesarearound40pps(f11,f14andf24)arealsoshortowsthatareonlyone-hoportwo-hopslong.GMPachievesfarbetterfairnessasshowninFig. 3-9C 65
PAGE 66
h1;2i h3;5i h4;5i 563.96 196.96 217.57 221.41 Table3-1. SimulationresultsonthetopologyinFig. 3-5 ow h1;2i h3;5i h4;5i 1 2 1 3 rate 527.58 225.40 121.90 377.20 Table3-2. SimulationresultsofweightedmaxmininFig. 3-5 ow 802.11 2PP GMP 131.86 164.75 188.76 176.04 240.85 179.21 1013.96 1025.54 0.547 0.919 0.946 0.999 Table3-3. SimulationresultsonthetopologyinFig. 3-6 ow 802.11 2PP GMP 43.31 145.46 347.81 145.94 43.33 134.26 86.67 132.38 43.39 135.44 86.70 133.04 43.36 141.69 346.96 149.07 1214.93 1674.13 0.125 0.888 0.514 0.998 Table3-4. SimulationresultsonthetopologyinFig. 3-7 802.11 2PP GMP 1672.65 2632.74 0.017 0.206 0.298 0.835 Table3-5. SimulationresultsonthetopologyinFig. 3-8 66
PAGE 67
Amoresophisticatedexampletoillustratethepurposeofthefourlocalconditions Figure3-2. Anexampleofrate-limitcondition Figure3-3. Thepathofaowwithanunsaturatedbueratthesource 67
PAGE 68
Per-destinationpacketqueueingisnecessarywhentheowspassinganodearedestinedfordierentdestinations. Figure3-5. Networktopologyofasimplescenario Figure3-6. Athree-linkstopology Figure3-7. Networktopology 68
PAGE 69
Networktopology B2PP CGMP RatesoftheowsonthetopologyinFig. 3-8 69
PAGE 70
Maximizingtheoperationallifetimeofasensornetworkisacriticalprobleminpractice.Manypriorworksdenethenetwork'slifetimeasthetimebeforetherstsensorinthenetworkrunsoutofenergy.However,whenonesensordies,therestofthenetworkcanstillwork,aslongasusefuldatageneratedbyothersensorscanreachthesink.Moreappropriately,weshouldmaximizethelifetimevectorofthenetwork,consistingofthelifetimesofallsensors,sortedinascendingorder.Forthisproblem,thereexistsonlyacentralizedalgorithmthatsolvesaseriesoflinearprogrammingproblemswithhigh-ordercomplexities.Thischapterproposesafullydistributedprogressivealgorithmwhichiterativelyproducesaseriesoflifetimevectors,eachbetterthanthepreviousone.Insteadofgivingtheoptimalresultinoneshotafterlengthycomputation,theproposeddistributedalgorithmhasaresultatanytime,andthemoretimespentgivesthebetterresult.Weshowthatwhenthealgorithmstabilizes,itsresultproducesthemaximumlifetimevector.Furthermore,simulationsdemonstratethatthealgorithmisabletoconvergerapidlytowardsthemaximumlifetimevectorwithlowoverhead. Therestofthischapterisorganizedasfollows.Section 4.1 givesthenetworkmodelandtheproblemstatement.Section 4.2 laysdownthetheoreticalfoundationforouralgorithm.Section 4.3 proposesourdistributedprogressivealgorithmformaximizingthelifetimevector.Section 4.4 presentsthesimulationresults.Section 4.5 summarizesthechapter. 4.1.1SensorNetworkModel 70
PAGE 71
LetNbethesetofsensornodes,amongwhichthesubsetSthatgeneratenewdataarecalleddatasources,whichmaybetheaggregationnodesrepresentinglocalclusters[ 1 17 ].Letgi;i2N,bethesourcerateatwhichnodeigeneratesnewdatapackets.gi>0ifi2S;gi=0ifi62S.Weassumethatthesourceratesaresetlowenoughtonotcausecongestioninthenetwork.Thesinkmayconsistofmultiplegeographicallydispersedbasestations.Assumethebasestationsareexternallyconnectedtoadatacollector.Itmakesnodierencewhichbasestationadatapacketisroutedto. Twonodesareneighborsiftheycanreceivepacketsfromeachother(tosupportDATA/ACKexchange).Theremaybemultipleroutingpathsfromeachnodetothesink.LetDibethesetofneighborsthatnodeiuseasthenexthopstothesink.Theyarecalleddownstreamneighborsofnodei.8j2Di,(i;j)iscalledanoutgoinglinkofi.LetUibethesetofupstreamneighbors,whichuseiasthenexthopontheirroutingpathstothesink.8k2Ui,(k;i)iscalledanincominglinkofi.Ifiisadownstreamneighborofk,thenkmustbeanupstreamneighborofi.LetE=f(i;j)j8i2N;j2Dig.Wecallthegraphconsistingofalltheselinksastheroutinggraphofthesensornetwork,whichcontainsallroutingpathsfromdatasourcestothesink. 71
PAGE 72
Leteibetheenergyavailableatnodei.Letbetheamountofenergythatanodespendsonreceivingadatapacketfromanupstreamneighbor,ibetheamountofenergythatnodeispendsonproducinganewdatapacket,ibetheamountofenergythatnodeispendsonsendingapacket.Theenergyconstraintisgivenbelow. WesayanodeiisexhaustedifXk2Uiv(k;i)+iv(i)+Xj2Diiv(i;j)=ei: Ifitrequiresperiodicmeasurementofmin/max/avgamongreadingsfromsourcesthathavenotexhaustedyetandremainreachabletothesink,thenanodewillsendapacketforeachsetofpacketsreceivedfromitsupstreamneighborsorgeneratedlocally.Theconstraintbecomes 72
PAGE 73
Thelifetimevectorofthesensornetworkisdenedas(ts;s2S)sortedinascendingorder. Eachfeasiblevolumescheduleproducesafeasiblelifetimevector.Allfeasiblelifetimevectorsformthelifetimespace.OnelifetimevectorTisgreaterthananotherT0ifTislexicographicallylarger|forsomex2[1::jSj],TandT0sharethecommon(x1)smallestelementsbutthexthsmallestelementinTisgreaterthanthexthsmallestelementinT0. Themaximumlifetimevectorproblemistondafeasiblevolumeschedulethatproducesthelargest(orsay,maximum)lifetimevector.Intuitively,itsgoalistorstmaximizethesmallestlifetimeofallsources,thenthesecondsmallest,andsoon. Oncewendthevolumescheduleforthemaximumlifetimevector,thenodesmustknowtheirpacketforwardingpoliciesthatwillrealizethevolumeschedule.Toimplementavolumeschedule,eachnodeisimplydoesthefollowing:1)itgeneratesnewpacketsatitssourcerategiforv(i)packets,and2)itforwardsthereceivedpacketstodownstreamneighborsinweightedroundrobin,usingthevolumesontheoutgoinglinksastheweights.Therefore,thepacketratesontheoutgoinglinksareproportionaltothevolumesonthelinks.Thisiscalledthevolume-rateproperty. wherer(i;j)isthepacketrateonlink(i;j). 73
PAGE 74
4{2 ):Identifyaroutingloopandndthelink(i;j)withthesmallestvolumev(i;j).Deductthelinkvolumesalongtheloopbyv(i;j)andremove(i;j).Repeattheaboveprocedureuntilallloopsareremoved.Theresultingvolumeschedulestillsatisesthevolumeconservationconstraintandtheenergyconstraint.Nosourcevolumehasbeenchanged,andthusthelifetimevectorproducedbythenewvolumescheduleontheacyclicroutinggraphremainsthesame.Furthermore,sincesomelinkvolumeshavebeenreduced,whichmayleaveroomforincreasingsomesourcevolumestoproducealargerlifetimevector.Thesamereasoningcanbeappliedtomodel( 4{3 )withaddeddetailsonhowtoreducevolumesalongacycleandonotherlinks.Butthekeypointisthesame|whenremovingacycle,linkvolumesonlyneedtobereduced. Theacyclicroutinggraphcanbeeasilyconstructedwhenpacketsareforwardedbasedonhopcountsorthenodes'geographiclocationstothesink.Forexample,Dimayconsistofalloraselectedsubsetofneighborsthatareclosertothesink(basedonthehopcountorEuclideandistancetotheclosestbasestation),andUimayconsistofalloraselectedsubsetofneighborsthatarefurtherawayfromthesink. Thevolumeofa(directed)pathisdenedastheminimumvolumeofthelinksonthepath.Apathintheroutinggraphiscalledaforwardingpathifitsvolumeisgreaterthanzero.Otherwise,itiscalledanon-forwardingpath. Nodes2Sisafeedingsourceofnodei2Nifthereisaforwardingpathfromstoi.Furthermore,nodesisarestrictedfeedingsourceofnodeiifthereisanexhaustednodeoneveryforwardingpathfromstoi.Nodesisanunrestrictedfeedingsourceofnodeiifthereisnoexhaustednodeonatleastoneforwardingpathfromstoi,wherethepath
PAGE 75
Wewillestablishthenecessaryandsucientconditionsformaximizingthelifetimevectorinatheorembelow.Belowweexplainabasictechniqueusedintheproof,calledvolumeshift.Understandingthistechniquewillalsohelponetounderstandthedesignofthealgorithm. ConsidertheroutinggraphinFig. 4-1 .Supposesandwaretwounrestrictedfeedingsourcesofnodei.LetP1andP2betwoforwardingpathsthatdonothaveanyexhaustednode.Weshowthatthelifetimeofanunrestrictedfeedingsourcecanbeincreasedattheexpenseofthelifetimeofanother.Todoso,wesimplydecreasethesourcevolumeofs,thendecreasethevolumesonthelinksofP1,increasethesourcevolumeofw,andnallyincreasethevolumesonthelinksofP2,allbythesametinyamount,whichshouldbesmallenoughsuchthatitsadditiononP2doesnotviolatetheenergyconstraint.Theaboveoperationiscalledavolumeshiftfromstowwithrespecttoi.Itiseasytoseethat,aftervolumeshift,thevolumescheduleremainsfeasibleandthelifetimeofsisdecreased,thelifetimeofwisincreased,whilethelifetimesofallothersourcesremainunchanged.Itisobviousthat,toimprovethelifetimevector,weshallalwaysperformavolumeshiftfromanodewithalargerlifetimetoanodewithasmallerlifetime. Notonlycanavolumeshiftbeperformedbetweentwounrestrictedfeedingsources,butalsoitcanbeperformedfromarestrictedfeedingsourceutoanunrestrictedfeedingsources,orfromanunrestrictedfeedingsourcestoapotentialsourcez,butnottheotherwayaround|morespecically,i)avolumeshiftcannotbeperformedfromanunrestrictedfeedingsourcestoarestrictedfeedingsourceubecausewecannotaddanyadditionalvolumetoP3thathasanexhaustednodex;ii)avolumeshiftcannotbeperformedfromapotentialsourceztoanunrestrictedfeedsourcesbecausethevolumeofanypathfromztoiiszeroandthusnothingcanbeshiftedout. 75
PAGE 76
1. Thereisanexhaustednodeoneverypathfromasourcetothesink. 2. Allunrestrictedfeedingsourcesofanodemusthavethesamelifetime,whichshouldbenolessthanthelifetimesoftherestrictedfeedingsourcesofthesamenode,andnogreaterthanthelifetimesofthepotentialsourcesofthesamenode. Proof:First,weprovethattheconditionsarenecessary.Ifafeasiblevolumescheduledoesnotmeeteithercondition,weshowthat,bymodifyingthevolumeschedule,wecanproducealargerlifetimevector.IftherstconditionisnottrueonapathPfromasourcestothesink,wecanimprovethelifetimeofsbyincreasingitssourcevolumeaswellasthevolumeofPbyatinyamount,whichresultsinalargerlifetimevector.Nextconsiderthesecondcondition. 4-1 )fromstowsuchthatthelifetimeofwisslightlyincreased(butstillbelowthatofs),whichresultsinalargerlifetimevector.Notethatthevolumeshiftonlychangesthelifetimesoftwonodes,sandw. Second,weprovethattheconditionsaresucient.Thelifetimespace,consistingofallfeasiblelifetimevectors,isconvexandcompact,whichcanbeseenfromthelinear(ormax)natureoftheenergyconstraint( 4{1 )andthevolumeconservationconstraint( 4{2 )or( 4{3 ),aswellasthelifetimedenition( 4{4 ).RadunovicandLeBoudecshowedthat,inaconvex,compactspace,amax-minvectorexists,andmoreoveritisuniqueandmustbelexicographicallylargestinthespace[ 58 ].Hence,weonlyneedtoshowthatafeasiblevolumeschedulethatmeetsthetwoconditionsproducesthemax-minvector,satisfyingthefollowingrequirement:Thelifetimetsofonesourcescannotbeincreased 76
PAGE 77
Basedontherstcondition,datasourcesshouldaggressivelysettheirsourcevolumestothehighestvaluesthattheirpathstothesinkallow. Thelifetimeofasources,whichisv(s)=gs,canbeinterpretedastheaveragevolumeassignedtoeachunitofrate.Thesecondconditionrequiresthateachunitofratereceivedbyanodeifromanunrestrictedfeedingsourcedeservesthesameamountofvolumeallocation.Inotherwords,forunrestrictedfeedingsources,nodeishouldallocatevolumesinproportiontotheirrates(thatireceivesandforwards).However,eachunitofratefromarestrictedfeedingsource(whichencountersanexhaustednodeonitsforwardingpath)mayreceivelessvolumeallocationatnodei.Moreover,asourceshouldalwaysdirectitspacketstopathsthathavehighestvolumeallocationperunitofrate. 77
PAGE 78
DPAbeginswithaninitialrateschedulethatcanbearbitrarilyset.Fromtheratescheduleandenergyavailabilityatthenodes,itcomputesavolume-bounddistributionbasedonthesecondconditioninTheorem 3 .Fromthevolume-bounddistribution,itsetsavolumeschedule,basedonwhichitwillinturnderiveanewrateschedule.Thiscompletestherstiterationofthealgorithm.AsshowninFig. 4-2 ,ineachsubsequentiteration,DPArepeatstheabovecomputationofanewvolume-bounddistribution(basedontherateschedulefromthepreviousiteration),thenanewvolumeschedule,andnallyanewrateschedule.Eachiterationproducesabettervolumeschedulewhoselifetimevectorislargerthanthepreviousone. Therateschedule,volume-bounddistribution,andvolumeschedulearestoredandcomputedinafully-distributedway.Eachnodeonlymaintainstherates,volumebounds,andvolumesofitsadjacentlinkswithaspacecomplexityofO(jDij+jUij).Becauseeachdirectedlinkissharedbyapairofupstream-downstreamnodes.Somepropertiesofthelinkwillbesetbytheupstreamnodeandthensenttothedownstreamnode,whileotherpropertieswillbesetbythedownstreamnodeandthensenttotheupstreamnode.Detailsaregivenbelow. Nodeiwillsetitsoutgoingrates,r(i;j);j2Di,bydistributingthetotalincomingrateamongtheoutgoinglinks.Itwilllearntheincomingrates,r(k;i);k2Ui,fromupstreamneighborskwhosetthoserates.(Wewanttostressthatthelinkrateshereareauxiliaryvariablesusedtofacilitatethecomputationofvolumes.Theyhavenothingtodowiththeactualdata-packetratesonthelinksatthetimewhenDPAisexecuted.In 78
PAGE 79
Nodeiwillsetitsoutgoingvolumesv(i;j)bydistributingthetotalincomingvolumeamongtheoutgoinglinks.Itwilllearntheincomingvolumesv(k;i)fromupstreamneighborskwhosetthosevolumes. Nodeiwillsetitsincomingvolumeboundsb(k;i)bydistributingitsforwardingcapacityamongtheincominglinks.Itwilllearntheoutgoingvolumeboundsb(i;j)fromdownstreamneighborsjwhosetthosebounds. Intherestofthesection,wewilldescribethedetailsofDPA,whichconsistsofInitializationphaseanditerativephasewitheachiterationhavingtwosteps.Therststepcomputesvolumeboundsbasedonlinkrates.Thesecondstepdetermineslinkvolumesfromvolumeboundsandthencomputesnewlinksrates,whichsetsthestageforthenextiteration. 79
PAGE 80
Intotal,atmostjNjINITpacketsandjNjRATEpacketsaretransmitted.EachnodeisendsoneINITofsizeO(1)andoneRATEpacketofsizeO(jDij).Theinitializationphasecompleteswithinthemaximumroundtriptimebetweenthesinkandanysourceinthenetwork. First,anodeishouldnotreceiveandforwardmorepacketsthanthedownstreamneighborscanhandle.Iftheapplicationmodelischaracterizedby( 4{2 ),thenthecombinedincomingvolumebound(setbyi)shouldnotexceedthecombinedoutgoingvolumebound(setbydownstreamneighbors). whereb(i;j)islearnedbyifromj.Iftheapplicationmodelischaracterizedby( 4{3 ),thentheconstraintbecomes maxfmaxk2Uifb(k;i)g;b(i)gXj2Dib(i;j)(4{7) Second,nodeishouldnotreceiveandforwardmorepacketsthanitsenergyallows. 80
PAGE 81
4{2 );maxfmaxk2Uifb(k;i)g;b(i)gb(i;j) 4{3 ): 4{6 )-( 4{7 ),b0(i;j)b(i;j),andthereforetheaboveconstraintismorerelaxthanonethatreplacesb0(i;j)withb(i;j). Aswehaveexplainedintheprevioussection,thesecondconditionofTheorem 3 requiresthatvolumeallocationshouldbemadeinproportiontotheincomingrates(whichmustbeadjustedforrestrictedfeedingsources,aswillbediscussedshortlyinStep2).Hence,wehavethefollowingrate-boundproperty. Ifr(k;i)=0,thenb(k;i)=0.Ifgi=0,thenb(i)=0. ThedistributedcomputationofStep1isdescribedasfollows:Thesinkbeginstheprocessofsettingvolumeboundsaftertherateinitializationphaseterminates(atthetimewhenthesinkreceivesRATEsfromallupstreamneighbors),orafterStep2completes(atthetimewhenthesinkreceivesVOL RATEpacketsfromallupstreamneighbors|tobedescribedinSection 4.3.4 ).ThesinksetsitsincomingvolumeboundstobeinniteandsendsaBOUNDpacketstoupstreamneighbors,carryingthevolumeboundsofitsincominglinks.AfteranodeireceivesBOUNDsfromalldownstreamneighborsj2Diandlearnsalloutgoingvolumeboundsb(i;j),itsetstheincomingvolumebounds,b(k;i);k2Ui,anditssourcevolumeboundv(i)aslargeaspossible,basedon( 4{9 )subjecttotheconstraintsof( 4{6 )-( 4{7 )and( 4{8 ).NodeithensendsitsincomingvolumeboundstotheupstreamneighborsinaBOUNDpacket. Intotal,jNjBOUNDpacketsaretransmitted.EachnodeionlytransmitsonepacketofsizeO(jUij). 81
PAGE 82
TherstconditionofTheorem 3 requiresustosetthesourcevolumeashighaspossible.Hence,weassign Inadditionto( 4{10 ),outgoinglinkvolumesarealsosubjecttothevolumeconservationconstraintin( 4{2 )or( 4{3 ).Anodecannotsendmorepacketsthanitreceives.Ifitdoesnotreceiveenoughincomingvolumes,itsoutgoingvolumesmayhavetobesetlowerthanwhatthevolumeboundsallow.Ifthevolumeconservationconstraintis( 4{2 ),tosatisfythisconstraint,nodeiassignsitsoutgoingvolumesasfollows. wherev(k;i)issetbyupstreamneighborkandlearnedbyifromk.Ifthevolumeconservationconstraintis( 4{3 ),nodeiassignstheoutgoingvolumestobe 4{10 ). First,weprovebyinductionthatusing( 4{12 )willsatisfytheboundconstraint( 4{10 ).ConsiderthebasecasewithUi=;.By( 4{12 ),( 4{6 )andthefactthatUi=;,we 82
PAGE 83
Nextwemaketheinductiveassumptionthatv(k;i)b(k;i);8k2Ui,andprovethecasewhenUi6=;.(ThisisavalidinductiveassumptionforaDAGroutinggraph,whichhasnoloopforcircularreasoning.)Togetherwith( 4{6 )and( 4{11 ),wehavev(i;j)=(Xk2Uiv(k;i)+v(i))b(i;j) Theinductionproofforthecaseof( 4{13 )issimilar. Thisresult,togetherwith( 4{8 ),ensuresthattheassignedvolumessatisfytheenergyconstraintrequiredin( 4{1 )|toseethis,onehastousethefactthatv(i;j)b0(i;j)dueto( 4{12 )-( 4{13 )and( 4{10 ),whereb0(i;j)isdenedin( 4{8 ).Consequently,theresultingvolumescheduleisfeasible. Afterwesetthelinkvolumes,weassignnewlinkratesbelowbasedontherate-volumepropertyin( 4{5 ),settingthestageforthenextiteration.Forapplicationmodel( 4{2 ), Forapplicationmodel( 4{3 ), 83
PAGE 84
4-3 ,thevolumeboundassignedbyionlink(w;i)foranunrestrictedfeedingsourceswillbefullyutilized.However,thevolumeboundassignedbyionlink(k;i)forarestrictedsourceumaynotbefullyutilizedduetoanupstreambottleneckxthatmaysetatighterboundonthesourcevolumeofu.Inthiscase,thevolumev(k;i),whichissetbykandconstrainedbythelimitedupstreamenergyatx,issmallerthanthevolumeboundb(k;i).Whenthishappens,weshallreduceb(k;i)tomatchv(k;i),andallowb(w;i)tobelarger,whichwillinturnallowstohavealargersourcevolumeandthusalargerlifetime.SincevolumeboundsaresetatStep1basedonlinkrates,wecanachievethereductionofb(k;i)byarticiallyreducingtherater(k;i). Morespecically,afterthelinkratesarecalculatedbasedon( 4{14 )-( 4{15 ),theymaybereducedbymultiplyingareductionfactorf(i)(2(0;1]),whichhasaninitialvalueof1andisupdatedateachiterationasfollows.Supposenodeiisnotadirectneighborofthesink.Ifiisexhausted,i.e.,Pk2Uiv(k;i)+iv(i)+Pj2Diiv(i;j)=ei,oritwasexhaustedinoneofthepreviousiterations,thenitupdatesf(i): whereB(i)andV(i)arethewould-bevolumeboundandvolumeonalloutgoinglinks,respectively,iftheratereductionhadnotbeenpreformedtoreducetheoutgoingvolumeboundinpreviousiterations.Clearly,thevalueoff(i)willstabilizeatanexhaustednodeionlywhenthevolumePj2Div(i;j)matchestheboundPj2Dib(i;j).Afterupdatingf(i),nodeireducestheoutgoingratesasfollows. 84
PAGE 85
4-3 ,bothiandxwillperformtheaboveoperation.Whenxdoesso,itsratereductionwillpropagatedownstream,causingthereductionofr(k;i),whichinturncausesthereductionofb(k;i)andtheincreaseofb(w;i). ThedistributedcomputationofStep2forsettingvolumes/ratesisanaturalcontinuationofStep1.AfteranodewithnoupstreamneighborreceivesBOUND(denedStep1)fromalldownstreamneighbors,itisabletoassignitssourcevolumeby( 4{11 )andoutgoingvolumesby( 4{12 )-( 4{13 ).Itthenupdatesthelinkratesby( 4{14 )-( 4{15 ),( 4{16 ),and( 4{17 ).Afterthat,itsendstheoutgoingvolumes/ratestothedownstreamneighborsbyaVOL RATEpacket.AfteranodeireceivesVOL RATEpacketsfromallupstreamneighborskandlearnsv(k;i),itisabletoassignitssourcevolumeby( 4{11 ),theoutgoingvolumesby( 4{12 )-( 4{13 ),andthenewoutgoingratesby( 4{14 )-( 4{15 ),( 4{16 ),and( 4{17 ).Itsendstheoutgoingvolumes/ratestodownstreamneighborsinVOL RATE.WhenthesinkreceivesVOL RATEfromallupstreamneighbors,itknowsthatStep2iscompleted. Step2transmitsjNjpackets.EachnodeisendsonlyonepacketofsizeO(jDij).Eachiteration,includingStep1andStep2,completeswithinthemaximumroundtriptimebetweenthesinkandanysourceinthenetwork. First,considerthecomputationofvolumebounds.Thetotalforwardingcapabilityofanode,whichisdeterminedby( 4{6 )-( 4{7 )and( 4{8 ),isdistributedasvolumeboundsbasedontherate-boundpropertyin( 4{9 ),whichessentiallyperformsvolumeshiftfromfeedingsourceswithlargervolumeperunitofrate(i.e.,largerlifetime)tothosewithsmallervolumeperunitofrate.Suchvolumeshiftincreasesthelifetimevector.Theonlyproblemisthatavolumeboundmaynotbefullyturnedintovolumeifthereisanupstreamexhaustednodewhichsetsatightervolumebound.Thisproblemissolvedby 85
PAGE 86
4{16 )untilthevolumematchesthebound. Second,thevolumeassignmentsin( 4{10 )and( 4{12 )-( 4{13 )areaggressiveinthesensethattheytrytofullyutilizeallvolumebounds,bysettingthesourcevolumesashighaspossibleandbyforwardingallincomingvolumesateachnode. Third,theratereductionin( 4{14 )-( 4{15 ),( 4{16 )and( 4{17 )articiallydecreasesthelinkratesifthevolumeboundsarenotfullyturnedintothevolumes.Insubsequentiterations,dueto( 4{9 ),decreasedratesleadtodecreasedvolumeboundsonthoselinks,allowingotherlinksthatcanfullyutilizetheirboundstohavehighervolumebounds. Insummary,thevolumeboundcomputationperformsvolumeshiftfromlarge-lifetimesourcestosmall-lifetimesources;thevolumecomputationandtheratereductiontechniqueensurethatthevolumeboundsarefullyutilized.Together,theyimprovethelifetimevectorasDPAexecutesthroughitsiterations.Asthelifetimevectormovesincreasinglyclosertoitsmaximumvalue,theroomforimprovementbecomessmallerandsmaller.OursimulationswillshowthatDPAconvergesrapidly. Proof:LetGbethesubgraphconsistingofallpathsfromsourcestotherstencounteredexhaustednodesortothesinkifnoexhaustednodesareencountered.RatereductionhasnoimpactonthelinkratesinsideG.WhenlinkvolumesarestabilizedinG,linkratesandvolumeboundsmustalsobestabilizedbecausetheirlinearinter-dependencyin( 4{5 ),( 4{9 ),( 4{12 )-( 4{13 )and( 4{14 )-( 4{15 ).Weprovebyinductionthat ConsiderthebasecasewithUi=;.Nodeiisnotexhaustedandhenceiv(i)+Xj2Diiv(i;j)
PAGE 87
4{11 )and( 4{12 )-( 4{13 ),itcanberewrittenasib(i)+Xj2Diib0(i;j)
PAGE 88
4{14 )-( 4{15 ),itsratewillbecontiguouslyshiftedawayfromfeedingi,andeventuallyturnsitselfintoapotentialsourceofi. WehaveprovedearliertherstconditionofTheorem 3 thatthereisanexhaustednodeoneverypathfromasourcetothesink.LetG0bethesubgraphconsistingofallpathsfromsourcestothelastencounteredexhaustednodes,andC0bethesetofthoselastencounteredexhaustednodes,whichformsacutofthenetworkthatseparatesthesinkfromallsources.WhenlinkvolumesarestabilizedinG0,linkratesandvolumeboundsmustalsobestabilizedbecausetheirlinearinter-dependencyin( 4{5 ),( 4{9 ),( 4{12 )-( 4{13 )and( 4{14 )-( 4{15 ).Weprovebycontradictionthat Suppose,9(k;i)2G0,v(k;i)
PAGE 89
By( 4{5 ),( 4{9 ),( 4{14 )-( 4{15 ),( 4{2 )-( 4{3 )and( 4{19 ),theratioofvolumetorateiskeptconstantonanypathsegmentinG0thatdoesnotcontainanexhaustednode(whichperformsratereduction).Itiseasytoseethatthelifetimeofarestrictedfeedingsourceuofanodeicanonlybeequaltoorsmallerthanthatofanunrestrictedsourcebecause,duetoratereduction,theradioofvolumetoratewilldecreasewhenwetraverseapathbackwardfromitouandcrossanexhaustednode. 4 ,weshallterminateDPAwhenithasstabilizedthelinkvolumes,whichcanbedetectedbyaddingaagthatistransitivelycarriedbythecontrolmessages.Theagisinitiallyunset.Anodesetstheagifitchangesalinkvolumebyanamountthatisnotnegligiblysmall.Itisuptotheapplicationrequirementtodecideonhowsmallisnegligible.Thesinkwillstopifitdoesnotreceiveaagthatisset.Alternatively,DPAmayalsobeterminatedarticiallyafteracertainnumberofiterations,orwhentheresultinglifetimevectormeetstheapplicationrequirement. Whiletheoodingdesignitselfmayappearnon-innovative,thenoveltyofDPAisinthedetailsthatestablishestheconstraintsandformulasfornodestoperformlocalizedoperations|iterativelycomputingtheirindividualvolumeboundsfromrates,volumesfromvolumesbounds,andratesfromvolumeswithreduction|yetglobally,asanetoutcome,produceaprogressivelybetterlifetimevector,approachingtotheoptimalresult. 89
PAGE 90
RATEofsizeO(jDij).Upstream/downstreamneighborsrepresentasubsetofallnodeswithinthecommunicationrangeofi.Thepacketsizeislimitedwhenwechooseasmallnumberofupstream/downstreamneighborsforroutingpurpose.WeperformedmanysimulationsinSection 4.4 ,whichshowsthatDPAconvergesquicklytowardstheoptimallifetimevector.Toachievenomorethan5%deviationfromtheoptimal,fornetworksof1,000nodes,lessthan25iterationsareneeded.Inaddition,theoverhead(i.e.numberofiterations)increasesslowlywithnetworksize. Ifthenetworkisdesignedtocollecttensofthousandsofdatapacketsfromeachsource,thesmalloverheadofDPA(intensofcontrolpacketspernode)isnegligible.Ifthenumberofiterationsispre-determined,wecantakethesmallenergyconsumptionofDPAintoaccountbyreducingthenodes'energy(ei)foranappropriateamount. Tokeepupwithchanges,DPAmaybere-executedtocomputeanewvolumeschedule.Thereisatradeobetweenoverheadandbetterlifetimevector.ThefrequencyofexecutingDPAisdependentontheamountofoverheadallowed.Forexample,supposethesinkcollectsaggregateinformationfromthenetworkperiodicallybasedontheapplicationmodelcharacterizedby( 4{3 ),anddatapacketsarelongerthancontrolpackets(INIT/RATE/BOUND/VOL RATE).IfDPAisallowedtoconsumenomorethan0.5%of 90
PAGE 91
Theonlyalternativesolution[ 1 17 ]intheliteratureiscentralized.Itismuchharderforacentralizedalgorithmtohandlenetworkdynamicsbecausethatrequiresthesinktocollectthecompletenetworkinformationbeforeeachexecution. 4-4 ,whereacirclerepresentsasourcenode,asquarerepresentsanon-sourcenode,andthetwonumbersbesideanodearetheinitialenergy(inJoules)andthesourcerate(inpackets/min),respectively.DPAitselfdoesnotdictatehowtheroutinggraphshouldbeconstructed.Instead,itcanworkwithanyroutinggraphthatcontainthepotentialroutingpathsitcanchoosefrom(seeSection 4.1 ).DPAworksattheapplicationlevel;itisindependentofwhichMACprotocolisused.Suppose==0:000012Joule/packetand=0:0000432Joule/packet,whicharechosenbasedontheparametersin[ 59 ]andwillbeusedinalloursimulations. Tables 4-1 showsthelifetimevectorsaftertherst,second,10th,and20thiterationsofDPA,aswellasthemaximumlifetimevector(MLV)inthelastcolumn,whichiscomputednumericallybasedonHou'scentralizedalgorithm[ 1 ].TheresultdemonstratesthatthesequenceoflifetimevectorsproducedbyDPAconvergesrapidlytowardsMLV.Table 4-2 showsthesourcevolumesthatareassignedbyDPAtothesourcenodesaftertherst,second,10th,and20thiterations,aswellastheoptimalsourcevolumesthatproduceMLV.Recallthatthesourcevolumeisthenumberofpacketsthatasourcecan 91
PAGE 92
ConsiderthelifetimevectorVxproducedbyDPAafterthexthiteration.WemeasurehowmuchVxdeviatesfromMLVbythefollowingtwometrics.Lettx(s)bethelifetimeofsourcesinVx.Lett(s)bethelifetimeofsinMLV.ThemaxdeviationofVxisdenedasmaxs2Sfjtx(s)t(s)j 4-5 showstheavg/maxdeviationsoflifetimevectorsproducedbyDPAon500-nodesensornetworks.Thedeviationsdropquicklytoaninsignicantlevelafterasmallnumberofiterations.Eachofthedatapointsusedtoproducetheguresinthissectionistheaverageof100simulationrunsondierentrandomnetworks.Table 4-3 presentssomedatapointsforFig. 4-5 .Forexample,theavg/maxdeviationsaremerely0.066and0.013respectivelyafter20iterations|thatmeans,intheworsecase,thelifetimeofany 92
PAGE 93
4-6 .ItshowsthattheoverheadforDPAtosatisfyatargetdeviation,whichismeasuredbythenumberofiterations,growsslowlywiththenetworksize.Recallthatanodesendsatmost2smallcontrolpacketsineachiteration.Evenforanetworkof3,000nodes,only12iterationsareneededtoachieveanavgdeviationof5%,and32iterationsareneededforamaxdeviationof5%. 1 ]basedoniterativelinearprogramming.OurDPAcanalsobeusedasacentralizedalgorithmwhentheinformationaboutthenetworkisavailableatthesink.ThetargetmaxdeviationforDPAissettobe0.025.Fig. 4-7 comparestherunningtimesofthetwoalgorithms.ItshowsthatDPAareordersofmagnitudefasterthanLP,andthegapwidenswhenthenetworksizeincreases. NextwecomparethecommunicationoverheadofthetwoalgorithmswhenLPisusedasacentralizedalgorithmwhileDPAisusedasadistributedalgorithm.ForLP,thesinkhastocollectnetworkinformation,including,foreachnode,sourcerate(4bytes),nodeenergy(4bytes),transmissionpower(4bytes),nodeID,andIDsofitsdownstreamneighbors(2byteseach).Thesinkhasalsotodownloadtheresultingvolumescheduletothenetwork,whichincludes,foreachnode,itssourcevolumeandthevolumesofitsoutgoinglinks(4byteseach).ForDPA,ineveryiteration,anodesendsoutthevolumes/ratesofitsoutgoinglinksandthevolumeboundsofitsincominglinks(4byteseach). 93
PAGE 94
4-8 showsthenodalcommunicationoverheadinascendingorder.Theoverheadismeasuredbythenumberofbytesthatanodehastotransmit.Clearly,somenodesinLP(attherightendofthegure)bearahugeburdenofcommunicationoverhead. TherightplotinFig. 4-8 showsthemaximumnodaloverheadwithrespecttonetworksize.ThemaximumnodaloverheadofLPincreasesmuchfasterthanthatofDPA. 17 ])thatisalinearprogrammingsolutionformaximizingtheminimumlifetimeofallsources,andMPR(Minimum-PowerRouting[ 45 60 ])thatisadistributedalgorithmforenergy-ecientrouting. FirstwerunDPA,SLP,andMPRon100-noderandomnetworks(withallnodesbeingsources).Fig. 4-9 comparesthelifetimevectorsproducedbythealgorithms.Eachcurverepresentsthelifetimevectorinascendingordergeneratedfromoneofthethreealgorithms.ThesmallestlifetimeinthevectorproducedbyDPAismorethan100%largerthanthatbyMPR.ForSLP,theresultshowsthatmaximizingtheminimumlifetimeofsourcesdoesnotmaximizethelifetimevectorofthenetwork.DPAproducesfarbettersourcelifetimesinthelowerthreequartersofthevector.Second,wecomparethealgorithmsonlargernetworks.Fig. 4-10 showstheavg/maxdeviationsofthelifetimevectorsproducedbySLPandMPRonnetworksof500to3,000nodes(with20%beingsources).ThedeviationsarelargewhencomparingwiththoseofDPA,whichcanbemadearbitrarilysmall. 94
PAGE 95
95
PAGE 96
1stiter. 2nditer. 10thiter. 20thiter. MLV k 41.9 41.9 41.3 41.9 41.9 v 41.9 41.9 41.3 41.9 41.9 u 36.9 63.2 131.2 129.4 125.8 x 33.6 60.7 109.0 121.6 125.8 m 129.9 146.2 158.9 157.3 157.3 j 108.8 154.2 160.2 157.3 157.3 w 151.0 140.5 156.7 157.3 157.3 q 335.5 200.5 239.5 251.7 251.6 Table4-1. Datasourcelifetimes(indays) sources 1stiter. 2nditer. 10thiter. 20thiter. MLV k 60.4 60.4 59.4 60.4 60.4 v 120.8 120.8 118.8 120.8 120.8 u 53.1 90.9 188.9 186.3 181.2 x 48.3 87.5 156.9 175.1 181.2 m 187.0 210.6 228.8 226.5 226.4 j 156.6 222.0 230.6 226.5 226.4 w 434.8 404.6 451.2 453.0 452.9 q 483.1 288.7 344.8 362.5 362.3 Table4-2. Datasourcevolumes(inthousandsofpackets) 10thiter. 20thiter. 30thiter. maxdev. 2.28 0.25 0.066 0.031 avgdev. 0.39 0.045 0.013 0.007 SomedatapointsusedtoproduceFig. 4-5 Figure4-1. ThereisnoexhaustednodeonP1orP2;nodessandwareunrestrictedfeedingsourcesofi.ThereisanexhaustednodexonP3;nodeuisarestrictedfeedingsourceofi.Thereisnoforwardingpathfromztoi;nodezisapotentialsourceofi. 96
PAGE 97
IterationsofDPA Figure4-3. Thereisnoexhaustednodefromstoi;nodesisanunrestrictedfeedingsourcesofi.Thereisanexhaustednodexfromutoi;nodeuisarestrictedfeedingsourceofi.Theupstreambottleneckxmaypreventsourceufromfullyutilizingthevolumeboundsetbyionlink(k;i). Figure4-4. Asimpleillustrativetestcase. 97
PAGE 98
MaxdeviationandavgdeviationoflifetimevectorwithrespecttothenumberofiterationsthatDPAhasperformed Figure4-6. DPAscaleswell.Itsoverheadgrowsslowlywiththenetworksize. Figure4-7. ComparisonofrunningtimebetweenLPandDPA 98
PAGE 99
Figure4-9. NetworklifetimesofDPA,SLPandMPR Figure4-10. AvgandmaxdeviationsofSLPandMPR 99
PAGE 100
Twoimportantproblemsinmultihopwirelessnetworksarestudied.Theyareend-to-endowratefairnessandlifetimefairness. Weproposetwoapproachestoachieveglobalend-to-endowratemaxmininmultihopwirelessnetworks.Therstapproachisacross-layerdesign.Ageneralizedmaxminmodelisrstproposedformultihopwirelessnetworks.Atthenetworklayer,ourdesignallocatesnetworkcapacitytoend-to-endowsformaxminbandwidthallocation.AttheMAClayer,itachievestheallocatedbandwidthsharesfortheowsthroughatwo-levelweightedfairqueuingalgorithm.Wedemonstratetheeectivenessoftheproposedsolutioninenhancingend-to-endfairness.Thesecondapproachproposedisafullydistributedapproach.Wetransformtheglobalmaxminobjectivetofourlocalconditionsandprovethat,ifthefourlocalconditionsaresatisedinthewholenetwork,thentheglobalmaxminobjectivemustbeachieved.Wethendesignadistributedrateadaptationprotocolbasedonthefourconditions.OurapproachdoesnotmodifythebackoschemeofIEEE802.11.Itreplacesper-owqueueingwithper-destinationqueueing.Mostimportant,itachievesfarbetterfairness(orweightedfairness)amongend-to-endowsthanexistingapproaches. Weproposeadistributedprogressivealgorithmformaximizingthelifetimevectorinawirelesssensornetwork,therstalgorithmofitskindforthisproblem.Thedesignofthealgorithmisbasedonthenecessaryandsucientconditionsthatwehaveprovedforproducingthemaximumlifetimevector.Withourprogressivealgorithm,aresultisavailableatanytimeandisgettingbetterasmoretimeisspent.Wedemonstratethatthealgorithmisabletoconvergerapidlytowardsthemaximumlifetimevectorwithlowoverhead. 100
PAGE 101
[1] Y.T.Hou,Y.Shi,andH.D.Sherali,\RateAllocationinWirelessSensorNetworkswithNetworkLifetimeRequirement,"Proc.ofACMMobiHoc'04,pp.67{77,2004. [2] X.L.HuangandB.Bensaou,\OnMax-MinFairnessandSchedulinginWirelessAd-hocNetworks:AnalyticalFrameworkandImplementation,"Proc.ofMobiHoc'01,LongBeach,California,October2001. [3] H.Luo,S.Lu,andV.Bharghavan,\ANewModelforPacketSchedulinginMultihopWirelessNetworks,"Proc.ofMobiCom'00,August2000. [4] B.Li,\End-to-EndFairBandwidthAllocationinMulti-hopWirelessAdHocNetworks,"Proc.ofIEEEICDCS'05,June2005. [5] D.BertsekasandR.Gallager,Datanetworks,2nded.Prentice-HallInc,1992. [6] J.ChangandL.Tassiulas,\Energyconservingroutinginwirelessad-hocnetworks,"Proc.ofIEEEINFOCOM'00,2000. [7] Q.Li,J.Aslam,andD.Rus,\Onlinepower-awareroutinginwirelessAd-hocnetworks,"Proc.ofACMMobiCom'01,pp.97{107,2001. [8] M.BhardwajandA.Chandrakasan,\Boundingthelifetimeofsensornetworksviaoptimalroleassignments,"Proc.ofIEEEINFOCOM'02,vol.3,p.1587C1596,2002. [9] K.Kalpakis,K.Dasgupta,andP.Namjoshi,\Maximumlifetimedatagatheringandaggregationinwirelesssensornetworks,"Proc.ofIEEEICN'02,2002. [10] G.ZussmanandA.Segall,\EnergyEcientRoutinginAdHocDisasterRecoveryNetworks,"Proc.ofIEEEINFOCOM'03,2003. [11] A.SankarandZ.Liu,\MaximumLifetimeRoutinginWirelessAd-hocNetworks,"Proc.ofIEEEINFOCOM'04,2004. [12] R.Madan,Z.Q.Luo,andS.Lall,\Adistributedalgorithmwithlinearconvergenceformaximumlifetimeroutinginwirelesssensornetworks,"Proc.oftheAllertonConferenceonCommunication,ControlandComputing,2005. [13] J.Zhu,S.Chen,B.Bensaou,andK.-L.Hung,\TradeobetweenLifetimeandRateAllocationinWirelessSensorNetworks:ACrossLayerApproach,"Proc.ofIEEEINFOCOM'07,2007. [14] Y.Wu,S.Fahmy,andN.B.Shro,\OntheConstructionofaMaximum-LifetimeDataGatheringTreeinSensorNetworks:NP-CompletenessandApproximationAlgorithms,"Proc.ofIEEEINFOCOM'08,2008. [15] Y.ShiandT.Hou,\TheoreticalResultsonBaseStationMovementProblemforSensorNetworks,"Proc.ofIEEEINFOCOM,April2008. 101
PAGE 102
D.BloughandP.Santi,\Investigatingupperboundsonnetworklifetimeextensionforcell-basedenergyconservationtechniquesinstationaryadhocnetworks,"Proc.ofACMMobiCom'02,p.183C192,2002. [17] Y.T.Hou,Y.Shi,andH.Sherali,\OnLexicographicMax-MinNodeLifetimeforWirelessSensorNetworks,"Proc.ofIEEEICC'04,vol.7,pp.3790{3796,June2004. [18] J.M.Fae,\BottleneckFlowControl,"IEEETransactionsonCommunications,vol.COM-29,no.7,pp.954{962,July1981. [19] F.BonomiandK.Fendick,\TheRate-BasedFlowControlFrameworkfortheAvailableBitRateATMService,"IEEENetwork,pp.9(2):25{39,March/April1995. [20] J.M.Kleinberg,Y.Rabani,andE.Tardos,\FairnessinRoutingandLoadBalancing,"IEEESymposiumonFoundationsofComputerScience,1999. [21] H.Balakrishnan,V.Padmanabhan,S.Seshan,andR.Katz,\AcomparisonofmechanismsforimprovingTCPperformanceoverwirelesslinks,"Proc.ofACMSIGCOMM'96,August1996. [22] H.ZhaiandY.Fang,\DistributedFlowControlandMediumAccessinMultihopAdHocNetworks,"IEEETransactionsonMobileComputing,vol.5,no.11,2006. [23] X.Chen,H.Zhai,J.Wang,andY.Fang,\TCPPerformanceoverMobileAdHocNetworks,"CanadianJournalofElectricalandComputerEngineering,vol.29,pp.129{134,2004. [24] F.Kelly,A.Maulloo,andD.Tan,\Ratecontrolincommunicationnetworks:shadowprices,proportionalfairnessandstability,"JournaloftheOperationalResearch,vol.49,1998. [25] S.H.LowandD.E.Lapsley,\OptimizationFlowControl,I:BasicAlgorithmsandConvergence,"IEEE/ACMTransactionsonNetworking,December1999. [26] L.MassoulieandJ.Roberts,\BandwidthSharing:ObjectivesandAlgorithms,"EEETransactionsonNetworking,vol.10,no.3,June2002. [27] Q.Gao,J.Zhang,andS.Hanly,\Cross-LayerRateControlinWirelessNetworkswithLossyLinks:Leaky-PipeFlow,EectiveNetworkUtilityMaximizationandHop-by-HopAlgorithms,"Proc.ofIEEEINFOCOM'08,April2008. [28] Z.FangandB.Bensaou,\FairBandwidthSharingAlgorithmsbasedonGameTheoryFrameworksinWirelessAd-HocNetworks,"Proc.ofIEEEINFOCOM'04,March2004. [29] X.WangandK.Kar,\DistributedAlgorithmsforMax-minFairRateAllocationinAlohaNetworks,"Proc.ofthe42ndAnnualAllertonConference,Urbana-Champaign,October2004. 102
PAGE 103
Y.YiandS.Shakkottai,\Hop-by-HopCongestionControloveraWirelessMulti-hopNetwork,"Proc.ofIEEEINFOCOM'04,HongKong,China,March2004. [31] Y.Xue,B.Li,andK.Nahrstedt,\OptimalResourceAllocationinWirelessAdHocNetworks:APrice-basedApproach,"IEEETransactionsonMobileComputing,vol.5,no.4,Aprile2006. [32] L.Chen,S.H.Low,M.Chiang,andJ.C.Doyle,\Cross-layerCongestionControl,Routing,andSchedulingDesigninAdHocWirelessNetworks,"Proc.ofIEEEINFOCOM'06,April2006. [33] U.Akyol,M.Andrews,P.Gupta,J.Hobby,I.Saniee,andA.Stolyar,\JointSchedulingandCongestionControlinMobileAd-HocNetworks,"Proc.ofIEEEINFOCOM'08,April2008. [34] Y.QiuandP.Marbach,\BandwidthAllocationinWirelessAd-HocNetworks:APrice-basedApproach,"Proc.ofIEEEINFOCOM'03,March2003. [35] X.LinandN.B.Shro,\TheImpactofImperfectSchedulingonCross-LayerCongestionControlinWirelessNetworks,"IEEE/ACMTrans.onNetworking,vol.14,no.2,April2006. [36] J.MoandJ.Walrand,\FairEnd-to-EndWindow-BasedCongestionControl,"IEEE/ACMTrans.onNetworking,vol.8,no.5,October2000. [37] H.Luo,J.Cheng,andS.Lu,\Self-CoordinatingLocalizedFairQueueinginWirelessAdHocNetworks,"IEEETransactionsonMobileComputing,vol.3,no.1,2004. [38] K.Kar,S.Sarkar,andL.Tassiulas,\AchievingProportionallyFairRatesUsingLocalInformationinAlohaNetworks,"IEEETransactionsonAutomatedControl,vol.49,no.10,2004. [39] L.TassiulasandS.Sarkar,\MaxminFairSchedulinginWirelessNetworks,"Proc.ofIEEEINFOCOM'02,June2002. [40] S.SarkarandL.Tassiulas,\End-to-endBandwidthGuaranteesThroughFairLocalSpectrumShareinWirelessAdhocNetworks,"IEEETransactionsonAutomaticControl,vol.50,no.9,September2005. [41] V.Gambiroza,B.Sadeghi,andE.W.Knightly,\End-to-EndPerformanceandFairnessinMultihopWirelessBackhaulNetworks,"Proc.ofMobicom'04,Philadel-phia,PA,USA,September-October2004. [42] S.ChenandZ.Zhang,\LocalizedAlgorithmforAggregateFairnessinWirelessSensorNetworks,"Proc.ofACMMobicom'06,September2006. [43] \IEEEStandardforTelecommunicationsandInformationExchangeBetweenSystems-LAN/MANSpecicRequirements-Part11:WirelessLANMediumAccessControl 103
PAGE 104
[44] J.P.amdS.Sahni,\AnOnlineHeuristicforMaximumLifetimeRoutinginWirelessSensorNetworks,"IEEETransactionsonComputers,vol.55,no.8,pp.1048{1056,August2006. [45] J.Gomez,A.Campbell,M.Naghshineh,andC.Bisdikian,\Conservingtransmissionpowerinwirelessadhocnetworks,"Proc.of9thInternationalConferenceonNetworkProtocols(ICNP),2001. [46] V.RodopluandT.Meng,\Minimumenergymobilewirelessnetworks,"IEEEJournalofSelectedAreasinCommunications,vol.17,no.8,pp.1333{1344,1999. [47] S.Singh,M.Woo,andC.Raghavendra,\Power-awareroutinginmobilead-hocnetworks,"Proc.ofMOBICOM(1998),pp.181{190,1998. [48] I.StojmenovicandX.Lin,\Power-awarelocalizedroutinginwirelessnetworks,"IEEETran.onParallelandDistributedSystems,vol.12,no.11,pp.1122{1133,2001. [49] R.Wattenhofer,L.Li,P.Bahl,andY.Wang,\Distributedtopologycontrolforpowerecientoperationinmultihopwirelessadhocnetworks,"Proc.ofINFOCOM(2001),2001. [50] Y.Bejerano,S.-J.Han,andL.E.Li,\Fairnessandloadbalancinginwirelesslansusingassociationcontrol,"Proc.ofACMMobiCom'04,2004. [51] J.Mosely,\Asynchronousdistributedowcontrolalgorithms,Ph.D.thesis,"Ph.D.dissertation,MIT,Dept.ofElectricalEngineeringandComputerScience,1984. [52] J.BennettandH.Zhang,\WF2Q:Worst-casefairweightedfairqueueing,"Proc.ofINFOCOM'96,pp.120{128,Mar1996. [53] D.ChiuandR.Jain,\AnalysisoftheIncrease/DecreaseAlgorithmsforCongestionAvoidanceinComputerNetworks,"JournalofComputerNetworksandISDN,vol.17,no.1,pp.1{14,June1989. [54] S.ChenandN.Yang,\CongestionAvoidancebasedonLight-WeightBuerManagementinSensorNetworks,"IEEETransactionsonParallelandDistributedSystems,SpecialIssueonLocalizedCommunicationandTopologyProtocolsforAdHocNetworks,vol.17,no.9,September2006. [55] P.Bose,P.Morin,I.Stojmenovic,andJ.Urrutia,\RoutingwithGuaranteedDeliveryinAdHocWirelessNetworks,"Proc.of3rdInt'lWorkshoponDiscreteAlgorithmsandMethodsforMobileComputingandCommunications(DialM'99),August1999. [56] B.KarpandH.Kung,\GPSR:GreedyPerimeterStatelessRoutingforWirelessNetworks,"Proc.ofACMMobiCom'00,August2000. 104
PAGE 105
C.E.PerkinsandE.M.Royer,\AdHocOn-demandDistanceVectorRouting,"Proc.ofthe2ndIEEEWorkshoponMobileComputingSystemsandApplications,February1999. [58] B.RadunovicandJ.L.Boudec,\AUniedFrameworkforMax-MinandMin-MaxFairnesswithApplications,"40thAnnualAllertonConferenceonCommunication,Control,andComputing,2002. [59] W.Heinzelman,\Application-SpecicProtocolArchitectureforWirelessNetworks,"Ph.D.Thesis,MIT,2000. [60] S.Doshi,S.Bhandare,andT.Brown,\Anon-demandminimumenergyroutingprotocolforawirelessadhocnetwork,"ACMMobileComputingandCommunica-tionsReview,vol.6,no.3,July2002. 105
PAGE 106
LiangZhangwasborninBeijing,China.HereceivedhisBachelorofEngineeringandMasterofEngineeringdegreesincomputerscienceandtechnologyfromTsinghuaUniversity,China,in1999and2002,respectively.Afterthat,hehadworkedinOracleR&DCenterinChinaforoneyear.In2003,hejoinedtheDepartmentofComputerandInformationScienceandEngineeringattheUniversityofFlorida,topursuehisPh.D.degree.HisadvisorisDr.ShigangChen.Hisresearchfocusedonfairnessinmultihopwirelessnetworks. 106
|