<%BANNER%>

Complexity and Delay Characterization of Wireless Scheduling Algorithms

Permanent Link: http://ufdc.ufl.edu/UFE0044112/00001

Material Information

Title: Complexity and Delay Characterization of Wireless Scheduling Algorithms
Physical Description: 1 online resource (97 p.)
Language: english
Creator: Boyaci, Cem
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2012

Subjects

Subjects / Keywords: coloring -- delay -- graph -- networking -- scheduling -- throughput -- wireless
Computer and Information Science and Engineering -- Dissertations, Academic -- UF
Genre: Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Recently, escalating number of wireless network users and popularity of  multimedia-driven applications give rise to an increasing demand  for high throughput and low delay that is to be met by wireless  data networks. The performance of wireless communication is  strongly tied to the underlying medium access control (MAC)  schemes that coordinate scheduling decisions  under interference constraints. In last two decades,  a few number of algorithms are proposed to address efficient  utilization of wireless medium. Among them there is a class  of throughput optimal algorithms which illustrate a diverse  spectrum of requirements on computational resources.  In this work our primary objective is to build a formal relation  between throughput optimality and corresponding  resource demands. Our secondary objective is to derive tight  packet delay bounds for optimal and sub-optimal wireless  scheduling algorithms. Under the protocol model, wireless scheduling problem can be  related to the weighted fractional chromatic number  of the underlying interference graph. In the second chapter  we formalize this relation, and consequently  show that a deterministic version of wireless scheduling  problem is computationally intractable, given queue sizes  are to be kept within practically implementable bounds, unless \NP=\ZPP. In the third chapter, we investigate the delay bounds of  $k$-throughput optimal wireless scheduling algorithms  under stationarity. We reveal a simple result that governs  the manipulation of the negative drift condition, and  tighten the previously provided results. It is also shown that  the same framework can be used to bound the moments  of queue size distribution at stationarity under a popular  suboptimal algorithm called Longest Queue First.  It is possible to derive delay bounds using a single  parameter or a multi-parameter approach. In the fourth  chapter we introduce a new single parameter characterization  of the delay bound under Maximum Weight Scheduling (MWS)  policy. We provide a performance comparison with the previously derived  bounds and underline the significance of our improvement. Next we show  , when the traditional quadratic potential function is employed in the  analysis, the new delay parameter is optimal. Our study reveals the  new parameter is an ideal choice in practice for deriving delay bounds  when a single parameter delay characterization of MWS is desired.  In the fifth chapter we generalize MWS policy using functions  that are integrable and differentiable. The generalization  enables us to use the functions in the weight calculations  and change the activation vectors accordingly. We refine the  sufficient conditions in literature for throughput optimality by  adding new requirements.and show there are additional requirements  for stability. The sixth chapter summarizes our results and a provides  a list of related open problems outlining immediate research  pursuit in the field of wireless scheduling.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Cem Boyaci.
Thesis: Thesis (Ph.D.)--University of Florida, 2012.
Local: Adviser: Xia, Ye.
Local: Co-adviser: Ungor, Alper.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2012
System ID: UFE0044112:00001

Permanent Link: http://ufdc.ufl.edu/UFE0044112/00001

Material Information

Title: Complexity and Delay Characterization of Wireless Scheduling Algorithms
Physical Description: 1 online resource (97 p.)
Language: english
Creator: Boyaci, Cem
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2012

Subjects

Subjects / Keywords: coloring -- delay -- graph -- networking -- scheduling -- throughput -- wireless
Computer and Information Science and Engineering -- Dissertations, Academic -- UF
Genre: Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Recently, escalating number of wireless network users and popularity of  multimedia-driven applications give rise to an increasing demand  for high throughput and low delay that is to be met by wireless  data networks. The performance of wireless communication is  strongly tied to the underlying medium access control (MAC)  schemes that coordinate scheduling decisions  under interference constraints. In last two decades,  a few number of algorithms are proposed to address efficient  utilization of wireless medium. Among them there is a class  of throughput optimal algorithms which illustrate a diverse  spectrum of requirements on computational resources.  In this work our primary objective is to build a formal relation  between throughput optimality and corresponding  resource demands. Our secondary objective is to derive tight  packet delay bounds for optimal and sub-optimal wireless  scheduling algorithms. Under the protocol model, wireless scheduling problem can be  related to the weighted fractional chromatic number  of the underlying interference graph. In the second chapter  we formalize this relation, and consequently  show that a deterministic version of wireless scheduling  problem is computationally intractable, given queue sizes  are to be kept within practically implementable bounds, unless \NP=\ZPP. In the third chapter, we investigate the delay bounds of  $k$-throughput optimal wireless scheduling algorithms  under stationarity. We reveal a simple result that governs  the manipulation of the negative drift condition, and  tighten the previously provided results. It is also shown that  the same framework can be used to bound the moments  of queue size distribution at stationarity under a popular  suboptimal algorithm called Longest Queue First.  It is possible to derive delay bounds using a single  parameter or a multi-parameter approach. In the fourth  chapter we introduce a new single parameter characterization  of the delay bound under Maximum Weight Scheduling (MWS)  policy. We provide a performance comparison with the previously derived  bounds and underline the significance of our improvement. Next we show  , when the traditional quadratic potential function is employed in the  analysis, the new delay parameter is optimal. Our study reveals the  new parameter is an ideal choice in practice for deriving delay bounds  when a single parameter delay characterization of MWS is desired.  In the fifth chapter we generalize MWS policy using functions  that are integrable and differentiable. The generalization  enables us to use the functions in the weight calculations  and change the activation vectors accordingly. We refine the  sufficient conditions in literature for throughput optimality by  adding new requirements.and show there are additional requirements  for stability. The sixth chapter summarizes our results and a provides  a list of related open problems outlining immediate research  pursuit in the field of wireless scheduling.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Cem Boyaci.
Thesis: Thesis (Ph.D.)--University of Florida, 2012.
Local: Adviser: Xia, Ye.
Local: Co-adviser: Ungor, Alper.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2012
System ID: UFE0044112:00001


This item has the following downloads:


Full Text

PAGE 1

COMPLEXITYANDDELAYCHARACTERIZATIONOFWIRELESSSCHEDULINGALGORITHMSByCEMBOYACIADISSERTATIONPRESENTEDTOTHEGRADUATESCHOOLOFTHEUNIVERSITYOFFLORIDAINPARTIALFULFILLMENTOFTHEREQUIREMENTSFORTHEDEGREEOFDOCTOROFPHILOSOPHYUNIVERSITYOFFLORIDA2012

PAGE 2

c2012CemBoyaci 2

PAGE 3

Tomymother 3

PAGE 4

ACKNOWLEDGMENTS IwouldliketoexpressmyappreciationforallthehelpIreceivedfrommyadvisorDr.YeXia.Theenlighteningdiscussionswehad,hisobservations,correctionsandcommentswereallintegraltomywork.IalsowouldliketothankDr.AlperUngorforhissupport,guidance,andpatiencethroughoutmyresearch.Iwouldalsoliketothankmycommitteemembers,Dr.YuguangFang,Dr.TamerKahveci,and,Dr.ShigangChenforalloftheirhelpandvaluablesuggestions.Myfamily'spersistentsupportprovidedmethestrengthandendurancetocontinueatthetimesofmiscellaneousobstacles.Aheartfulthankyougoesto:CanBoyac,CemilBoyac,Nabi,andAytenBalkanl.Finally,thepeoplewhowerealwaysthereforme,myfriends.OnurDursun,Gunes.Yakupoglu,Eris.Aslan,KaganAbidin,BekirArslan,OguzhanTopsakal,HaleErten,FerhatAy,MehmetOnurBaykan,CihanS.ahin,GunhanGulsoy,BradleyNolan,GoktugC.nar,BesimSolakandmanymoreIcannotlisthere.AdditionallyIamthankfultoBoLiforhisfriendshipandforbeingmyprimarycomradeintheprocess. 4

PAGE 5

TABLEOFCONTENTS page ACKNOWLEDGMENTS .................................... 4 LISTOFTABLES ....................................... 7 LISTOFFIGURES ....................................... 8 ABSTRACT ........................................... 9 CHAPTER 1INTRODUCTION .................................... 11 1.1WirelessSchedulingProblem ............................ 11 1.2ProposedModelsandRelatedChallenges ..................... 12 1.2.1WirelessInterferenceModels ........................ 12 1.2.2SystemModel ................................ 13 1.3ResearchOverview ................................. 15 1.3.1OptimalAlgorithms ............................. 15 1.3.2SuboptimalAlgorithms ........................... 16 1.4Contributions .................................... 17 1.4.1EstablishingHardnessofAchievingLowDelay .............. 17 1.4.2BoundsonSumofHigherOrderQueueDistributionMomentsatSta-tionarity ................................... 18 1.4.3SingleParameterCharacterizationofDelayBounds ............ 18 1.4.4GeneralizedMWSPolicy .......................... 18 2COMPLEXITYOFACHIEVINGLOWDELAY .................... 20 2.1Background ..................................... 20 2.2Preliminaries .................................... 22 2.2.1WirelessLinkScheduling .......................... 23 2.2.2FractionalColoringProblem ........................ 24 2.2.3MinimumSpanPacketScheduling ..................... 25 2.3RelevanceofFractionalColoringtoScheduling .................. 27 2.4ComplexityofSchedulingProblem ......................... 31 2.5QueueDynamicsunderDeterministicArrival ................... 33 2.6ACloserLooktotheCapacityRegionunderProtocolModel ........... 36 2.7CapacityRegionandLocalPooling ........................ 38 2.8EffectsofRedeningLoad-slackofArrivalVectors ................ 42 2.9ResultsandDiscussion ............................... 47 3STEADYSTATEDELAYCHARACTERISTICSOFAPPROXIMATEMWSPOLICY 48 3.1Background ..................................... 48 3.2SystemModel .................................... 49 5

PAGE 6

3.3GraphTheoreticalResults ............................. 51 3.4DelayBoundsofApproximateMWSPolicy .................... 54 3.5LQFasanApproximateMWSPolicy ....................... 57 3.6MiscellaneousResultson-LocalPooling .................... 60 3.6.1GraphswithArbitrarilySmallSet-LocalPoolingFactors ........ 61 3.6.2ComputationalComplexityofCalculatingG ............... 63 3.6.3SpecialStructuresThatPreserveG .................... 65 3.7ResultsandDiscussion ............................... 67 4CHARACTERIZINGTHEEXPECTEDQUEUE-SUMBOUNDOFTHEMWSPOLICY .......................................... 68 4.1Background ..................................... 68 4.2DelayCharacterizationUsingaSingleParameter ................. 68 4.2.1ChallengesinComputing() ........................ 70 4.2.2PerformanceComparisonBetween()and() .............. 73 4.3TightnessofDelayParameters ........................... 75 4.3.1DerivingDelayBoundsforaGeneralizedMWSPolicyUsingaQuadraticPotentialFunction .............................. 75 4.3.2EstablishinganIntuitiveMeaningfor():AStepTowardsSinglePa-rameterDelayCharacterizationofMWSPolicy .............. 78 4.4ResultsandDiscussion ............................... 80 5GENERALIZINGMWSPOLICY ............................ 82 5.1Background ..................................... 82 5.2ChoosingWeightFunctioninMWSforStability .................. 83 5.3ResultsandDiscussion ............................... 88 6CONCLUSIONS ..................................... 90 6.1SummaryofResults ................................. 90 6.2FutureResearch ................................... 91 REFERENCES ......................................... 92 BIOGRAPHICALSKETCH .................................. 97 6

PAGE 7

LISTOFTABLES Table page 2-1Atabularcomparisonofnegativeresultsfrom[ 49 ]andTheorem 2.4 .......... 44 7

PAGE 8

LISTOFFIGURES Figure page 1-1Networkgraph,interferencegraphandthematrixrepresentationoftheinterferencerelation .......................................... 14 2-1Interferencegraphforasix-cyclenetworkanditsschedulevectors ........... 37 2-2Capacityregionandarrivalratevectorsindifferentsections ............... 38 3-1IllustrationontightnessofLemma 17 .......................... 53 3-2Hypercubesofdifferentdimensions ........................... 63 3-3IllustrationofGandG' .................................. 65 4-1Anasymmetricnetworkrealizingthequeue-sumboundthrough()beinganim-provementovertheboundthrough() .......................... 71 4-2Vertextransitivitydoesnotrequireorimplymembershipinclass ........... 74 4-3IllustrationofK-cliqueandanodelooselyconnectedforK=6 ............ 74 5-1Conictgraphofathreelinknetworkillustratingthelinetopology ........... 88 8

PAGE 9

AbstractofDissertationPresentedtotheGraduateSchooloftheUniversityofFloridainPartialFulllmentoftheRequirementsfortheDegreeofDoctorofPhilosophyCOMPLEXITYANDDELAYCHARACTERIZATIONOFWIRELESSSCHEDULINGALGORITHMSByCemBoyaciAugust2012Chair:YeXiaCochair:AlperUngorMajor:ComputerEngineeringRecently,escalatingnumberofwirelessnetworkusersandpopularityofmultimedia-drivenapplicationsgiverisetoanincreasingdemandforhighthroughputandlowdelaythatistobemetbywirelessdatanetworks.Theperformanceofwirelesscommunicationisstronglytiedtotheunderlyingmediumaccesscontrol(MAC)schemesthatcoordinateschedulingdecisionsunderinterferenceconstraints.Inlasttwodecades,afewnumberofalgorithmsareproposedtoaddressefcientutilizationofwirelessmedium.Amongthemthereisaclassofthroughputoptimalalgorithmswhichillustrateadiversespectrumofrequirementsoncomputationalresources.Inthisworkourprimaryobjectiveistobuildaformalrelationbetweenthroughputoptimalityandcorrespondingresourcedemands.Oursecondaryobjectiveistoderivetightpacketdelayboundsforoptimalandsub-optimalwirelessschedulingalgorithms.Undertheprotocolmodel,wirelessschedulingproblemcanberelatedtotheweightedfractionalchromaticnumberoftheunderlyinginterferencegraph.Inthesecondchapterweformalizethisrelation,andconsequentlyshowthatadeterministicversionofwirelessschedulingproblemiscomputationallyintractable,givenqueuesizesaretobekeptwithinpracticallyimplementablebounds,unlessNP=ZPP.Inthethirdchapter,weinvestigatethedelayboundsofk-throughputoptimalwirelessschedulingalgorithmsunderstationarity.Werevealasimpleresultthatgovernsthemanipulationofthenegativedriftcondition,andtightenthepreviouslyprovidedresults.Itisalsoshownthatthesameframeworkcanbeusedtoboundthemomentsofqueue 9

PAGE 10

sizedistributionatstationarityunderapopularsuboptimalalgorithmcalledLongestQueueFirst.Itispossibletoderivedelayboundsusingasingleparameteroramulti-parameterapproach.InthefourthchapterweintroduceanewsingleparametercharacterizationofthedelayboundunderMaximumWeightScheduling(MWS)policy.Weprovideaperformancecomparisonwiththepreviouslyderivedboundsandunderlinethesignicanceofourimprovement.Nextweshow,whenthetraditionalquadraticpotentialfunctionisemployedintheanalysis,thenewdelayparameterisoptimal.OurstudyrevealsthenewparameterisanidealchoiceinpracticeforderivingdelayboundswhenasingleparameterdelaycharacterizationofMWSisdesired.InthefthchapterwegeneralizeMWSpolicyusingfunctionsthatareintegrableanddifferentiable.Thegeneralizationenablesustousethefunctionsintheweightcalculationsandchangetheactivationvectorsaccordingly.Werenethesufcientconditionsinliteratureforthroughputoptimalitybyaddingnewrequirements.andshowthereareadditionalrequirementsforstability.Thesixthchaptersummarizesourresultsandaprovidesalistofrelatedopenproblemsoutliningimmediateresearchpursuitintheeldofwirelessscheduling. 10

PAGE 11

CHAPTER1INTRODUCTION 1.1WirelessSchedulingProblemInwirelessmediumsimultaneouslinktransmissionscauseinterferenceinthelackofcoor-dination.Inpracticethismeanstwodevicescannotconductsuccessfulwirelesscommunicationatthesametime,interferingeachotherstransmission,iftheirdistanceisbelowahardwaredenedthreshold.Inasmallwirelessnetworkwithfewdevicestheimpactofusinganoptimalmediumaccessschememightnotbeimmediate.Howeverasthenumberofdevicesandlinksinthenetworkincreasestheperformanceofwirelesscommunicationsystemsprimarilyrelyontheefcientutilizationofthesharedtransmissionmedium.Today'stypicalwirelessrouterscansupporthundredsofuserssimultaneously.Additionallythedemandonhighperformancewire-lesscommunicationisstillincreasingasthewirelesslocalareanetworks(WLAN)becomingthemostpopularmethodofaccessingtheInternet.Mediumaccesscontrol(MAC)protocolsaddresstheefcientmediumutilizationproblembydevisingchannelaccessmechanismsthatoptimizeobjectivesofchoice.MostMACalgorithmstargetavoidingtransmissioncollisionswhilemakingsensibledecisionsonthesetoflinkstoactivate.Throughput,whichingeneralmeansaveragerateofsuccessfuldelivery,hasbeencon-sideredasoneofthemostprevalentnetworkperformancemetricsinthelasttwodecades.Accordingly,thereisamassivebodyofresearchstrivingtodevisethroughputoptimalMACpoliciesthatareresourceefcientaswell.Moreprecisely,lowcomputationalcomplexityandlowdelayarethedesiredfeaturesofawirelessschedulingalgorithmtodeemitapplicableinpractice.Inthisdissertation,ourmainobjectiveistounderstandlimitationsonachievingthedesiredfeaturesofawirelessschedulingalgorithm.Towardsthisend,werstshowthatitisthe-oreticallyhardtoachievecomputationallyefcientalgorithmswithlowdelay.Nextweanalyzemomentsofthequeuesizedistributionatstationarityunderapproximatemaxweightscheduling(WMS)togainfurtherinsightonthedelaycharacteristicsofawiderangeofwirelessschedulingalgorithms. 11

PAGE 12

1.2ProposedModelsandRelatedChallenges 1.2.1WirelessInterferenceModelsWirelessinterferenceisprimarilymodeledusingtwodifferentapproaches.Physicalinterferencemodelcapturespowercontrol,andfeaturesadjustabletransmissionratesthataffectsinterference.Inthismodeltwointerferinglinkscantransmitsimultaneouslygiventheyhavereducedthetransmissionratessufciently.Theinterferenceinthismodelactslikeacontinuousentitythatisdeterminedbythetransmissionrateswhichareingeneralafunctionofthepowerallocation.Protocolmodelontheotherhandassumesapairwiseinterferenceinteractionbetweenlinksanddoesnotallowsimultaneousactivationofinterferinglinks.Inotherwords,protocolmodelsimpliesthecomplexinteractionsdrivenbypowerallocationviasimplyassumingmaximumpowerisusedineachtransmission.Theprotocolmodelhasdifferentvariantswhichcapturegeneralizationsofsimplelimita-tions.Onewaytogeneralizethemodelistoconsiderthepathbetweentwowirelessdevicesthathastheminimumnumberoflinksastheconstrainingfactor.IntheK-hopinterferencemodelthelinksinthenetworkinterfereifandonlyiftheshortestpathbetweentwolinksislessthanorequaltoKinthenetworkgraph.Asanexamplethe2-hopinterferencemodelsuccessfullycapturesthelimitationsimposedbyRTS/CTSmechanismofIEEE802.11,oneofthemostpopularWLANcommunicationstandardstoday.Intheprotocolmodel[ 16 ],interferencebetweenlinkscouldbecapturedbyasymmetric0-1matrix.ForalinksetE,the(i;j)thentryofthejEjjEjinterferencematrixis1iflinksiandjcannottransmitsimultaneouslyandbothsucceed,and0otherwise.Equivalently,theinterferencerelationshipcanalsoberepresentedbyaninterferencegraph(orconictgraph),inwhichanoderepresentsaphysicallinkandanedgerepresentstheexistenceofinterferencebetweentwophysicallinks.Figure 1-1 illustratesawirelessnetworkgraphwithunit-discgraphtopology,itscorrespondinginterferencegraphforthe2-hopinterferencemodelandtheassociatedsymmetric0-1matrix,insequence. 12

PAGE 13

Itisimportanttonotethatindependentsetsofinterferencegraphrepresentfeasiblesetsoflinksthatcanbeactivatedsimultaneouslybywirelessschedulers.Howeverfor1-hopinterferencemodel,usinganinterferencegraphisnotdesiredsincethematchingsintheoriginalnetworkgraphmaptoindependentsetsintheinterferencegraphand,forthemostpart,workingwithmatchingsiseasierthanworkingwithindependentsets.HoweverwhenK>1usinganunderlyinginterferencegraphisjustiedin[ 51 ]byshowingsimilarityinthestructuresofK-hopmatchingsofthenetworkgraphandindependentsetsoftheinterferencegraph.Moreoverwhentheinterferencegraphisapplicable,theeldofgraphtheoryoffersmanyresultsthatarereadilyapplicable.Asanexample;hardnessofschedulingisinvestigatedfor2-hopinterferencemodelin[ 2 ]usinginterferencegraphs.Laterin[ 51 ]thecomplexityresultsaregeneralizedforK2. 1.2.2SystemModelTheinterferencemodelisingeneralappliedtoanunderlyingsystemmodelthatcapturesnetworktrafcwhichconsistsofarrivals,storageandserviceofnetworkpackets.Overallthewirelessnetworkismodeledasatime-slottedqueuedsystem.Weconsiderasetofqueuesoperatingindiscrete-time.Ateachtimeslot,anumberofpacketsfromdifferentqueuescanbeserved,accordingtoascheduleconstrainedtoliewithinaprespeciedsetwhichcapturesinterferenceandschedulingconstraintssomeofwhichareexplainedinthepreviousdiscussion.Whenaqueueisactivatedforserviceapacketfromthequeue,ifthereisone,leavesthenetwork.Newpacketsmayalsoarrivetoeachofthosequeues,accordingtoanexternalarrivalprocesswith,generallyunknowntothescheduler,niterate.Thenetworkdeparturesareassumedtobe0-1,meaningapacketisservedonlyifitsqueueisactivatedbytheschedulerandthereisatleastonepackettoserveinthatqueue.Itisalsosafetoassumearrivalsarenormalizedaccordingtotheservicerates.Itisimportanttonotethatthemodeldescribedaboveisnotonlyapplicabletowirelessnetworks.Indeedthisgeneralmodelhasbeenusedtodescribevarioussettings,includinginput-queuedswitches,akeyfunctionalcomponentofInternetrouters,wheresimultaneouspackettransferfromingressportstoegressportsareconstrainedbyunderlyinghardware.Accordingly 13

PAGE 14

ANetworkgraph BInterferencegraph 2666640111010110110111110100110377775CMatrixrepresentationFigure1-1. Networkgraph,interferencegraphandthematrixrepresentationoftheinterferencerelation.A)Networkgraphwithunit-disktopology.B)Theinterferencegraphpertainingto2-hopinterferencemodel.C)Matrixrepresentationofthewirelessinterference. 14

PAGE 15

theresearchoninput-queuedswitches,whichincorporatewirednetworksetup,isthepredecessorofthewirelessschedulingresearch. 1.3ResearchOverviewIntheirseminalwork[ 52 ],Tassiulasetal.denedthestabilityofthesystemasthepositiverecurrenceofunderlying(irreducible)Markovchain.ThestatespaceofthatMarkovchainisformedbythepossiblequeuesizesandiscountableassumingthediscretearrivalanddepartureprocesses.Therearealsostrongerdenitionsofstabilitysuchastheexpectedvalueofqueuesizesumtobeniteinlimit.Initsintuitiveform,stabilityreferstoqueuesizesbeingboundedonaverage.Anoptimalalgorithmwithagivenstabilitynotionsetstheschedulingdecisionsateachtime-slot(inatime-slottedsystem)suchthat,looselyspeaking,allaveragearrivalratevectorsthatcouldbestabilizedbysomealgorithmarestableundertheoptimalone. 1.3.1OptimalAlgorithmsIn[ 52 ],Tassiulasetal.proposedathroughputoptimalschedulingpolicythat,whenappliedtoprotocolmodel,activatesaweightedmaximumindependentset(WMIS)ofthequeue-weightedinterferencegraphateachtime-slot.HowevertheiralgorithmisnotapplicabletolargenetworkssinceWMISproblemisNP-hardingeneral.Laterin[ 53 ],alineartimeprobabilisticalgorithmwasprovedtoachievethroughputoptimality.Althoughfromthecomplexityperspective,thetwoalgorithmsdiffersignicantlyin[ 27 ]itisshownthatthelattertriestotracktheformerusingarandomizedsearchscheme.ThefollowingresearchrevealedaperiodiccalculationofWMISwouldbesufcienttoachievestabilitywithoutaddressingtheaveragequeuesizebehaviorunderthatsetting[ 47 ].Thealgorithmin[ 52 ]iswidelyreferredasWMISschedulingpolicyinliterature.AnothercommonnameforthesamealgorithmisMaximumWeightScheduling(MWS)policy.Bothtermsareusedinterchangeablyinthiswork.ItisnotnecessaryforanoptimalalgorithmtosolveWMISproblem.Therearerandomizedoptimalalgorithmswhich,giventheaveragearrivalrates,computeactivationprobabilitiesforapolynomialnumberofindependentsetssuchthatthearrivalratesarecoveredbytheactivations.Inasense,thiscorrespondstosolvingtheoptimizationversionofrate-weightedfractional 15

PAGE 16

chromaticnumberproblemderivedusingtheinterferencegraph.HoweverthisapproachofsolvingtheWMISproblemrequirescentralcontrolmechanismsoverthecommunicationnetworkaswellastheknowledgeofaveragearrivalrates.Thelatestresearchstrivestocreatedistributedschemestoachievethroughputoptimality.In[ 21 ]authorsproposedafullydistributedadaptivecarriersensemultipleaccess(CSMA)algorithmtoachievingthatobjective.Thealgorithmachievesthroughputoptimalitywhilekeepingthemessagecomplexitysufcientlylow.Howeverthequestionwhetherthealgorithmproposedin[ 21 ]canreplacetraditionalMAClayerprotocolslikeIEEE802.11dependsonvariouscharacteristicsofthemethodsuchas:transientandstationaryexpectedqueuesizes,momentsofqueuesizedistributionatstationarity,andthecomplexityofimplementingthescheme.Tothatend,thisdissertationconcentratesonalloftheconcernslistedabove. 1.3.2SuboptimalAlgorithmsThehardnessofWMISproblemledresearcherstodevisesuboptimalalgorithmsthatperformwellinpractice.Typicallythosealgorithmsareveryefcientintermsofcomputationalresourcesandtheyareeasytoimplementinpracticalsettings.Variousattemptstocomeupwithsimpleyetefcientalgorithmsconcentrateonmakingsimplelocaldecisionswithoutbeingtoofarawayfromtheoptimalchoice.Thealgorithmslistedin[ 7 ],[ 56 ],[ 25 ],and[ 23 ]areexamplesofthoseattempts.Thosealgorithmstypicallystabilizeascaledthroughputregion.AmongtheclassofsuboptimalalgorithmstheLongestQueueFirst(LQF),whichalsoknownas,GreedyMaximalScheduling(GMS)[ 23 ],isdistinguishedduetoitssimplicityandaptnessfordistributedimplementation.PerformanceofLQFisstudiedin[ 8 ]and[ 24 ]meticulously.In[ 8 ]theauthorsintroducetheconceptoflocal-poolinggraphs,aclassofnetworkinterferencegraphsunderwhichthroughputoptimalityofLQFalgorithmisguaranteed.In[ 24 ]thenotionoflocalpoolingisgeneralizedforarbitrarygraphsviatheconceptof-localpooling.Weprovidedmiscellaneousresultsontopologicalfeaturesof-localpoolingfactorinChapter 3 .Thereisabodyresearchonthroughputcharacterizationofthesuboptimalalgorithms,how-evertothebestofourknowledgetwoimportantaspectsofwirelessschedulingalgorithms:the 16

PAGE 17

delaycharacteristicsandhighermomentsofthequeuelengthdistributionatstationarityarenotwellstudied.WederiveboundsforhighermomentsofqueuelengthdistributionatstationarityunderapproximateMWSpolicyinChapter 3 .AdditionallywehighlighteffectivenessofLQF,asasuboptimalalgorithm,whenqueue-summomentsatstationarityareconcerned.Combinedwithitssimplicity,ourresultdeemsLQFtobeanidealsuboptimalalgorithminpracticalsettings. 1.4ContributionsAsaresultofourmotivation,westudydelaycharacteristicsofschedulingalgorithmsinwirelesscommunicationnetworks.Thefollowingsectionssummarizeourcontributionsinthisdissertation. 1.4.1EstablishingHardnessofAchievingLowDelayThewiderangeofcomputationalresourcerequirementsofthroughputoptimalalgorithmshasbeenoneofthepuzzlingaspectsofthewirelessschedulingresearchinlasttwodecades.TherstthroughputoptimalalgorithmhadanNP-Completesub-problem,whilesubsequentdevelopmentsrevealedalineartimerandomizedalgorithmprovidingmaximumthroughput.Weanalyzedthetraditionalqueuebased,time-slottedsystemutilizingthecomputationalcomplexitytheory.Ourapproachisparalleltothemethodologyprovidedin[ 1 ]sothatanetworkowproblemisattackedbysimulatingaqueueingnetworkandconductingbinarysearchusingthesimulationresult.Tothebestofourknowledgethisworkistherstattempttogetherwithsimultaneousindependentworkprovidedin[ 49 ]torevealanticipatedtrade-offbetweenper-time-slotcomplexityandqueuesizes.Ourwork'simmediateconsequenceisapointerforfutureresearchersintheeldofwirelessscheduling:athroughputoptimalalgorithmcouldonlybeeffectiveeitherintermsofcomputationalresourcesorintermsthetransientexpectedqueuesizes.AquesttosearchforefcientalgorithmsisidenticaltothequesttondapolynomialtimealgorithmforanNP-Completeproblem.Accordinglyonecanprefertodesignalgorithmstotargetspecialtopologieswherethenegativeresultnolongerapplies.Additionallyitmightbepossibletoabandonthroughputoptimalitytoworkaroundthislimitation.HoweverthatshouldbedonecarefullysinceadeeperanalysisofmaintheoreminChapter 2 revealsanalgorithm 17

PAGE 18

whichisefcientonlyfortheuniformarrivalvectorsstillfacesthelimitationsthatarementionedabove. 1.4.2BoundsonSumofHigherOrderQueueDistributionMomentsatStationarityInChapter 3 weprovideaframeworktostudyhighermomentsofthequeuelengthdistri-butionatstationarityofthesystemwhenanapproximateWMISschedulerisutilized.DelayboundsofWMISschedulingisstudiedin[ 26 ].Ourresults,buildingonthesameframeworkofquadraticpotentialfunction,improvetheboundpresentedthereinandextendtheresultsforapproximatealgorithmsincludingtheoriginalwhichhasanapproximationratioof1.Weprovidetheboundsontheimprovementfactorusinggraphtheoreticalresults.Werelatetheproductofweightedfractionalchromaticnumberandweightedindependencenumbertotheinnerproductoftheweightvectorsthroughaninequality.ThatinequalityplaysakeyroleinderivationofthedelayboundsforapproximateWMISscheduling.Usingthatrelationandhigherorderpotentialfunctionssuchascubicfunction,weendupachievingboundsforsumofhighermoments.Thosebounds,inturn,canbeusedtoboundsumofqueuevariances.Webelieveabetterunderstand-ingofthequeuevarianceatstationaritycanbeintegraltodesigninghigherefciencywirelessnetworkhardware. 1.4.3SingleParameterCharacterizationofDelayBoundsDelayboundsofgeneralizedMWSschedulingisstudiedin[ 15 ].Buildingontheirframe-workweintroduceanewboundforMWSpolicyandprovethattheweightselectionin[ 15 ]isoptimalforminimizingdelay,whenthequadraticpotentialfunctionisbeingutilizedintheanalysis.Wealsoproveournewboundistightthroughthesamemethodology.ThoseresultstogetherestablishaframeworktocharacterizeandassesstightnessonthedelayboundofMWSalgorithm:asingleparameteronefornativeMWSpolicyandamulti-parameteroneforthegeneralizedversionconsideredin[ 15 ]. 1.4.4GeneralizedMWSPolicyVariationsofMWSalgorithmcanprovidethroughputoptimalitywhilealsooptimizingparametersofchoice.Resultsgivenin[ 15 ]motivateustofurtherinvestigatealessconstrained 18

PAGE 19

generalizationofMWSpolicythroughwhichalargeclassoffunctionscanbeusedtocomputelinkweightsinMWSalgorithm.Werenethesufcientconditionsforstabilitythatwasgivenin[ 27 ].Asaresult,newrequirementsareaddedtotheprevioussufcientconditions.Weconsiderasetofexamplesandillustrateaspectrumoffunctionsmovingfrominadequacyforestablishingstabilitytotheabilitytoprovidestability. 19

PAGE 20

CHAPTER2COMPLEXITYOFACHIEVINGLOWDELAY 2.1BackgroundGivenaninterferencemodel,thischapterinvestigateshowtoschedulethetransmissionsofthelinkssothatsomechosenobjectivecanbeachieved.Forsimplicity,weconsideratime-slottedsystemwheretimeisdividedintoxed-sizedslots,thepacketsareallidenticalinsize,andeachlinkhasthecapacityoftransmittingonepacketinatimeslotifaccessisgranted.Undertheprotocolmodel,afeasiblescheduleisasubsetofthelinkssuchthatnotwolinksinthesetinterferewitheachother.Suchasetoflinkscanbeactivatedforsimultaneoustransmissiononthesametimeslot.Notethatascheduleisfeasibleifandonlyifitcorrespondstoanindependentsetoftheinterferencegraph.Acanonicalschedulingproblemisasfollows.Supposethearrivalratesoftrafctothelinksfallintothecapacityregion[ 41 ],which,looselyspeaking,isthesetofarrivalratesthatareschedulablebysomealgorithm.Findaschedulingpolicy,whichcanproduceafeasiblescheduleineachtimeslot,sothatthenetworkqueuesremainstable.Variousdenitionsofstabilityhavebeen(ormaybe)developed,suchasthequeuesbeingboundedortheMarkovianqueueprocessesbeingpositiverecurrent.Themotivationofthischapteristounderstandtheinherentcomplexityofthescheduling-for-stabilityproblem.Manyauthorshavestudiedvariousversionsofthisschedulingproblem.Theproblemappearstobeverydifcultingeneral.TassiulasandEphremidesshowedthatifaschedulecorrespondingtoaWMISisusedineachtimeslot,thenthenetworkofqueuescanbestabilizedforanyarrivalratevectorinsidethecapacityregion(inwhichcase,wesaythecapacityregionisachieved),foraclassofMarkovprocesses[ 52 ].FindinganWMISisaverydifcultprobleminitself(e.g.,stronglyNP-hardandcannotbeapproximatedbyaconstantratio).Thishasledtothebeliefthattheoverallschedulingproblemisdifcult.However,wehavefoundveryfewstudiesthataddresstheinherentdifcultyoftheoverallschedulingproblem(asopposedtotheone-stepsub-problemoneachtimeslot).Later,Tassiulasshowedthatthecapacityregion 20

PAGE 21

canbeachievedbyasimplerprobabilisticalgorithmrunningineachtimeslot[ 53 ].However,hedidcommentinformallythattheprobabilisticalgorithmtriestosolvetheWMISsub-problembutovermultipletimeslots.Thecommentwaslaterformalizedandgeneralizedin[ 57 ].Atthispoint,onemighthypothesizethattheindependentsetsub-problemisunavoidableandanyschedulingpolicythatemploysasimpleone-step(per-time-slot)algorithmmustpayfortheinherentcostofschedulinginotherways.Mostlikely,thecostismerelyamortizedovertime,andthismayleadtopoortransientperformanceofthepolicy,suchaslargequeuesorslownessinreachingsystemsteadystate.Recentdevelopmenthasmadeitevenmorerelevanttostudytheinherentcomplexityofscheduling.Muchofthelatestworkconcentratesondevisingevensimpleranddecentralizedone-stepalgorithmsthatstabilizethefullcapacityregion[ 46 ][ 9 ][ 42 ][ 6 ][ 20 ][ 37 ][ 45 ].However,ifthecomplexityoftheschedulingproblemisirreducible,thenonemaywonderwherethehiddencostisinthesealgorithms.Arelatedstreamofwirelessschedulingresearchconcentratesondevisingsimpleone-stepalgorithmsthatareguaranteedtoachievesomesubsetsofthecapacityregion,foranarbitrarynetwork[ 35 ][ 34 ][ 25 ][ 56 ][ 24 ][ 22 ].Again,acomplexityquestionexistshere:Istheschedulingprobleminherentlyeasyoneachofthesesubsets?In[ 8 ][ 58 ][ 5 ],theauthorsshowedthat,iftheinterferencegraphsatisesaconditioncalledlocalpooling,thentheLongestQueueFirst(LQF)policy(whichissimple)achievesthefullcapacityregion.But,itisunknownwhethertheschedulingproblemforthisspecialclassofnetworksiseasyornot.In[ 24 ],theauthorsdenedthe-localpoolingfactorforanarbitrarynetwork,where0<1,andshowedthatLQFachieves,whereisthecapacityregion.Severalstudieshaveimplicitlyaddressedsimilarconcernsofthischapter.Inparticular,theauthorsof[ 57 ]examinedthetradeoffsamongthecomplexityoftheone-stepalgorithm,theachievablerateregionandthedelay.There,theone-stepalgorithmwasassumedtobetheWMISscheduleoritsapproximations.Thedelayandcomplexityrelationshipwasalsobrieysurveyed.Otherworksthatexaminedthedelayissueofschedulinginclude[ 44 ],[ 43 ],[ 48 ].Theyinfactexaminedthequeuesizesinsteadoftheactualdelays.Theupperboundofthequeue 21

PAGE 22

sizesdeterminesthedelayproperty.Somelineartrade-offswereobservedbetweendelayandcomplexity.Thefollowingisasummaryofourmainresultsandcontributions.Westartwithformulatingthecapacityregionusingthefractionalcoloringproblem.Althoughin[ 17 ]asimilarstudyisconducted,weconcentrateonahardersettingwherea(maximal)feasiblescheduleateachstepisamaximalindependentsetoftheinterferencegraph.Wenextconsiderthescheduling-for-stabilityproblemunderdeterministicarrivalsandstudyitscomplexity.Weincorporatethequeuesizeboundintothecomplexityresults.Weshowthat,ingeneral,itisimpossibletohavesmallqueuesizesusingsimpleone-stepalgorithmsunlessZPP=NP.Tothebestofourknowledge,thisisoneofthefewformalconnectionstothetheoryofcomputationalcomplexityforthestabilityproblems.Wealsousetheapproachin[ 52 ]todemonstratethatanalgorithmthatusesanWMISsolverasaone-stepsubroutineisstorage-wiseefcient.Wethenexaminethegeometryofthecapacityregion.Finally,werelatethefractionalcoloringproblemandthenewinterpretationofthecapacityregiontothenotionofset-localpooling.Webelievethattheresultsofthischapterareusefulforcontinuedinvestigationontheinherentcomplexityofwirelessscheduling.Thischapterisstructuredasfollows.InSection 2.2 ,weprovidethenotationanddenitionsofrelatedproblems.InSection 2.3 ,wedenethecapacityregionintermsofthefractionalcoloringproblem,andhence,connectschedulingtocoloring.InSection 2.4 ,weanalyzethecomplexityofefcientlystabilizinganetworkunderdeterministicarrivals.WeshowthattheWMIS-basedalgorithmisstorage-wiseefcientinSection 2.5 .InSection 2.6 ,weinvestigatedifferentinterestingsectionsofthecapacityregion. 2.2PreliminariesWeconsiderawirelessnetworkwithone-hoptrafc.Thatis,anydataistransmittedonlyonce,andafterthat,itleavesthenetwork.WemodelthewirelesslinksetL'sinterferencerelationswithaninterferencegraph,G=(V;E),whereeachlinkl2Lintheoriginalnetworkcorrespondstoanodev2V.Twonodesv1;v22VareconnectedinG,ifandonlyifthecorrespondinglinksintheoriginalnetworkinterferewitheachother.Weassumeasymmetric 22

PAGE 23

interferencerelationthusGisundirected.Afeasiblescheduleisdenedtobeasetofnon-interferingnodesinG.Amaximalscheduleisafeasibleschedulethatcannotincludeanymorenodeswithoutviolatingthefeasibilityconstraint.AfeasiblescheduleinanetworkcorrespondstoanindependentsetinG.ThesetofallfeasibleschedulesisdenotedbyIL;thissetcanalsobeinterpretedasallindependentsetsofthegraphG.WedenotethesetofallmaximalschedulesbyML,andusem2MLtodenoteasinglemaximalschedule.Throughoutthischapter,weusethewordscheduletorefertoafeasiblemaximalschedule.Whenapplicable,MLisregardedasajVjjMLj0-1matrix.Eachcolumnofthematrixisa0-1vectorrepresentationofamaximalindependentsetofG,with1indicatingthatthecorrespondingnode(inG)isselected(andthelinkisactiveintheschedule)and0otherwise.Acolumnvectorthathas1'sinallentriesisdenotedby~e.Weadopttheconventionaldenitionofthecapacityregion:=fj0forsome2Co(ML)g:Theinteriorofthecapacityregionisdenedas:o=fj0
PAGE 24

westepbackalittleanddeneamorebasicversionoftheschedulingproblem.Itisourbeliefthatthisbasicversioncontainsessentialelementsaboutthecomplexityofvariousstabilityproblems.Wenowdeneastaticversionoftheschedulingproblem.Givenanarrivalratevector,theschedulingproblemistondaconvexcombinationoftheschedulessuchthattheresultingserviceratedominatesthearrivalrate.Formally,givena(rational)arrivalratevector,ndanon-negativetimesharevectorsuchthatMLandPjMLji=1i1(fasterthanunittime).Ourconjectureisthatthecomplexityofvariousscheduling-for-stabilityproblemsisdeterminedbytheshapeoftheconvexpolytopeoftheabovestaticfeasibilityproblem.Forinstance,ifisstrictlyinsidethecapacityregionandifwecannda(servicerate)vectorthatstrictlydominatesandsatisestheabovefeasibilityproblem,thenthearrivaltrafccanbestabilizedfornearlyalldenitionsofstabilityencounteredintheliterature.Hence,thefocusofthechapterisaboutthisstaticproblemanditsvariants.Weusetodenoteasolutiontotheaboveschedulingproblem.TheproblemandsolutionsbothdependonthegraphG.Here,wesuppressthedependenceonGinthenotation. 2.2.2FractionalColoringProblemThechromaticnumberofagraphG,(G),istheminimumnumberofcolorstopainttheverticessothattheconnectednodesdonotsharethesamecolor.Duetoitsrelationtothevertexcoveringproblems,thechromaticnumberproblemhasbeenextensivelystudiedinthegraphtheoryeld.ThechromaticnumberproblemisNP-complete.Itcanberepresentedwithanintegerprogramming(IP)formulation.Alinearprogramming(LP)relaxationofthisIP,whichisknownasthefractionalcoloringproblem,isalsocentraltothefractionalpackingandcoveringproblemsingraphs.Here,weprovideaformulationfortheweightedfractionalcoloringproblem. 24

PAGE 25

Denition1. Letw2QjVjbea(component-wise)non-negativeweightvectorandconsiderthefollowingoptimizationproblem:f(G;w)=minimizeXm2MLmsubjecttoMLw0:Theoptimalvaluef(G;w)iscalledtheweightedfractionalchromaticnumber.Weletw2QjVjdenoteasolutiontothefractionalcoloringproblem.Theweightedfractionalchromaticnumberofagiveninterferencegraphinthewirelessschedulingcontextcanbeinterpretedasthefastestwayofservingqueueddata(withoutaddi-tionalarrivals)intheamountgivenbywifwecanhopbetweenschedulesininnitesimaltimeslots.In[ 17 ],asimilarproblemisstudiedforthe1-hopinterferencemodel,wheretheschedulesaremaximalmatchings.Theauthorsshowedthat,underthatsetting,theschedulingproblemadmitspolynomialalgorithms.In[ 17 ],threeinterpretationsoftheoptimizationproblemaredis-cussed.Althoughthisworkconcentratesontheindependentsetpolytope,therstinterpretationin[ 17 ]andourperspectivearecloselyrelated. 2.2.3MinimumSpanPacketSchedulingGivenaninterferencegraphGandinitialqueuesizesatallthelinks,denotedbythevectorw,wedenetheminimumspanpacketscheduling(MSPS)problemtobethesmallestnumberoftimeslotstodrainthepacketsfromthequeuesundertheinterferenceconstraintsandunittransmissioncapacityateachlink.ThefollowingisanIPformulationofthisproblem: 25

PAGE 26

\(G;w)=minimizeXm2MLcmsubjecttoMLcwcm2Z+;whereZ+denotesnon-negativeintegers.Now,considerreningthetimeslotswhilekeepingtheinitialamountofqueueddataxed.Whenreducingthesizeofthetimeslots,wealsomaketheunitpacketlengthsmaller.IfwechoosethetimeslottobeRtimessmaller,thentheinitialqueuesizeswillbecomeRtimeslargerintermsofpacketcount.Theoptimizationproblemcanbewrittenasfollows.\(G;w;R)=minimizeXm2MLcmsubjecttoMLcRwcm2Z+:Thefollowinglemmaindicatesifwemakethetimeslotsizeinnitesimallysmallandscaletheoptimalvaluedownbythesamescalingfactor,thesolutionoftheaboveproblemwillconvergetothesolutionofthefractionalcoloringproblem. Lemma1. limR!1\(G;w;R) R=f(G;w). Proof. Letwbeasolutioncorrespondingtof(G;w).Sincew2QjVjandisnon-negative,wecanconcludew2QjMLjandisnon-negative.Wedenotethereducedformofeachrationalnumberwibyai biforsomeai;bi2Z+;bi6=0,whereaiandbiarerelativelyprime.Lettheleastcommonmultipleofthebi'sbeB.Similarly,denotethereducedformofeachwibyci diforsomeci;di2Z+;di6=0,wherecianddiarerelativelyprime.Denotetheleastcommonmultipleofthedi'sbyD.Foranyj2Z+andj1,theoptimizationproblemforderivingf(G;jBDw) 26

PAGE 27

hasanintegeroptimalsolution,andhence,f(G;jBDw)=\(G;w;jBD).NowsupposekBDR<(k+1)BDforsomek2Z+.Notethat \(G;w;kBD)\(G;w;R)\(G;w;(k+1)BD):Hence, f(G;kBDw)\(G;w;R)f(G;(k+1)BDw):DividingeachtermabovebyRandusingthelinearityoff(G;w)inw,wehave f(G;kBDw R)\(G;w;R) Rf(G;(k+1)BDw R):Usingthecontinuityoff(G;w)inw,asR!1,bothf(G;kBDw R)andf(G;(k+1)BDw R)approachf(G;w). Asimilarresultforthenon-weightedcasewasderivedin[ 28 ]wheretheauthorsstudiedhowtominimizethespectrumusageinFDMAnetworks. 2.3RelevanceofFractionalColoringtoSchedulingUndertheprotocolmodel,theschedulingproblemassumesastrongcombinatorialstructure.Inthissection,wewillbridgeschedulingwiththefractionalcoloringproblemtohighlightitscombinatorialnature.Theresultswillbeusefulforthefollowingsections.Theorem 2.1 statesthatthecapacityregioncanbedenedintermsoftheweightedfractionalchromaticnumber,wheretheweightvectoristhearrivalratesonthelinks. Theorem2.1. =fj0f(G;)1g: Proof. Takeanyarrivalvector2.Bydenition,thereexists2Co(ML)suchthat.Since2Co(ML),thereexistssuchthatML=,where0andPjMLji=1i=1.Hence,isafeasiblesolutiontotheoptimizationproblemforf(G;)anditscomponentssumto1.Therefore,f(G;)1.Forthereversedirection,takeanarrivalvectorsuchthat0f(G;)1.Assumeisasolutioncorrespondingtof(G;).Thatis,MLand 27

PAGE 28

B=PjMLji=1i1.IfB=0,then=0,whichisin.IfB6=0,thenML BandML B2Co(ML).Thisindicates2. BytheproofforTheorem 2.1 ,weconcludethatasolutiontothefractionalcoloringproblemcanbethoughtasatime-sharingvectorofschedulesfortheschedulingproblem.Similarly,wecandenetheboundaryofthecapacityregionintermsoftheweightedfractionalchromaticnumber.But,weneedatechnicalfactrst.Fact1ForanyJL,thereexists2Co(ML)suchthatj>0forallj2J.Theproofissimpleandisomitted.Next, Theorem2.2. 2()]TJ /F7 11.955 Tf 11.96 0 Td[(o)()f(G;)=1. Proof. Werstshowif2()]TJ /F7 11.955 Tf 12.14 0 Td[(o),thenf(G;)=1.SincebyTheorem 2.1 ,f(G;)1forall2,weonlyneedtoshowwhen0f(G;)<1,2o.Iff(G;)=0,then=0andhence2o.Considerthecase0jjforj=2J.Now,takeany^2Co(ML)withtheproperty^j>0forj=2J.ThisispossiblebyFact1.ConstructanewvectorinCo(ML):~=(1)]TJ /F3 11.955 Tf 12.29 0 Td[()+^,where0<<1.Whenissufcientlysmall,wecanhave~>0and~>.Hence,2o.Next,supposef(G;)=1.ByTheorem 2.1 ,2.Supposeisino.Then,thereexists2Co(ML)suchthat>.Thisinturnimpliesthatthereexists>0suchthat>(1+),or1 1+>.SupposesatisesML=,where0andPjMLji=1i=1.Then,1 1+isafeasiblesolutiontotheoptimizationproblemforndingf(G;),butitscomponentssumtolessthanone.Thiscontradictstheassumptionf(G;)=1.Therefore=2o. Corollary1. o=fj0f(G;)<1g. 28

PAGE 29

Aninterpretationofthetheoremsandcorollaryaboveisasfollows.Thecapacityregionconsistsofvectorsthatcanbeservicedinunittimeorlessbysomemixingofschedules.Theinteriorcontainsvectorsthatareschedulableinstrictlylessthanunittime.Theboundaryofthecapacityregion,)]TJ /F7 11.955 Tf 12.64 0 Td[(o,consistsofvectorsthatleavethecapacitypolytopewhenmultipliedbyascalaroftheform1+forany>0.Thefollowingformalizesthisinterpretationoftheboundary. Lemma2. 2()]TJ /F7 11.955 Tf 11.95 0 Td[(o)()2and(1+)=2forall>0. Proof. ByTheorem 2.2 ,2()]TJ /F7 11.955 Tf 12.26 0 Td[(o)impliesf(G;)=1.Suppose(1+)2forsome>0.Bythelinearityoff(G;w)inw,f(G;(1+))=(1+)f(G;).ByTheorem 2.1 ,0(1+)f(G;)1,whichimpliesf(G;)1 1+,contradictingtheassumption.Conversely,suppose2andthereisno>0suchthat(1+)2.Suppose2o.Then,thereexists2Co(ML)suchthat>.Hence,thereexists>0suchthat>(1+).Then,(1+)2,whichisacontradiction. Avectorinsidethecapacityregionbecomesincreasinglydifculttoserviceasitapproachestheboundary.Typicalsuboptimalschedulingalgorithmsintheliteratureachieveapartofthecapacityregionderivedbyscalingthecapacityregionuniformlybysomeless-than-1constant(e.g.,[ 24 35 56 ]).Motivatedbythisfact,weassociateanumberwitheveryarrivalratevector. Denition2. Foranynon-negativeratevector,let ,supf j 2g:(2)Notethat =1ifandonlyifisontheboundaryofthecapacityregion.If2,alargervalueof indicatesisdeeperinsidethecapacityregion.Wenextdemonstratetherelationshipbetweenthisnumberandtheweightedfractionalchromaticnumber. Lemma3. Foranon-negativevector,f(G;)=1 Proof. Bylinearity,wehavef(G; )= f(G;).ByLemma 2 isontheboundaryofthecapacityregion.ByTheorem 2.2 ,f(G; )=1.Hence, f(G;)=1. 29

PAGE 30

Thisresultassociatesthefractionalcoloringproblemwiththedifcultyofservicinganarrivalratevector.Notethatasolutiontothefractionalcoloringproblemisalsoasolutiontotheschedulingproblem(whichisafeasibilityproblem)butnotviceversa.Accordingly,weturntothequestionofhowscheduling(andhopefully,achievingstability)andtheweightedfractionalcoloringproblemarerelatedwhenanarrivalratevectorisclosetheboundary.Thefollowinglemmaindicatesthatasolutiontotheschedulingproblemapproachesasolutiontothefractionalcoloringproblemastheratevectorgetssufcientlyclosetotheboundary. Lemma4. Foranarrivalvector2,any(seethedenitionofthisnotationinSection 2.2.1 )isa approximationtotheoptimizationproblemforf(G;). Proof. ByLemma 3 ,f(G;)=1 .Nowtakeanysolutionoftheschedulingproblem,,bydenitionisafeasiblesolutiontotheproblemforndingf(G;)andPjMLji=1i1=f(G;) .Thisconcludestheproof. Oneimplicationisthatanalgorithmthatstabilizesanarrivalratevectoryieldsanapproxi-mationoftheweightedfractionalcoloringproblem2.Furthermore,theapproximationratiotendsto1asthearrivalratevectorgetsclosertotheboundary.Theresultestablishesapartialequiva-lencebetweenthetwoproblems.Theyarethesameforthevectorsnearorontheboundary,butarepotentiallyverydifferentforvectorsfurtherfromtheboundary.The(unweighted)fractionalcoloringnumberofagraphG,denotedbyf(G),isthesolutiontotheweightedversionwhentheweightvectorconsistsofallones,i.e.,f(G)=f(G;~e).Considertheproblemofachievingthemaximumidenticalrateonalllinks.Wedenote 2Strictlyspeaking,weonlyshowedthatthesolutiontothestaticschedulingproblemgivessuchanapproximation. 30

PAGE 31

thisrateby(G).ThefollowingistheLPformulationfornding(G).(G)=maximizezsubjecttoMLz~ez;0: Lemma5. (G)=1 f(G). Proof. ByLemma 2 andtheoptimality,(G)~eisontheboundaryof.ByTheorem 2.2 ,weconcludef(G;(G)~e)=1.Observef(G;(G)~e)=(G)f(G;~e)=1. 2.4ComplexityofSchedulingProblemInSection 2.3 ,weestablishedaconnectionbetweenthefractionalcoloringandtheschedul-ingproblems.Next,wewillusethederivedresultstorevealcomplexitycharacteristicsoftheschedulingproblemandstablealgorithms.Wewillfocusonadeterministicsystem,wherethearrivalspertimeslotareconstantforeverylink.ThearrivalratesareassumedtobeinQjLjandalllinkshaveunittransmissioncapacity.Astablealgorithminthissettingisdenedtobeanyalgorithmthat,giveninitiallyemptyqueues,keepsthequeuesizesboundedthroughouttheexecutionforallarrivalratesintheinteriorofthecapacityregion.Forastablealgorithm,theupperboundonthequeuesizesmightbeafunctionofthearrivalrateandtheencodinglengthoftheproblem.Later,wewillrelatetheupperboundtothedistancetotheboundaryforeacharrivalrate.Theexistenceofastablealgorithmimpliestheexistenceofavectorthatsolvesthewirelessschedulingproblem.However,givenasolutiontotheschedulingproblem,applyingitasanalgorithmateachtimeslotmightnotbestraightforward.Inadditiontothefrequenciesoftheschedules,whentoactivateeachschedulealsoplaysakeyrole.Additionally,althoughthetwoarerelated,thenotionofstabilityrequirestheconditionthatthesystemrunsforever.Todealwiththeseissuesandtocontinueourinvestigationofcomputationalcomplexity,wewill 31

PAGE 32

investigatestablealgorithmsthatnotonlyboundthequeuesizesbutalsoensurethattheboundispolynomialintheinput'sencodinglength.Beforewestart,wedeneaquantitycalledload-slackdenotedbyforeacharrivalratevector,whichisameasureofdistancetotheboundary.Itwillbeusedtocharacterizethequeuesizeupperboundofastablealgorithm. ,supfj+~e2g:(2) Denition3. Analgorithmstabilizesanarrivalratevectorcomputationallyefcientlyifthealgorithmisstableandateachiteration,theoperationofthealgorithmtakesapolynomialnumberofstepsinjLj. Denition4. AnalgorithmstabilizesanarrivalratevectorP-storage-wiseefcientlyifthealgorithmisstableandthequeuesizesareupperboundedbyP(jLj;())]TJ /F6 7.97 Tf 6.59 0 Td[(1),wherePisapolynomialfunction. Theorem2.3. ThereisnoalgorithmthatstabilizesallarrivalratevectorsinocomputationallyandP-storage-wiseefcientlyforanypolynomialP,unlessZPP=NP. Proof. Assumethereexistsanalgorithmthatstabilizesanydeterministicarrivalratevector2ocomputationallyandP-storage-wiseefciently.Then,allqueuesareupper-boundedbyP(jLj;())]TJ /F6 7.97 Tf 6.58 0 Td[(1)duringtheoperationofthenetwork.LetusdenotethisnumberwithPforshort.Also,eachtimeslottakesatmostP(jLj)computationsteps,forsomepolynomialfunctionP.Givenanapproximationfactor,wecanusethealgorithmasanoracletocreateafullypolynomialtimeapproximationscheme(FPTAS)forthefractionalcoloringproblemf(G).(Asimilartechniqueisusedfortheconcurrentowproblemin[ 1 ].)AssumewerunthealgorithmforRtimeslots.Then,thetotaltrafcinsertedintoqueueiisRiforeveryi.Ifallthequeuesareupper-boundedbyP,theamountofthetrafcservedatqueueiintheRtimeslotsisnolessthanRi)]TJ /F18 11.955 Tf 12.04 0 Td[(P.TheaverageservicerateatqueueiisatleastRi)]TJ /F19 7.97 Tf 6.59 0 Td[(P R.LettingR=P)]TJ /F6 7.97 Tf 6.59 0 Td[(1,weensurequeueiisservedatanaverageratei)]TJ /F3 11.955 Tf 11.5 0 Td[(~e.Therefore,wecanrunthestablealgorithm 32

PAGE 33

forP)]TJ /F6 7.97 Tf 6.59 0 Td[(1timeslotsandeitherachieveatimesharevectorsuchthatML)]TJ /F3 11.955 Tf 12.46 0 Td[(~eorconcludethatisnotinthecapacityregionthuscreatinganoracle.ByLemma 5 ,thearrivalratevector1 f(G)~eisontheboundaryofthecapacityregion.Bysettingtheprecisionto(6jLj))]TJ /F6 7.97 Tf 6.58 0 Td[(1,wecanconductbinarysearchandndanumberzsuchthatz1 f(G)z+(2jLj))]TJ /F6 7.97 Tf 6.58 0 Td[(1usingtheoracleforO(log()]TJ /F6 7.97 Tf 6.59 0 Td[(16jLj))times.Since1f(G)jLj,f(G)1=z(1+)f(G).Thus,thetotalrunningtimeofthealgorithmisP(jLj)O(log()]TJ /F6 7.97 Tf 6.58 0 Td[(16jLj))P(jLj;)]TJ /F6 7.97 Tf 6.59 0 Td[(16jLj))]TJ /F6 7.97 Tf 6.59 0 Td[(16jLj,whichispolynomialininputG.However,in[ 10 ],itisproventhatf(G)cannotbeapproximatedwithin(jLj1)]TJ /F5 7.97 Tf 6.58 0 Td[()forany>0,unlessZPP=NP.Therefore,devisingaFPTASherewouldequalizeNPandZPPaswell. ThefractionalcoloringproblemisknowntobeNP-complete.Grotscheletal.[ 13 ][ 14 ]establishedanequivalencebetweenthefractionalcoloringandWMISproblems.Formally,onecansolvefractionalcoloringbyusingasolverfortheWMISproblemapolynomialnumberoftimes.Additionally,anapproximatesolverfortheWMISproblemcanbeusedtoderiveanapproximationalgorithmforthefractionalcoloringproblem.TheapproximationratiooftheWMISsolverisdirectlycarriedtothelatterproblem'ssolution[ 18 ][ 19 ].Thisresult,however,shouldbementionedtogetherwiththefactthattheWMISproblemisNP-completeanddoesnotadmitapolynomialtimeapproximationalgorithmwithanapproximationratiojVjforsome>0[ 40 ]. 2.5QueueDynamicsunderDeterministicArrivalInthissection,wewilldemonstrateastorage-wiseefcientschedulingalgorithm:TheWMIS-basedalgorithmisP-storage-wiseefcientforP=3 2jLj())]TJ /F6 7.97 Tf 6.59 0 Td[(1+p jLj.Wedenotethequeuesizeoflinki2LattimetbyQi(t).ThearrivalanddepartureratesatlinkiandtimetareAi(t)andDi(t),respectively.NotethatAi(t)=iandDi(t)=min(1;Qi(t)+Ai(t))ui(t),whereui(t)isanindicatorfunction,equalto1iflinkiisscheduledfortransmissionattimetand0otherwise.Thesystempotentialisdenedtobethesumofallqueuesizesquares,(Q)=PjLj1Q2i.Wewillshowwhenthequeuesizesgrowbeyondathreshold,thesystempotentialdropsinthe 33

PAGE 34

nexttimeslot.Thisinturnprovesthatthesystempotentialisboundedandconsequentlythequeuesizesareboundedaswell.Thequeueevolutionisgivenby:Qi(t+1)=Qi(t)+Ai(t))]TJ /F3 11.955 Tf 11.95 0 Td[(Di(t):Thepotentialdifferencebetweentimestandt+1isdenotedby(t),(Q(t+1)))]TJ /F7 11.955 Tf 11.8 0 Td[((Q(t)).Then,(t)=Xi2LQ2i(t+1))]TJ /F12 11.955 Tf 11.96 11.36 Td[(Xi2LQ2i(t)=Xi2L[Qi(t+1)+Qi(t)][Qi(t+1))]TJ /F3 11.955 Tf 11.96 0 Td[(Qi(t)]=Xi2L[2Qi(t)+Ai(t))]TJ /F3 11.955 Tf 11.96 0 Td[(Di(t)][Ai(t))]TJ /F3 11.955 Tf 11.95 0 Td[(Di(t)]=Xi2L[2Qi(t)(Ai(t))]TJ /F3 11.955 Tf 11.96 0 Td[(Di(t))]+Xi2L[Ai(t))]TJ /F3 11.955 Tf 11.95 0 Td[(Di(t)]2:Thesecondtermaboveisbounded:Pi2L[Ai(t))]TJ /F3 11.955 Tf 12.37 0 Td[(Di(t)]2jLj.Wenextinvestigatetherstterm.Xi2L[2Qi(t)(Ai(t))]TJ /F3 11.955 Tf 11.95 0 Td[(Di(t))]=2Xi2L[Qi(t)Ai(t)])]TJ /F7 11.955 Tf 11.96 0 Td[(2Xi2L[Qi(t)Di(t)]:ThedeparturevectorD(t)=(D1(t);D2(t);:::;DjLj)Tisdeterminedbythechosenscheduleattimet,m(t)=argmaxm2MLQ(t)Tm.Notethatthedepartureisdominatedbythebinaryrepresentationofthechosenschedule,D(t)m(t). Lemma6. 0Q(t)Tm(t))]TJ /F3 11.955 Tf 11.96 0 Td[(Q(t)TD(t)jLj. Proof. SinceQ(t);m(t);D(t)0andD(t)m(t),Q(t)TD(t)Q(t)Tm(t).Asusual,wedenotethei-thcomponentofm(t)bymi(t).RecallDi(t)=min(1;Qi(t)+Ai(t))ui(t)= 34

PAGE 35

min(1;Qi(t)+Ai(t))mi(t).Hence,Qi(t)mi(t))]TJ /F3 11.955 Tf 11.96 0 Td[(Qi(t)Di(t)=Qi(t)mi(t))]TJ /F3 11.955 Tf 11.96 0 Td[(Qi(t)min(1;Qi(t)+Ai(t))mi(t):WhenQi(t)=0orQi(t)1ormi(t)=0,wehaveQi(t)mi(t))]TJ /F3 11.955 Tf 12.29 0 Td[(Qi(t)Di(t)=0.Theonlyinterestingcaseiswhen00and2Co(ML)suchthatA(t)+~e.SinceQ(t)0,Q(t)TA(t)+Q(t)T~eQ(t)T.Moreover,Q(t)Tm(t)Q(t)Tforall2Co(ML).Therefore,Q(t)TA(t)+Q(t)T~eQ(t)Tm(t).Xi2L[2Qi(t)(Ai(t))]TJ /F3 11.955 Tf 11.95 0 Td[(Di(t))]=2Xi2L[Qi(t)Ai(t)])]TJ /F7 11.955 Tf 11.95 0 Td[(2Xi2L[Qi(t)Di(t)]2[Q(t)Tm(t))]TJ /F3 11.955 Tf 11.96 0 Td[(Q(t)T~e)]TJ /F7 11.955 Tf 11.95 0 Td[((Q(t)Tm(t))-222(jLj)])]TJ /F7 11.955 Tf 26.9 0 Td[(2Q(t)T~e+2jLj:Hence,thepotentialdifferencecanbeboundedas(t))]TJ /F7 11.955 Tf 23.91 0 Td[(2Q(t)T~e+3jLj:Fortimet,wedenotethetotalqueuesumasQ(t)=Pi2LQi(t).WhenQ(t)>3jLj 2,(t)<0.Also,if(Q(t))>(3jLj 2)2,thenQ(t)>3jLj 2.Foranytimet,themaximumpotential 35

PAGE 36

increaseisboundedby2Q(t)+jLj(whichiswheneachqueueisincreasedby1).Ifthepotentialismorethan(3jLj 2)2atthebeginningoftimet,itwilldropafteronetimeslot.Themaximumpotentialthatthesystemcaneverreachisboundedby(3jLj 2)2+23jLj 2+jLj(3jLj 2+p jLj)2.Aboundedpotentialtranslatesintoboundedqueuesizes.Themaximumqueuesizeduringtheoperationofthesystemisboundedby3jLj 2+p jLj.Inconclusion,byusingasolverfortheWMISproblem,wecanachieveastorage-wiseefcientalgorithm. 2.6ACloserLooktotheCapacityRegionunderProtocolModelInthissection,wewillcharacterizethedifferentsectionsofthecapacityregionandprovidesamplearrivalratevectorsthatarerepresentativeofeachsection.Ingeneral,thecapacityregionconsistsofarrivalratevectorsthatcanbeservedbysomealgorithm.Theboundarycanbeconsideredasthesetofarrivalratevectorsthatarethemostchallengingtoserve.Inearliersections,weprovidetwocharacterizationsoftheboundary.Therstonerelatestothefractionalchromaticnumberandthesecondonestatesthatthevectorsontheboundarycannotbeextendedfurtherandyetremaininthecapacityregion.Bytheseconddenition,ifavectorisontheboundary,noothervectorinthecapacityregioncandominateitinitsdirection.However,itispossiblethat,fortwovectorsontheboundary,onedominatestheother.Therefore,itisworthtosingleoutthepartoftheboundarywhereavectorisnotdominatedbyanyothervectorinthecapacityregion.Thispartcontainsthemostchallengingarrivalratevectorstoserve.Wecallthisregiontheupper-boundaryofthecapacityregion,anddenoteitbyU().Thedenitionoftheupper-boundaryexcludesthosevectorsonthetraditionalboundarythataredominatedbysomevectorsinthecapacityregion.Formally,U()=f2jTforallT2andT6=g:Thenotationmeansnot. 36

PAGE 37

~M1T=(1,0,1,0,1,0) ~M2T=(0,1,0,1,0,1) ~M3T=(1,0,0,1,0,0) ~M4T=(0,1,0,0,1,0) ~M5T=(0,0,1,0,0,1) Figure2-1. Interferencegraphforasix-cyclenetworkanditsschedulevectors. Toillustratethesignicanceofdifferentregions,wewilluseasmallnetworkandexamineafewinterestingarrivalratevectors.Considerthe6-cyclenetworkwhoseinterferencegraphisgiveninFigure 2-1 .Theschedulesofthisnetworkarelistedonitsleft.Consideranarrivalratevectorontheupper-boundary,(1 2;1 2;1 2;1 2;1 2;1 2)T.Thisvectorcanberepresentedasaconvexcombinationoftheschedules,1 2~M1+1 2~M2.Itisimpossibletodominatethisvectorbyanyotherconvexcombinationoftheschedules.Next,considertheratevector(1 2;1 2;1 2;1 2;1 2;0)T.Itisontheboundary.However,thepreviousvectordominatesthisone.Therefore,thisvectorsitsontheboundarybutnotontheupper-boundary.Nowmoveinsidethecapacityregionandconsidertheratevector(1 3;1 3;1 3;1 3;1 3;1 3)T.Wecanprovidetwodifferentrepresentationsofthisvector:1 3~M1+1 3~M2or1 2~M3+1 2~M4+1 2~M5.Notethattheformerrepresentationismoreefcientthanthelattersincethecoefcientshaveasmallersum.Thisvectorisintheconvexhullofthe(maximal)schedulesbutnotontheboundary.InFigure 2-2 ,weprovideasimpliedillustrationofacapacityregiontodemonstratetheconcepts.Thesolidthickerlineindicatestheupper-boundaryandthedashedlineisfortherestoftheboundary.1isontheboundarybutalsodominatedbysomevectorinthecapacityregion.Therefore,1isnotontheupper-boundary.3cannotbedominatedbyanyothervectorinthecapacityregionanditisontheupper-boundary.2isaconvexcombinationoftheschedules,butitcanstillbescaledupwithoutleavingthecapacityregion;hence,itisbothinCo(ML)andin 37

PAGE 38

Figure2-2. Capacityregionandarrivalratevectorsindifferentsections. theinteriorofthecapacityregion.4isalsointheinteriorofthecapacityregionbutitisnotinCo(ML).4hasrelativelylowarrivalrates;itmaybeeasytoschedulesuchanarrivalratevector. 2.7CapacityRegionandLocalPoolingTheshapeofthecapacityregioncanaffectthedifcultyoftheschedulingproblem.Sincealltheothervectorsinthecapacityregionaredominatedbytheupper-boundary,aninvestigationoftheupper-boundarycanbecrucial.Foranyarrivalratevector,wecandeneanorm-likefunctionwithrespecttothecapacityregion,whichwewilllooselycallschedulenorm. Denition5. GivenalinksetL,theschedulenormofanarrivalratevector,denotedjjjjs,isgivenbyjjjjs=inffkj92Co(ML);kg:Here,isinterpretedasaserviceratevector.Ifthelinksareendowedwithqueueddataproportionaltoandtherearenofurtherarrivals,thenjjjjsistheshortesttimetodrainalldataamongallpossibleschedules.Wenextshowthattheschedulenormisequaltotheweightedfractionalchromaticnumber. 38

PAGE 39

Lemma7. Foranyarrivalratevector,jjjjs=f(G;). Proof. Thecaseof=0istrivial.Weonlyconsiderthecasewhere6=0.SinceCo(ML)isaclosedset,itcanbeshowneasilythesetinDenition 5 ,fkj92Co(ML);kg,isclosed,andtheinmumisachieved.Hence,foranyarrivalratevector,thereisa2Co(ML)suchthatjjjjs.Since2Co(ML),=MLforsome0andPjMLji=1i=1.Leti=jjjjsiforeachi;isfeasiblefortheoptimizationproblemtondf(G;).Hence,jjjjsf(G;).Toshowtheotherinequality,welet=ML f(G;),whereisanoptimalsolutiontotheproblemofndingf(G;).Then2Co(ML)andf(G;).Therefore,jjjjsf(G;). Corollary2. Thecapacityregionisthesetcontainingallthevectorswhoseschedulenormsarelessthanorequalto1.Theinteriorofthecapacityregionisthesetcontainingallthevectorswhoseschedulenormsarelessthan1.In[ 34 ],thelengthofavectorrespecttoaconvexregionisdenedas:jjxjj=1 supfkjk0;kx2g: (2) Lemma8. ThedenitionoftheschedulenormisidenticaltothedenitioninEquation 2 ,ifwelet=,i.e:jjxjjs=jjxjj: Proof. Notethatjjxjj=1= xbyDenition 2 .ByapplyingLemma 3 andLemma 7 ,weobtainthedesiredequality. Set-localpoolingisstudiedin[ 32 ].Ithasmanyinterestingpropertiesandisrelatedto-localpoolingdenedin[ 24 ].Wewillshowtherelationshipbetweentheset-localpoolingfactorandtheschedulenorm.LetLbeanynon-emptylinkset.Forconvenience,letL=fjLL;forallL;L2Co(ML)g: 39

PAGE 40

ThecomplimentofLiscL=fjL>L;forsomeL;L2Co(ML)g: Denition6. Givenanon-emptysetoflinksL,wesayLhasaset-localpoolingfactorLifthefollowingholds.L:=supfj2Lg:=inffj2cLg:Ithasbeenshownthattheset-localpoolingfactorisequaltotheoptimalsolutionofthefollowingproblem.min;L;LsubjecttoLLL;L2Co(ML):LetKL.ForanyjKj-dimensionalvector,wecandeneanextendedjLj-dimensionalvectorLbysettingalltheextendedcomponentstobe0.Conversely,foranyjLj-dimensionalvector,wecancreateajKj-dimensionalvectorbyrestriction,andwedenoteitby[]K. Lemma9. Theset-localpoolingfactorsatises:K=minfjjLjjsj2Co(MK)g. Proof. Forany2Co(MK),0
PAGE 41

Thelemmashowsthattheset-localpoolingfactorisequaltotheminimumschedulenormofallvectorsinCo(MK).Theset-localpoolingfactorforsetKcontainsinformationabouttheshapeofCo(MK),asillustratedbyCorollary 3 Corollary3. WhenL=1,any2Co(ML)satisesjjjjs=1. Lemma10. ForagivensetLandanytwovector1;22Co(ML),wehaveL(jj1jjs+jj2jjs)jj1+2jjs. Proof. Supposethereexisttwovectors1;22Co(ML)andL(jj1jjs+jj2jjs)>jj1+2jjs.Then,2L>jj1+2jjsorL>jj1=21+1=22jjs.Since1=21+1=222Co(ML),thiscontradictsLemma 9 Lemma11. Theupper-boundarysatisesUCo(ML). Proof. BythedenitionofU,any2Usatises2.Then,forsome2Co(ML):Thus,=,whichimpliesthat2Co(ML)andthelemmaholds. Lemma12. Theupper-boundarysatisesU=Co(ML)ifandonlyifL=1. Proof. First,wewillprovethatifU=Co(ML)thenL=1.SupposeL<1.Thereexisttwovectors;2Co(ML)suchthatL.Then,62U,whichisacontradiction.Next,wewillshowthatifL=1,thenU=Co(ML).FromLemma 11 ,wealreadyknowthatUCo(ML).WeonlyneedtoshowCo(ML)U.SupposeCo(ML)6U,thenthereexistsa2Co(ML)and62U.Since2,theremustbea2Co(ML)suchthatand6=.Then,thereisapartitionofsetLinto,sayPandQ,suchthatP\Q=;,P[Q=Land[]Q=[]Q;[]P>[]P.IfQ=;,thentheremustbea<1suchthat.ThiscontradictsL=1.SupposeQ6=;.Ifforalli2Q,i=i=0,thenagain,theremustbea<1suchthat.Wenextconsidertheremainingcase.Pickani2Qsuchthati=i>0.Forthisi,theremustbeaschedulesinwhichlinkiisnotactivated.Wechooseasmallenoughandlet0=(1)]TJ /F3 11.955 Tf 12.22 0 Td[()+s.Since;s2Co(ML),wehave02Co(ML).Next,setP P[figand 41

PAGE 42

Q Q)-175(fig.ThisoperationcanberepeateduntileitherQ=;or[]Q=[0]Q=0.Then,theremustbea<1suchthat0,whichcontradictsL=1. ThislemmaimpliesthatwheneverL=1,theupper-boundaryisconvexandequaltoCo(ML).Then,anyrayfromtheorigincanonlyintersectCo(ML)atonepoint. 2.8EffectsofRedeningLoad-slackofArrivalVectorsInthissectionwederiveresultsforamodiedload-slackdenition.Thenewdenitionprovidesamoreintuitiveperspective,asitcomesclosertorepresentingthegapbetweenthenormalizedservicerateandthesystemloadincorrespondencetotheinterferencegraph.Westartwithjointlydeningtheconceptsofstorageefciencyandtheload-slack,().Weusearrivalvectorfortheaveragearrivalratevectorthroughoutthesection. Denition7. AschedulingalgorithmhasstorageefciencyforaveragearrivalratevectorifitsatisesthefollowinginequalitywiththeinitialconditionQ(0)=0:E[Q(t)Te]P(N;())]TJ /F6 7.97 Tf 6.58 0 Td[(1)8t20;1;::: (2)where()=1)]TJ /F3 11.955 Tf 12.24 0 Td[(f(G;),Pisapolynomiallyboundedfunction,andNistheencodinglengthoftheinterferencegraph(equivalentlyprobleminput). Denition8. Wecallanarrivalratevectoruniformifitisintheformk~efork>0.Thefollowinglemmawillprovideitselfusefulinderivingapproximationbounds: Lemma13. If00then1 zisan-approximationtothef(G),i.e.:f(G)1 z(1+)f(G). Proof. 1 f(G)z+kf(G)1 z+k(1+)f(G)(1+) z+k(1+) z+z=1 z: 42

PAGE 43

Nowwearereadytostatethemaintheoremwiththenewdenitionoftheload-slack: Theorem2.4. ThereisnoschedulingalgorithmthathasstorageefciencyandcomputationefciencyforuniformarrivalvectorsunlessP=NP. Proof. AssumewearegivenagraphGandanapproximationratio2(0;1].AtransientqueueboundP=P(N;(0))]TJ /F6 7.97 Tf 6.59 0 Td[(1)ondeterministicarrivals,where0= 4jLj,canbeusedtocompute-approximationtothefractionalchromaticnumber,Xf(G),asfollows.RuntheefcientschedulingalgorithmforP)]TJ /F6 7.97 Tf 6.59 0 Td[(1roundswhilekeepingaveragearrivalrateuniformamongqueues.TherearetwopossibilitieseitherallqueueswillstaybelowPthroughouttheexecutionoratleastonequeuepassesthethreshold.Nextwegoovereachcaseindetail:Case1:Qi(t)P8i;t:ThequeueiwillhaveP)]TJ /F6 7.97 Tf 6.58 0 Td[(1iamountofowinsertedandhasatleastservedtheamountP)]TJ /F6 7.97 Tf 6.59 0 Td[(1i)]TJ /F3 11.955 Tf 12.61 0 Td[(P.Ifyourunthealgorithmwithidenticalarrivalsatqueuesi.e.=ke;k>0andqueuesdonotexceedPduringP)]TJ /F6 7.97 Tf 6.59 0 Td[(1rounds;theaveragescheduleactivationswillformavector,,thatalsocouldbeinterpretedastime-sharingfactorofschedulessuchthatML()]TJ /F3 11.955 Tf 11.49 0 Td[()~e.Whichinreturnguarantees()]TJ /F3 11.955 Tf 11.49 0 Td[()~eisinthecapacityregionor)]TJ /F3 11.955 Tf 11.49 0 Td[(1 f(G).Case2:Qi(t)>P9i;t:IfanytimeduringtheprocedurethequeuesizesaretoexceedPitimplies1 f(G))]TJ /F3 11.955 Tf 11.38 0 Td[(0orotherwisethequeuebound,P,shouldhavebeenobserved.Bydiscretizingtheinterval(1 jLj;1]usingtheprecisionparameter,abinarysearchproce-durecanbeconducted.Leta;btobepositiverealnumberswherefortheidenticalarrivalvectora~eCase1holdsandfortheidenticalarrivalvectorb~eCase2holds.Additionallyletb=a+.Nowwehave: 43

PAGE 44

Table2-1. Atabularcomparisonofnegativeresultsfrom[ 49 ]andTheorem 2.4 .Therstcolumnrepresentstheregiontowherearrivalvectorsbelong;thesecondcolumnrepresentsthequeueboundachievementofwhichleadstoP=NP. RegionQueueBound Shahetal.[ 49 ]2p logN NoP(N)Theorem 2.4 oP(N;())]TJ /F6 7.97 Tf 6.59 0 Td[(1) a)]TJ /F3 11.955 Tf 11.96 0 Td[(1 f(G)b+0a)]TJ /F3 11.955 Tf 11.96 0 Td[(1 f(G)a++0a)]TJ /F3 11.955 Tf 11.96 0 Td[(1 f(G)(a)]TJ /F3 11.955 Tf 11.96 0 Td[()+++0:Theprocedurewilllocateanumberzsuchthat:f(G)1=z(1+)f(G)whichisthedenitionof-approximationforaminimizationproblemwithoptimalvaluef(G).Setz=a)]TJ /F3 11.955 Tf 12.32 0 Td[(,=0 2,=0 2andk=1 2jLjinLemma 13 ,andtheratioisachieved.Thepresentedmethodisapolynomial-time-approximationalgorithmtothefractionalcoloringproblem.Howevercombiningresultsfrom[ 54 ]and[ 49 ],weconcludesuchanalgorithmwillimplyP=NP. Corollary4. ThereisnoschedulingalgorithmthathasstorageefciencyforallvectorsintheinteriorofthecapacityregionwithpolynomialPandcomputationefciencyunlessP=NP.Authorsin[ 49 ]providedasimilarresultwhichdisplaysintractabilityofachievinglowqueuedelayswithsimpleper-time-slotcomputationswithinalinearlyscaled-downsubsetofthecapacityregion. Denition9. ([ 49 ])Apolicyhasthepoly-queuepropertyifthereexistsapolynomialp(n)suchthatforeveryn-nodegraph,andeveryarrivalratevectorforwhich()1 (n),theresulting 44

PAGE 45

queuesizessatisfysupE[hQ();1i]p(n); (2)for(n)=n 2p logn.Notethat()=f(G;),wheretherstfunctionusesgraphGimplicitly. Theorem2.5. (Theorem3.1in[ 49 ])Forthecaseofdeterministicarrivalsthereexistsnopoly-queuedeterministicpolicythatrunsinpolynomialtime,unlessP=NP.BothTheorem 2.4 andTheorem 2.5 suggestthatachievingpolynomiallyboundedqueuesizeswithpertime-slotsimple(polynomial)algorithmsischallenging.Table 2-1 summarizestheresults.Next,withthenewload-slackinmind,wewillinvestigatetheconsequencesofhavingapolynomialtimeoracleforWMISproblem.OurprimarymotivationistodemonstratethatwithaWMISsolver,evenifitsapproximate,polynomiallyboundedqueuesizes,whicharetransientinnaturefordeterministicarrivals,canbeachieved.AnalyzingperformanceofMWSviaLyapunovfunctionstypicallyrequiresrelatingtwoinnerproducts:A(t)TQ(t)andD(t)TQ(t).Lemma 14 achievesthistaskbycomparinginnerproductoftwonon-negativevectorsbytheassociatedgraphtheoreticalfunctions. Lemma14. Tf(G;)(G;)8;2RjVj+. Proof. Let2RjMLj+(mi2ML),tobeanoptimalsolutiontof(G;)i.e.,PjMLji=1imiandPjMLji=1i=f(G;),thenwehaveT(jMLjXi=1imi)T=jMLjXi=1imTijMLjXi=1i(G;)=f(G;)(G;): Thislemmacanbeinterpretedasanextensionofcorollary7.5.3in[ 12 ]toarbitraryweightvectors.Inthefollowingdiscussionwe'llillustratehowLemma 14 canbecentraltoderivingthedriftconditionandinSection 3.4 itwillbeusedforextendingtheboundonexpectedWMISsizetothequeuesum. 45

PAGE 46

Westartwithdeningthewirelessschedulingpoliciesink-WMISclass.Forthisclass,schedule(activationvector)m(t)attimetsatises:m(t)TQ(t)k(G;Q(t))where(G;Q(t))=maxm2MLmTQ(t).ConsiderthequadraticLyapunovfunction:RjVj!Rsuchthat(Q(t))=PjVji=1Q2i(t).Let(Q(t))=(Q(t+1)))]TJ /F7 11.955 Tf 11.95 0 Td[((Q(t)).(Q(t))=Q(t+1))]TJ /F3 11.955 Tf 11.96 0 Td[(Q(t)TQ(t+1)+Q(t)=A(t))]TJ /F3 11.955 Tf 11.95 0 Td[(D(t)T2Q(t)+A(t))]TJ /F3 11.955 Tf 11.96 0 Td[(D(t)=2A(t))]TJ /F3 11.955 Tf 11.95 0 Td[(D(t)TQ(t)+A(t))]TJ /F3 11.955 Tf 11.96 0 Td[(D(t)TA(t))]TJ /F3 11.955 Tf 11.95 0 Td[(D(t): (2)SecondtermintherighthandsideofEquation 2 canbebounded:A(t))]TJ /F3 11.955 Tf 9.47 0 Td[(D(t)TA(t))]TJ /F3 11.955 Tf -457.95 -23.91 Td[(D(t)jLj.Nextweboundtherstterm:2A(t))]TJ /F3 11.955 Tf 11.95 0 Td[(D(t)TQ(t)2TQ(t))]TJ /F3 11.955 Tf 11.95 0 Td[(k(G;Q(t))2f(G;)(G;Q(t)))]TJ /F3 11.955 Tf 11.95 0 Td[(k(G;Q(t))2(f(G;))]TJ /F3 11.955 Tf 11.96 0 Td[(k)(G;Q(t)): (2)CombiningEquation 2 andEquation 2 wehave:(Q(t))2(f(G;))]TJ /F3 11.955 Tf 11.95 0 Td[(k)(G;Q(t))+jLj: (2)Letk()=k)]TJ /F3 11.955 Tf 11.57 0 Td[(f(G;)then(Q(t)))]TJ /F7 11.955 Tf 21.92 0 Td[(2k()(G;Q(t))+jLj.When(G;Q(t))jLj 2k(),(Q(t))willdropinthenextround.NotethatbyLemma 14 ;Q(t)Te>f(G)jLj 2k()implies(G;Q(t))jLj 2k().Alsoif(Q(t))(f(G)jLj 2k())2thenQ(t)Te>f(G)jLj 2k().Thatimplieswhenpotentialexceeds(f(G)jLj 2k())2itwilldropinthenextround.Potentialincreaseforanytime-slottisboundedfrom 46

PAGE 47

aboveby2Q(t)+jLj.Themaximumpotentialthatthesystemcaneverreachisboundedby(f(G)jLj 2k())2+2f(G)jLj 2k()+jLj.Boundedpotentialfunctiontranslatesintoboundedqueuesizes:themaximumqueuesizeduringtheoperationofthesystemisboundedbyf(G)jLj 2k()+p jLj. 2.9ResultsandDiscussionInthischapter,weinvestigatedthecapacityregionofawirelessnetworkinameticulousmanner.Inparticular,wedenedthecapacityregionusingtheweightedfractionalcoloringproblemandusedthederivedrelationstoaddressthecomplexityofschedulingdeterministicarrivals.Thechoiceofdeterministicarrivalswasduetoitsimmediateconnectiontotherandom-izedcounterpartandtheeaseofutilizingthecomplexitytheory.Underthedeterministicsetting,weshowedthattheWMIS-basedalgorithmstabilizesthequeueswithefcientbounds.Wealsoinvestigatedinterestingsectionsofthecapacityregionanddenedtheupper-boundary.Were-latedthe-localpoolingfactortotheweightedfractionalcoloringproblem.Wecharacterizedtheupper-boundaryunderthesetlocalpoolingcondition.Wealsoderivedresultsforanewlydenedload-slack,whichmeasurestheconceptofsystemloadmoreconventionally.ItisdisplayedthatwithanapproximateWMISsolverthepolynomialqueueboundscanbeachievedwithinascaledportionofthecapacityregion.Therearemanyremainingquestions.First,intheschedulingcontext,thenecessityoftheWMISproblemincomputingfractionalcoloringnumberhasnotbeenestablishedyet.Second,theapplicabilityoftheresultstothestateofartcontrolalgorithmsrequiresconnectingdeterministicarrivalswithprobabilisticones.Third,webelievethatsetlocalpooling(andlocalpoolingingeneral)isrelatedtothecomplexityoftheWMISproblemingraphs.Webelievethosequestionsformtheimmediateconcernssurroundingthischapter. 47

PAGE 48

CHAPTER3STEADYSTATEDELAYCHARACTERISTICSOFAPPROXIMATEMWSPOLICY 3.1BackgroundPerformanceofwirelesscommunicationsystemsrelyontheefcientutilizationofthesharedtransmissionmedium.Simultaneouslinktransmissionsisknowntocauseinterferenceinthelackofcoordination.Mediumaccesscontrol(MAC)protocolsaddressthisproblembydevisingchannelaccessmechanismsthatoptimizeobjectivesofchoice.Throughputhasbeenconsideredasoneofthemostprevalentnetworkperformancemetricsinthelasttwodecades.Nevertheless,recentlythereisagrowinginterestonthedelayboundsofwirelessschedulingalgorithms[ 26 ][ 15 ].Thischapterstudiesthedelayboundsoftheapproximatemaximumweightscheduling(k-MWS)policy.Intheirseminalwork[ 52 ],Tassiulasetal.proposedathroughputoptimalschedulingpolicythat,whenappliedtoprotocolmodel,activatesaWMISofthequeue-weightedinterferencegraphateachtime-slot.(Alsocommonlyknownasmaximumweightscheduling(MWS)policy).HowevertheiralgorithmisnotapplicabletolargenetworkssinceWMISproblemisNP-hardingeneral.Laterin[ 53 ],Tassiulasprovidedalineartimeprobabilisticalgorithmachievingthroughputoptimality.Althoughfromcomplexityperspective,twoalgorithmsdiffersignicantly,looselyspeaking,thelattertriestosimulatetheformerusingarandomizedsearchscheme[ 27 ].Howeverwhenthelatteralgorithmisusedqueuesizesaresubjecttogrowexponentially,leadingtoanexorbitantmemoryusage.Accordinglyrecentresearchfocusesondelayboundsofthroughputoptimalalgorithms[ 26 ][ 15 ].Tothebestofourknowledgethereislittleprogressinderivingboundsfortransientexpectedqueuelengths.Asaresultresearchersconcentrateonthesteady-stateboundsofexpectedqueuesizes.Theprecedingresearchonswitchschedulingprovidedmanyresultsthatareapplicabletothewirelessschedulingwithminormodications.In[ 48 ]authorsinvestigatedtheexpectedqueuesizeboundsatstationarityforvirtualoutputqueuedswitches.Theirframeworkcorrespondstoaspecialcaseofthegeneralinterferencemodel,wherenodeexclusiveinterferencemodelisusedonacompletebipartitegraph.Later 48

PAGE 49

in[ 26 ]ageneralizationforarbitrarygraphsisprovidedfortheMWSalgorithm.ThelatesteffortinthisdomainintroducedavariantofMWStoachievetightexpectedqueueboundsatstationarity[ 15 ].TocounterthedifcultyofcomputingWMISateachtime-slot,astreamofresearchconcentratesondevisingalgorithmsthataresimpleyetefcientinpractice.Thelongestqueuerst(LQF)policyisknownforitssimplicity,highperformance,andaptnessfordistributedimplementation,acharacteristicneededforschedulingalgorithmsinthelackofcentralcontrol[ 24 ].PerformanceofLQFisstudiedfromtheperspectiveofthroughputin[ 7 ],[ 8 ],and[ 24 ].However,tothebestofourknowledge,delaycharacteristicsofLQFpolicyinwirelessnetworksisstillnotaddressedformally.Thefollowingisthesummaryofourmainresultsandcontributionsinthischapter: Weprovideadelayboundforthek-MWSpolicy.Ourresultimprovestheanalyticalboundderivedin[ 26 ]whenk=1andgeneralizestheboundfor0
PAGE 50

relationswithaninterferencegraph,G=(V;E),whereeachlinkl2Lintheoriginalnetworkcorrespondstoanodev2V.Twonodesv1;v22VareconnectedinG,ifandonlyifthecorrespondinglinksintheoriginalnetworkinterferewitheachother.WeassumeasymmetricinterferencerelationthusGisundirected.Afeasiblescheduleisdenedtobeasetofnon-interferingnodesinG.Amaximalscheduleisafeasibleschedulethatcannotincludeanymorenodeswithoutviolatingthefeasibilityconstraint.AfeasiblescheduleinanetworkcorrespondstoanindependentsetinG.WedenotethesetofallmaximalschedulesbyML,andusem2MLtodenoteasinglemaximalschedule.Throughoutthechapter,weusethewordscheduletorefertoafeasiblemaximalschedule.Whenapplicable,MLisregardedasajVjjMLj0-1matrix.Eachcolumnofthematrixisa0-1vectorrepresentationofamaximalindependentsetofG,with1indicatingthatthecorrespondingnode(inG)isselected(andthelinkisactiveintheschedule)and0otherwise.Acolumnvectorthathas1'sinallentriesisdenotedbye.Weadopttheconventionaldenitionofthecapacityregion:=fj0forsome2Co(ML)g:Theinteriorofthecapacityregionisdenotedbyo.Weassumeatime-slottedsystem.Thepacketsarrivingateachlinkarequeued.Eacharrivalstreamisi.i.d.intime.Thedistributionofthenumberofpacketsarrivingtolinkiattimet,Ai(t),istimeinvariant.Assumesecondmoments,E(A2i),ofthearrivalprocessesareboundedbyC.ThearrivalvectorforallqueuesattimetisdenotedbyA(t).SimilarlythequeuelengthvectorisdenotedbyQ(t).Alinkcanbeactivatedinatimeslot,onlyifthequeueisnon-empty.Atmostonepacketcanbeservedataqueueinagiventimeslot.Foreachlinki,functionDi(t)indicateswhetherornotlinkireceivedserviceattimeslott.Notethat:Di(t)=8>><>>:1;ifQi(t)1andiisscheduled.0;otherwise. 50

PAGE 51

D(t)denotesthevectorofdeparturesattimet.Theevolutionofthequeueisasfollows:Q(t+1)=Q(t)+A(t))]TJ /F3 11.955 Tf 11.96 0 Td[(D(t):Inthischapterweinheritthedenitionofstabilityfrom[ 52 ],whereitisdenedastheunderlyingirreducibleDiscreteTimeMarkovChain(DTMC)beingpositiverecurrent.Athroughputoptimalalgorithmkeepsthequeuesstableforanyaveragearrivalratefromtheinteriorofo. 3.3GraphTheoreticalResultsInthissectionourprimaryfocuswillbeonbuildingsupportinglemmasfortheperformanceevaluationoftwoparticularalgorithms.Therstalgorithmisthek-MWSpolicy(k2(0;1]),whichpicksaschedule,denotedbymk(t),thatapproximatesthemaximuminnerproductwiththequeuelengthsatthecurrenttimesloti.e.,mk(t)TQ(t)kmaxm2MLmTQ(t).Itisknownthatthek-MWSwouldstabilizeaveragearrivalrateswithinko[ 57 ].ThesecondalgorithmistheLQFpolicywhichschedulesthelongestqueueamongtheavailablelinks,anddiscardstheneighboringlinks.Thisprocedureisrepeateduntilallthequeuesareeitherscheduledordiscarded.We'llcallanyschedulethatisproducedbythisprocedureaLQFschedule.Whenprotocolmodelisassumed,capacityregioncanbecharacterizedusingtheconceptoffractionalcoloring(seeDenition 1 ).Computingf(G;w)isNP-hardtoapproximateingeneral.Wedenotetheweightedindependencenumber(totalweightofWMIS)ofagraphwith(G;w),wherew2RjVj+istheweightvectorasusual.Wedenotef(G;e)and(G;e)withf(G)and(G)respectively.Let(G;;)bequeuelengthsumofaLQFschedule,where2RjVj+isthevectorrepresentingthequeuelengths.TheLQFscheduleischosenasgivenabovewithanadditionalparameter,apermutationofvertices,toprovidetie-breakingifnecessary.We'llbederivingresultsindependentof.Inordertoeliminatewe'lldeneafunctiontorepresentminimumqueue-sumLQFscheduleoverallpossiblepermutations,; 51

PAGE 52

Denition10. (G;)=min2(G;;).In[ 7 ]and[ 24 ]theconceptofinterferencedegreeisusedtocharacterizetheperformanceofLQF.Wedenotetheclosedneighborhoodofanodev2VbyN[v].AlsoforasubsetofnodesSV,G[S]denotesthevertexinducedsubgraphofG[ 55 ].Thefollowingisthedenitionoftheinterferencedegree: Denition11. =maxv2V(G[N[v]]).Inwords,interferencedegreeisthemaximumnumberofnodesthatcouldbeactivatedsimultaneouslyintheneighborhoodofanode.In[ 7 ],itisshownwhenanaveragearrivalratevectoriswithin1 o,LQFwillstabilizethenetwork.ThefollowinglemmarelatesLQFschedulingtoWMISandjustiestheunderlyingperformancebound: Lemma15. (G;)(G;)82RjVj+. Proof. Omittedforbrevity. ThegreedyscheduleselectionofLQFprovidesaapproximationtotheWMISproblem.We'lldenoteanLQFscheduleachieving(G;Q(t))withmLQF(t).Lemma 16 withCorollary 5 demonstratethataLQFscheduleproducedwithqueueweightsQ(t)isstillaapproximationtoWMISwhenthequeuesizesareraisedtoapowern2Z+: Lemma16. (G;Qn(t))mLQF(t)TQn(t)8Qn(t)2RjVj+;n2Z+. Proof. Firstwenotice,sinceexponentiationisanon-decreasingfunctionwhenQ(t)2ZjVj+andn2Z+,mLQF(t)isstillaLQFschedulepickedusingweightsQn(t)undersomepermutationofvertices.Combinethisfactwiththedenitionof(G;Qn(t))andtheresultfollows. Corollary5. 1 (G;Qn(t))mLQF(t)TQn(t).HowevertheboundprovidedinCorollary 5 forLQFisnotcommontoanyk-MWSpolicy.WeprovidethegeneralizedboundsinLemma 17 : Lemma17. kn ((G))n)]TJ /F6 7.97 Tf 6.59 0 Td[(1(G;Qn(t))mk(t)TQn(t)8Q(t)2ZjVj+;n2Z+;k2(0;1]. 52

PAGE 53

Proof. (G;Qn(t))((G;Q(t)))n(1 kmk(t)TQ(t))n(1 k)n(mk(t)TQ(t))n(1 k)n((G))n)]TJ /F6 7.97 Tf 6.58 0 Td[(1mk(t)TQn(t);wherethelastinequalityfollowsfromthefacts:forai;n;l2Z+,(Pli=1ai)n(l)n)]TJ /F6 7.97 Tf 6.58 0 Td[(1Pli=1(ai)nandcardinalityofanyscheduleisupperboundedby(G). Figure3-1. IllustrationontightnessofLemma 17 TheboundprovidedinLemma 17 couldbetightasillustratedinFigure 3-1 .InFigure 3-1 forthecurrenttime-slottthequeuesizesareasfollows;Q1(t)=6;Q2(t)=Q3(t)=Q4(t)=1.Fork=1 2,theactivationvectormk(t)=f0;1;1;1gcorrespondstoaschedulethatwouldbeactivatedbythek-MWSpolicy.TheschedulefromMWSpolicyontheotherhandistheactivationvectorm(t)=f1;0;0;0g.Also(G;Qn(t))=6n.Interferencedegree,forthegivengraphis3.Readercanverifyforn=2;3;:::,theboundgiveninLemma 17 istight.LQFpolicyontheotherhanddemonstratesaninvariantbehaviorwhenexponentiationisinvolved.ThatpropertywillproveitselftobeusefulinSection 3.5 whenwearetoboundthehighermomentsofthequeuesizedistributionatstationarity. 53

PAGE 54

3.4DelayBoundsofApproximateMWSPolicyInthissectionwederiveboundsonstationaryexpectedqueue-sumsintheguidanceofLemma 14 .ThedelayboundfollowsusingLittle'sLaw[ 36 ].Ourderivationfollowsfrom[ 26 ],however,weimprovetheboundsbyusingafractionalcoloringscheme.ConsiderthequadraticLyapunovfunction:RjVj!Rsuchthat(Q(t))=PjVji=1Q2i(t).(Q(t+1)))]TJ /F7 11.955 Tf 11.95 0 Td[((Q(t))=Q(t+1))]TJ /F3 11.955 Tf 11.95 0 Td[(Q(t)TQ(t+1)+Q(t)=A(t))]TJ /F3 11.955 Tf 11.96 0 Td[(D(t)T2Q(t)+A(t))]TJ /F3 11.955 Tf 11.95 0 Td[(D(t)=2A(t))]TJ /F3 11.955 Tf 11.96 0 Td[(D(t)TQ(t)+A(t))]TJ /F3 11.955 Tf 11.95 0 Td[(D(t)TA(t))]TJ /F3 11.955 Tf 11.96 0 Td[(D(t)=2A(t))]TJ /F3 11.955 Tf 11.96 0 Td[(D(t)TQ(t)+A(t)TA(t)+D(t)TD(t))]TJ /F7 11.955 Tf 11.96 0 Td[(2A(t)TD(t): (3)Ifthenetworkisstableunderagivenpolicy;astimetendstoinnityQ(t);A(t);D(t)constitutesapositiverecurrentDTMC[ 52 ].Similarto[ 31 ],atstationarity,duetopolynomialmomentsofthequeuelengthsbeingnite[ 30 ],foralowerboundedpolynomialpotentialfunctionL:RjVj!Rwehavelimt!1E[L(Q(t+1)))]TJ /F3 11.955 Tf 12.35 0 Td[(L(Q(t))]=0.Sincesatisesthecondition,whenthesystemreachesstationarity:E[(Q(t+1)))]TJ /F7 11.955 Tf 11.96 0 Td[((Q(t))]=2E[(A(t))]TJ /F3 11.955 Tf 11.96 0 Td[(D(t))TQ(t)]+E[A(t)TA(t)]+E[D(t)TD(t)])]TJ /F7 11.955 Tf 11.96 0 Td[(2E[A(t)TD(t)] (3)=0:NextweboundtheadditivetermsinEquation 3 .UsingLemma 14 wecanobtain: 54

PAGE 55

E[A(t)TQ(t)jQ(t)]=TQ(t)f(G;)(G;Q(t)):Bytakingexpectationonbothsideswehave;E[A(t)TQ(t)]f(G;)E[(G;Q(t))]: (3)SinceDi(t)2f0;1g;D2i(t)=Di(t).Additionallythesystemisatthestationarityi.e.,E[A(t)]=E[D(t)],thusE[D(t)TD(t)]=E[jVjXi=1Di(t)]=jVjXi=1i: (3)AlsoAi(t)isindependentofDi(t)leadingto;E[A(t)TD(t)]=E[A(t)]TE[D(t)]=jVjXi=12i: (3) Lemma18. Underk-MWSpolicy,foranarrivalrate2ko,atstationarityonehas:E[jVjXi=1Qi(t)]f(G) 2k)]TJ /F3 11.955 Tf 11.95 0 Td[(f(G;)B1whereB1=PjVji=1i+E[A2i(t)])]TJ /F7 11.955 Tf 11.96 0 Td[(22i. Proof. Underk-MWSpolicywehaveE[D(t)TQ(t)]kE[(G;Q(t))]:CombiningwithEquation 3 weget 55

PAGE 56

2E[A(t)TQ(t))]TJ /F3 11.955 Tf 11.96 0 Td[(D(t)TQ(t)]2(f(G;))]TJ /F3 11.955 Tf 11.96 0 Td[(k)E[(G;Q(t))]: (3)WeobtainthenextresultbyusingEquation 3 ), 3 ,and, 3 inEquation 3 :02(f(G;))]TJ /F3 11.955 Tf 11.96 0 Td[(k)E[(G;Q(t))]+jVjXi=1i+jVjXi=1E[A2i(t)])]TJ /F7 11.955 Tf 11.95 0 Td[(2jVjXi=12i:Notingf(G;)
PAGE 57

Ourresultimprovestheboundin[ 26 ]byafactorof(G) f(G),wheref(G)istheintegralcoloringnumberofG.Nextweprovideacomparisonbetweenmagnitudesof(G)andf(G).In[ 29 ]authorsprovedthedifferencebetween(G)andf(G)canbearbitrarilylarge: Lemma19. Foranyintegern2,thereexitsauniquelycolorable,vertextransitivegraphG,suchthat(G))]TJ /F3 11.955 Tf 11.96 0 Td[(f(G)>n)]TJ /F7 11.955 Tf 11.96 0 Td[(2.Howeverthatresultcountsonthegrowingsizeofthegraphconsidered.Lovaszin[ 39 ]providedaninequalitytoupperboundtheintegralcoloringnumberwhichrelatestothefractionalcounterpartasfollows: Lemma20. f(G)(G)<(1+ln((G)))f(G).Inotherwords,byLemma 19 ,ourboundcanbearbitrarilytighterthantheresultsprovidedin[ 26 ].HoweverLemma 20 statesthatwhentheboundsarenormalizedbythegraphsize,theimprovementfactorisatmost1+ln((G)). 3.5LQFasanApproximateMWSPolicyInthissectionwederiveanalyticalboundsforE[PjVji=1Q2i(t)],whenk-MWSpolicyisutilized,andtheaveragearrivalratesatises:2k2 (G)o(Thescalingcanbeconsideredasathroughputcompromisetoachievetighterbounds).CombinedwiththeresultfromSection 3.4 ,boundsonqueuelengthvarianceatstationaritycouldbeobtained.AdditionallywhenLQFschedulingpolicyisemployed,wendthat,upperboundsforhighermomentsofstationaryqueuesumdistributioncanbecomputedwithinasubsetofthestabilityregion,1 o,withoutcompromisingthroughput.ConsiderthecubicLyapunovfunction:RjVj!Rsuchthat(Q(t))=PjVji=1Q3i(t).Atstationarity,wehave: 57

PAGE 58

E[(Q(t+1)))]TJ /F7 11.955 Tf 11.96 0 Td[((Q(t))]=3E[(A(t))]TJ /F3 11.955 Tf 11.96 0 Td[(D(t))TQ2(t)]+3E[jVjXi=1((Ai(t))]TJ /F3 11.955 Tf 11.95 0 Td[(Di(t))2Qi(t)]+E[jVjXi=1(Ai(t))]TJ /F3 11.955 Tf 11.96 0 Td[(Di(t))3] (3)=0:NextweboundadditivetermsinEquation 3 .DeneB2=PjVji=1E[A3i(t)])]TJ /F7 11.955 Tf -407.63 -23.91 Td[(3iE[A2i(t)]+32i)]TJ /F3 11.955 Tf 11.96 0 Td[(ithen;E[jVjXi=1(Ai(t))]TJ /F3 11.955 Tf 11.96 0 Td[(Di(t))3]=B2: (3)Rememberingsecondmomentofthearrivalprocessisboundedwehave;E[jVjXi=1((Ai(t))]TJ /F3 11.955 Tf 11.96 0 Td[(Di(t))2Qi(t)]E[(C+1)jVjXi=1Qi(t)](C+1)f(G) 2k)]TJ /F3 11.955 Tf 11.96 0 Td[(f(G;)B1: (3)SimilartothederivationinEquation 3 ,byLemma 14 ,E[jVjXi=1(Ai(t)Q2i(t)jQ(t))]=TQ2(t)(G;Q2(t))f(G;):Additionallyunderk-MWSpolicywehave;E[jVjXi=1Di(t)Q2i(t)jQ(t)]k2 (G)(G;Q2(t)): (3)Thusweobtain; 58

PAGE 59

E[jVjXi=1(Ai(t))]TJ /F3 11.955 Tf 11.96 0 Td[(Di(t))Q2i(t)]E[(G;Q2(t))](f(G;))]TJ /F3 11.955 Tf 20.47 8.09 Td[(k2 (G)): (3)CombiningEquation 3 3 ,and, 3 inEquation 3 :E[(G;Q2(t))]3(k2 (G))]TJ /F3 11.955 Tf 11.96 0 Td[(f(G;))3 2(C+1)f(G) k)]TJ /F3 11.955 Tf 11.95 0 Td[(f(G;)B1+B2:When2k2 (G)o;E[(G;Q2(t))](C+1)f(G)B1 2+B2(k)]TJ /F3 11.955 Tf 11.96 0 Td[(f(G;)) 3 (k2 (G))]TJ /F3 11.955 Tf 11.96 0 Td[(f(G;))(k)]TJ /F3 11.955 Tf 11.95 0 Td[(f(G;)):ByapplyingLemma 14 againwenallyget;E[jVjXi=1Q2i(t)]=E[eTQ2(t)]f(G)E[(G;Q2(t))]f(G)(C+1)f(G)B1 2+B2(k)]TJ /F3 11.955 Tf 11.96 0 Td[(f(G;)) 3 (k2 (G))]TJ /F3 11.955 Tf 11.96 0 Td[(f(G;))(k)]TJ /F3 11.955 Tf 11.95 0 Td[(f(G;)):NextweprovidetighterboundsforEPjVji=1Q2i(t)underLQFwithoutathroughputcompromise(within1 o).ConsiderCorollary 5 withk=1 andn=2i.e.,1 (G;Q2(t))mLQF(t)TQ2(t): 59

PAGE 60

When21 oandLQFschedulingpolicyisused,byusingthederivationprovidedaboveandconsideringLQFinEquation 3 weget:E[jVjXi=1Q2i(t)]f(G)(C+1)f(G)B1 2+B2(1 )]TJ /F23 10.909 Tf 10.91 0 Td[(f(G;)) 3 (1 )]TJ /F23 10.909 Tf 10.91 0 Td[(f(G;))2:FinallyitispossibletoderiveupperboundsonthequeuevarianceusingtheboundsonEPjVji=1Q2i(t)andEPjVji=1Qi(t).Theframeworkwe'vepresentedcanbeextendedtothehighermomentsofqueuesizedistributionatstationaritywithlittleeffort,givenallassociatedmomentsofthearrivalprocessarebounded. 3.6MiscellaneousResultson-LocalPoolingInthissectionwewillinvestigatethefollowingquestionsthatrelatetothestate-of-artperformancemetricofLQFpolicy,-localpoolingfactor.Adetaileddiscussiononthesignicanceof-localpoolingfactorcanbefoundin[ 32 ]. Isthereanylower-boundforthe-localpoolingfactoringraphs? Whatisthecomputationalcomplexityofcalculatingset-localpoolingfactor(whichformsthebasisforthedenitionoftheoverall-localpoolingfactor)? Whatkindofstructures,whentheyareaddedtoagraph,donotaffect-localpoolingfactoringraphs?Westartwithcitingalistofresultsfrom[ 33 ]thatwillbeusedforansweringquestionsgivenabove.Inresultslistedbelowusethefollowingnotation:xisaweightassignmentonthenodesoftheinterferencegraph,s2MLisamaximalindependentsetintheinterferencegraph,x(s)isthetotalweightofthesets,andG(V;E;I)istheinterferencegraphcomposedofvertexsetV,edgesetE,andacollectionofindependentsetsI. 60

PAGE 61

Lemma21. (Lemma6in[ 33 ])Foralinkl2E,listhesmallestLforallLEthatcontainsl,i.e.,l=minfLEjl2LgL (3) Lemma22. (Lemma1in[ 33 ])ForanetworkgraphG=(V;E;I),thefollowingholds.(G)=minl2El: (3)Equation9in[ 33 ]:L=w=maxx0minixTsi maxixTsi: (3) Lemma23. (Lemma5in[ 33 ])Listheoptimalvalueofthefollowingoptimizationproblem.maxx0;wws.t.xTMLeTc(L)xTMLweTc(L):Equation36in[ 33 ]:G=maxx0mins2MGx(s) maxs2MGx(s): (3) Lemma24. (Lemma24in[ 33 ])IfagraphGisvertex-transitive,thentheoptimizationproblemgiveninEquation 3 isachievedbyanequalweightassignmentonthenodes. 3.6.1GraphswithArbitrarilySmallSet-LocalPoolingFactorsWewillprovethatset-localpoolingfactorscanbearbitrarilysmallusingaspecialclassofgraphs,thehypercubes.Asimilarresultwasfoundindependentlyin[ 3 ]usingadifferentclassofgraphs.ThehypercubegraphQnisaregulargraphwith2nvertices.ThehypercubegraphQncanbeconstructedbylabelingthe2nvertices0;1;2;:::;2n)]TJ /F7 11.955 Tf 12.6 0 Td[(1andconnectingtwoverticeswhenevertheHammingdistanceofthebinaryrepresentationsofthelabelsisequalto1.ThisconstructionisillustratedinFigure 3-2 .ThevertexsetofQnisdenotedbyVQn. 61

PAGE 62

Lemma25. AmaximumcardinalityindependentsetofQnhasatleast2n)]TJ /F6 7.97 Tf 6.59 0 Td[(1elements. Proof. Wewillinductivelyconstructanindependentsetofthedesiredcardinality.Theinductionhypothesisis:Qnhasamaximalindependentset,ZVQn,suchthatVQn)]TJ /F3 11.955 Tf 12.55 0 Td[(Zisamaximalindependentsetaswell.Forn=1,thestatementholds(forn=0,thestatementwouldholdwithproperdenitions).ItisknownthatQn+1canbeconstructedbyusingtwohypercubesQnandconnectingcorrespondingverticestogether.TakeQ1nandQ2n,twocopiesofQn,toconstructQn+1.WewillshowQn+1hasamaximalindependentsetZ,whosecomplement,VQn+1)]TJ /F3 11.955 Tf 12.49 0 Td[(Z,isamaximalindependentsetaswell.Bytheinductionhypothesis,Q1ncanbepartitionedintotwomaximalindependentsets,Z1andVQ1n)]TJ /F3 11.955 Tf 12.08 0 Td[(Z1.LetZ2bethemaximalindependentsetofQ2ncorrespondingtoZ1.Now,observeZ=Z1[(VQ2n)]TJ /F3 11.955 Tf 12.53 0 Td[(Z2)isamaximalindependentsetanditscomplementVQn+1)]TJ /F3 11.955 Tf 12.39 0 Td[(Z=VQn+1)]TJ /F7 11.955 Tf 12.39 0 Td[((Z1[(VQ2n)]TJ /F3 11.955 Tf 12.39 0 Td[(Z2))=Z2[(VQ1n)]TJ /F3 11.955 Tf 12.39 0 Td[(Z1)isamaximalindependentsetaswell.Thustheresultfollows.WeconcludethatQncanbepartitionedintotwomaximalindependentsetsandoneofthemmusthaveatleast2n)]TJ /F6 7.97 Tf 6.58 0 Td[(1elements. Lemma26. AminimumcardinalitymaximalindependentsetofQnhasatmost2n=(n+1)elementswhenn=2k)]TJ /F7 11.955 Tf 11.95 0 Td[(1,wherek2Z+. Proof. In[ 38 ],itisshownwhenevernisintheformn=2k)]TJ /F7 11.955 Tf 12.53 0 Td[(1forsomepositiveintegerk,Qnhasaperfectdominatingset.Inperfectdomination,everyvertexcanonlybedominatedbyasinglenode.SinceeverynodeinQnhasdegreen,thereareatmost2n=(n+1)elementsinaperfectdominatingset.Notethatanindependentdominatingsetisamaximalindependentset. Lemma27. Letan=2n)]TJ /F7 11.955 Tf 11.96 0 Td[(1,thenQan21)]TJ /F5 7.97 Tf 6.59 0 Td[(n. Proof. Ahypercubeisvertex-transitive.ByLemma 24 ,Qancanbederivedunderanequalweightassignmentonthevertices.LetMQanbesetofallmaximalindependentsetsofQan.By 62

PAGE 63

Lemma 25 ,maxm2MQanjmj2an)]TJ /F6 7.97 Tf 6.58 0 Td[(1.ByLemma 26 ,minm2MQanjmj2an=(an+1).Since Qan=minm2MQanjmj maxm2MQanjmj2an=(an+1) 2an)]TJ /F6 7.97 Tf 6.59 0 Td[(1=21)]TJ /F5 7.97 Tf 6.59 0 Td[(n;thelemmaholds. Corollary6. Letan=2n)]TJ /F7 11.955 Tf 11.95 0 Td[(1.Then,limn!1Qan=0. Figure3-2. Hypercubesofdifferentdimensions. 3.6.2ComputationalComplexityofCalculatingGAgain,considertheinterferencegraph(orsubgraph)G=(V;E).ByLemma 23 ,theset-localpoolingfactor,G,istheoptimalvalueofthefollowingoptimizationproblem.maxw (3)subjecttox0MGe0 (3)x0MGwe0 (3)x;w0: (3) 63

PAGE 64

Here,MGisamatrixwherethecolumnsareallthemaximalindependentsetsofG,ande=(1;1;:::;1)0ofanappropriatedimension.WewillinvestigatethecomputationalcomplexityofcalculatingGbysolvingtheaboveoptimizationproblem.AseparationoraclefortheabovelinearprogramisaproceduretotestwhetheragivenvectorisintheconvexregiondenedbytheconstraintsgiveninEquation 3 throughEquation 3 ,andifnot,ndaviolatingconstraint.LetSEPORC(G)denotesuchaseparationoracle.WewillrstshowSEPORC(G)isNP-hardbygivingareductionfromtheminimumcardinalitymaximalindependentsetproblemtoSEPORC(G). Lemma28. SEPORC(G)isNP-hard. Proof. Thedecisionversionoftheminimumcardinalitymaximalindependentset(MCMIS,alsoknownastheminimumindependentdominatingset)problemaskswhetherthereisamaximalindependentsetofsizeKorless,where1KjVj.ItisknownthatthisproblemisNP-complete(page190,[ 11 ]).WewillprovideaTuringreductionfromtheMCMISproblemtoSEPORC(G).Givenavectorx2QjVjandw2Q,SEPORC(G)candecidethemembershipof(x;w)intheconvexregiondenedbytheconstraintsgiveninEquation 3 throughEquation 3 .Usingthisseparationoracle,wewillcreateasolverfortheMCMISproblem(thusprovidingaTuringreduction).Wesetxi=1 jVjfori=1;:::;jVjandsetw=(K+1)=jVj.Thevector(x;w)alwayssatisesconstraintinEquation 3 .Hence,feeding(x;w)toSEPORC(G)willtelluswhether(x;w)satisesconstraintinEquation 3 ornot.Ifyes,anMCMISisatleastofthesizejVjw=K+1;ifno,theMCMIShaslessthanorequaltoKelements. Remark:Intheproof,onlythefeasibilityaspectoftheseparationoracleisused.TheproofreallysaysthefeasibilityproblemisNP-hard.But,SEPORC(G)isatleastashardasthefeasibilityproblem. Lemma29. ComputingGbysolvingtheoptimizationproblemgiveninEquation 3 throughEquation 3 isNP-hard. 64

PAGE 65

Proof. In[ 13 ],theauthorshaveestablishedthatthecomplexityoftheseparationoracleandthatoftheoriginaloptimizationproblemarepolynomiallyequivalent.CombinethisfactwithLemma 28 Remark:ThestatementofLemma 29 shouldnotbeunderstoodasndingGisNP-hard.ItmeansthatsolvingtheoptimizationproblemgiveninEquation 3 throughEquation 3 isNP-hard,whichinvolvesndingbothanoptimalsolutionandtheoptimalvalue. 3.6.3SpecialStructuresThatPreserveGInagraph,wewillcallanoderasuper-nodeifitisconnectedtoallothernodes.Asuper-nodeinaninterferencegraphcorrespondstoalinkinthenetworkthatinterfereswithallotherlinks.SupposewestartwithaninterferencegraphG=(V;E).Wewillinvestigatetheeffectofinsertingasuper-nodetoG.Wewillshowthatthelocalpoolingfactor(forgraphs)isnotalteredbytheinsertion.Wedenotethegraphaftertheinsertionofthesuper-nodebyG0=(V0;E0).Withslightabuseofnotation,wedenotethegraphlocalpoolingfactorsby(G)and(G0)forGandG0,respectively(insteadofthenotationthatusesthenetworkgraph).GivenasubsetTV,wedenotetheset-localpoolingfactorcorrespondingTbyT. Figure3-3. IllustrationofGandG' Lemma30. (G)=(G0). Proof. Wewillconsidercomputingtheset-localpoolingfactorT,TV,accordingtoEquation 3 ,whichwere-writenextusingsomenewnotations.LetMTbethesetofmaximumindependentsetsfortheinterferencesubgraphassociatedwithT.Givenaweightassignmentxonthenodesandasetofnodessand,letx(s)bethetotalweightofthesets.We 65

PAGE 66

knowthatT=maxx0mins2MTx(s) maxs2MTx(s): (3)Alsonote(bycombiningLemma 22 andLemma 21 ) (G)=minTVT:(3)Letthesuper-nodebedenotedbyr.Then,V0=V[r.ConsiderthecasewhereT0V0andr2T0.LetT=T0)-245(frg.WewillshowTT0.LetxbeanoptimalweightassignmentonthenodesinTthatachievesT.Letabeanarbitraryvalueon[mins2MTx(s);maxs2MTx(s)],andassigntheweightatonoder.Thevectory=(x0;a)0isaweightassignmentonthenodesinT0.Next,notethatMT0=MT[fsg.Hence, mins2MT0y(s)=min(mins2MTx(s);a)=mins2MTx(s);and maxs2MT0y(s)=max(maxs2MTx(s);a)=maxs2MTx(s):Asaresult, T0mins2MT0y(s) maxs2MT0y(s)=mins2MTx(s) maxs2MTx(s)=T:(3)Next,(G0)=minT0V0T0=min(minT0V0;r2T0T0;minT0V0;r=2T0T0)=min(minT0V0;r2T0T0;(G)): (3)ByEquation 3 andtheone-to-onecorrespondencebetweenT0andTthroughT0=T[frg,whereT0V0,r2T0andTV,wegetminT0V0;r2T0T0minTVT=(G): (3) 66

PAGE 67

CombiningEquation 3 andEquation 3 ,wehave(G0)=(G): 3.7ResultsandDiscussionInthischapterweprovideimproveddelayboundsfortheapproximateMWSalgorithminwirelessnetworks.Weillustrateusingourframeworkitispossibletoboundqueue-sumvariance,withinadiminishingsubsetofthecapacityregion.AdditionallyweshowhighermomentsofqueuelengthsumunderLQFcanbeboundedwithin1 withoutathroughputcompromise.Wealsoconsidermiscellaneoustopologicalresultsonlocal-poolingfactorofgraphs.Ourworkcontributestotheeldofsub-optimalschedulingalgorithmsandparticularlytothestudiesontheperformanceofLQF.Ourfutureworkconsistsofextendingthepresentedframework,forhighermoments,tothewholestabilityregionofoftheapproximateMWSpolicy.Itisknownallpolynomialmomentsofthequeuesumdistributionatstationarityarenite[ 30 ].ItmightbeinterestingtondpotentialfunctionstoexploitLemma 14 forderivingboundsonhighermoments.AlsoLQFisknowntobethroughputoptimalwhentheunderlyingconictgraphsatiseslocalpoolingcondition[ 8 ].Anopenquestioniswhether,forlocalpoolinggraphsunderLQF,ourframeworkadmitsboundsforhighermomentswithinthewholestabilityregionofthenetwork.Webelieveifitistrue,thederivedboundscancharacterizeperformanceofLQFmoreprecisely. 67

PAGE 68

CHAPTER4CHARACTERIZINGTHEEXPECTEDQUEUE-SUMBOUNDOFTHEMWSPOLICY 4.1BackgroundThroughputoptimalityhasbeenoneoftheprimaryconcernsofwirelessnetworkre-searchersinthelasttwodecades.Itispossibletoachievethroughputoptimalityviaawiderangeofalgorithms.Recenteffortsinwirelessschedulingresearchconcentrateonndingdelayef-cientthroughputoptimalalgorithms.InthischapterweconcentrateonageneralizationofMWSpolicy,whichscalestheindividualqueuesizeswithconstantstogetqueueweightsthatareusedtocomputeaWMIS.ThisversionoftheMWSpolicyisstudiedin[ 15 ].Authorsshow,underthatgeneralization,anysetofpositiveconstantswillensurestability.Nexttheyinvestigatetheperformanceimprovementondelayusingthealgorithmwithoptimizedparameters.TheprimaryobjectiveoftheauthorsthereinistoprovethegeneralizedMWSpolicyisabletoprovidequeueboundsatsystemstationaritythatisatleastasgoodasanyrandomizedpolicythatassignscon-stantactivationprobabilitiestotheschedules.InotherwordsauthorsshowedthatthegeneralizedMWSpolicyisbetterthananyalgorithmthatusesaprobabilitydistributionoverallpossibleschedules.InthischapterourobjectiveistocharacterizethedelayperformanceofMWSpolicyusingasingleparameter.Towardsthatend,wewillderiveaparameterthatimprovesonthepreviousresultsandistightwhenquadraticpotentialfunctionisbeingutilizedforproducingthedelaybounds.Weprovideacomparisonbetweenthepreviousparameterandthenewoneunderlyingthetopologieswhereaperformanceimprovementisachievedwiththenewchoice.Aneasyestimationofthedelayperformanceisprovidedforthecompleteness.Wealsoextendouranalysistothealgorithmconsideredin[ 15 ],andprovetheweightselectionproposedin[ 15 ]isanoptimalonefromdelayperspective.Intheprocessweidentifytheimportantparametersinthequadraticpotentialfunctionframeworkandprovidetechniquestoassesstheiroptimality. 4.2DelayCharacterizationUsingaSingleParameterSystemmodelinthischapterfollowsdirectlyfromchapter 3 .Wereferreadertosection 3.2 fordetailedsystemmodelandnotations. 68

PAGE 69

InChapter 3 weprovidedthestationaryqueuesumboundf(G) 21)]TJ /F5 7.97 Tf 6.58 0 Td[(f(G;)BwhereB=PjVji=1i+E[A2i(t)])]TJ /F7 11.955 Tf 12.8 0 Td[(22i.Wesignifythedeningfactorinthatinequalitywith()=1)]TJ /F5 7.97 Tf 6.59 0 Td[(f(G;) f(G).Nextwewillshowthatanotherscalar()suchthatf(G;+e())=1,canreplace()providingabetterbound.Firstweshow()isanupper-boundfor(). Lemma31. ()()2RN+. Proof. Westartwithasimpletriangleinequalitydenedwithrespecttothefractionalcoloringnumber:f(G;)+f(G;)f(G;+)Forany;2RN+: (4)UsingEquation 4 anddenitionof()weget:f(G;)+f(G;()e)f(G;+e())f(G;)+()f(G)1()1)]TJ /F3 11.955 Tf 11.96 0 Td[(f(G;) f(G)=(): Weusequadraticpotentialfunctionagain.We'llshowthersttermontherighthandsideofEquation 2 canbeboundedusing().Forsome2Co(ML)and>,usingthedenitionof()wehave:2A(t))]TJ /F3 11.955 Tf 11.96 0 Td[(D(t)TQ(t)2)]TJ /F3 11.955 Tf 11.96 0 Td[(TQ(t)2)]TJ /F7 11.955 Tf 11.96 0 Td[((+()e)TQ(t))]TJ /F3 11.955 Tf 35.87 0 Td[(()eTQ(t): 69

PAGE 70

AtstationaritywehaveE[(Q(t))]=0.andE[(A(t))]TJ /F3 11.955 Tf 11.95 0 Td[(D(t))T(A(t))]TJ /F3 11.955 Tf 11.96 0 Td[(D(t))]=B.UsingEquation 2 ,wehave:0=E[2(A(t))]TJ /F3 11.955 Tf 11.96 0 Td[(D(t))TQ(t)]+B0)]TJ /F7 11.955 Tf 21.92 0 Td[(2()E[eTQ(t)]+BE[eTQ(t)]B 2(): (4)Thenewparameter,(),providesabetterboundingmechanismforthestationaryqueuesumvaluewhichinturncanbeusedtobounddelaythroughLittle'sLaw.Nextwedetailontheissuespertainingtothecomputationof().Westartwithaformulationthatequivalentlydenes().Anoptimizationproblemtocalculate()canbeposedasfollows,consideranarrivalratevector2o:max()s.t.ML+()e;NXi=1i=1;i0: 4.2.1ChallengesinComputing()Computing()isNP-hardthroughreductiontothefractionalcoloringproblem.Considerthefollowingreduction:givenagraphG,onecancompute(~e),whichisequalto1)]TJ /F3 11.955 Tf 12.52 0 Td[(f(G).Thusaproceduretocompute()canbeusedtosolvethefractionalchromaticnumberproblemforanygraphG.Alsonotethattheequality(~e)=1)]TJ /F3 11.955 Tf 12.65 0 Td[(f(G)holdsforonlyvector~eandisnotcorrectingeneral.Duetointractabilityoftheproblem;weprovideasimpleframeworktoestimate().Considerthefollowingequationsfor2o(G)thatrelatestoitsschedulingnorm(seeSection 2.7 forthedenition): 70

PAGE 71

minf1;NXiigf(G;)maxii:ItisNP-hardtocomputef(G;)ingeneral.AlsoapolynomialtimealgorithmtoprovideapproximationwithinN 2p logNimpliesP=NP[ 54 ].ConsequentlywewillconcentrateonaspecialcaseofarrivalvectorswherePNii<1.Nowconsiderthefollowingnumber()=1)]TJ /F25 7.97 Tf 6.59 5.98 Td[(PNii N.SinceNf(G)and1)]TJ /F12 11.955 Tf 12.42 8.96 Td[(PNii<1)]TJ /F3 11.955 Tf 12.42 0 Td[(f(G;)weconclude()<()thusB 2()canbeusedtoprovideanupperboundforstationaryqueue-sumunderMWS.Computationof()canbeperformedinlineartime.NotethatB 2()B 2()thustheexpectedqueue-sumboundprovidedinEquation 4 isatleastasgoodasthepreviousboundin[ 4 ].Nextwepresentanexamplewheretheimprovedboundisrealizedi.e.B 2()3thequeueboundisstrictlyimproved. 71

PAGE 72

Nextweinvestigatetheclassofgraphsforwhichthequeueboundsareidenticalusingtheoldparameterandthenewone.We'lldenotethisclasswithsymbolanddemonstratevariousclassesthatarecontainedwithin. Denition12. ThegraphclasscontainsallgraphsGsuchthat82o(G):()=(). Lemma32. AgraphGisinifandonlyifintheLinearProgrammingformulationoff(G)alloptimalsolutionshaveML=e. Proof. SupposeagraphGisin.Then82o(G):()=().AssumethereexitsanindexisuchthatanoptimalsetofdecisionvariablesforLPformulationof(seeDenition 1 )f(G)has(ML)i>1.Leti=(ML)i)]TJ /F6 7.97 Tf 6.58 0 Td[(1 f(G).Nowconsiderthearrivalvectorf:i=iandj=0ifj6=ig.Notethatf(G;+1 f(G)e)=f(G;1 f(G)e)=1whichimplies()=1 f(G).However()=1)]TJ /F5 7.97 Tf 6.59 0 Td[(i f(G)<(),contradictingtheassumptionGisin.Thusalloptimalsetofdecisionvariablesforf(G)satisesML=e.Supposeforalloptimalsolutionsoff(G)foragraphGhasML=e.AssumeG6=thenforsome2o(G)()>()orequivalentlyf(G;)+f(G;()e)>f(G;+e()).Howeverforoptimalsolutionsoff(G;ke)fork>0thereexistsnoslackintheinequalitiesoftheformMLke.Thusforall>0wehavef(G;)+f(G;()e)=f(G;+e())contradictingtheassumptionG6=. Werstneedtointroducesomedenitions(see[ 55 ]foramoredetaileddiscussion).Weconsiderundirectedgraphswithnoloopsandnomorethanoneedgebetweenanytwodifferentnodes,i.e.,thesimplegraphs. Denition13. AnisomorphismfromagraphG=(VG;EG)toagraphH=(VH;EH)isabijectionf:VG!VHsuchthat(u;v)2EGifandonlyif(f(u);f(v))2EH. Denition14. AnautomorphismofagraphGisanisomorphismfromGtoG. Denition15. AgraphG=(V;E)isvertex-transitiveifforeverypairu;v2Vthereisanautomorphismthatmapsutov. Lemma33. Theclassofvertextransitivegraphsiscontainedwithin. 72

PAGE 73

Proof. Firstweshowforvertextransitivegraphsinalloptimalsolutionsoff(G),ML=ewhereisthevectorofoptimaldecisionvariables.LPformulationoff(G)isgiveninDenition 1 :ThedualLPwhichisknownasfractionalweightedcliquenumberisthesolutionofthefollowingLP:wf(G;)=maxTs.t.(ML)Te;i0:LetGbeavertextransitivegraph.Forvertextransitivegraphsitisgivenin[ 12 ]thatwf(G)admitsanoptimalsolutionwherealli=1 (G)i=1;:::;N.Bycomplementaryslacknesstheorem,sincei>0wehaveatleastoneoptimalsolutiontotheprimalproblemwhereML=e.Assumethereexistsanotheroptimalsolution,i,forwhich(MLi)i>1.Nowconsidernodevi;vj2V:j6=i.Bythedenitionofvertex-transitivity,thereisanautomorphismfthatmapsvitovjandforthatmappingthereisanewsetofoptimaldecisionvariablesijforwhich(MLij)j>1.Nowifweapplytheautomorphismtogeteachnodekandgetikfork=1;:::;Nwegetacollectionofdecisionvariablesforeach(MLik)k>1.ConsiderthesumMLPNk=11 Nik>ehoweversinceeachikisoptimal;PNi=11 NeTik=f(G).ThusthevectorPNi=11 Nikcansatisfyconstraintsforf(G)andcanbescaleddownwithinconstraintstoachieveanewsolutionf(G)suchthatf(G)
PAGE 74

ACyclegraphwith5nodes BGraphwithanasymmetrictopologyFigure4-2. Vertextransitivitydoesnotrequireorimplymembershipinclass.A)Cyclegraphwith5nodes,C5,isvertex-transitiveandbelongstoclass.B)Thegraphdisplayedisnotvertex-transitivebutbelongstoclass. Figure4-3. IllustrationofK-cliqueandanodelooselyconnectedforK=6. ConsideragraphthatconsistsofaK-clique(forKinpositiveintegers)andasinglenodeconnectedtoonlyoneofthenodesinK-clique.Figure 4-3 illustratesthisgraphwhenK=6.WewillrefertoitasGinthefollowingdiscussion.Nowconsiderarrivalratevector=f1)]TJ /F6 7.97 Tf 15.56 4.7 Td[(2 K;0;:::;0gwherethelooselyconnectednoderepresentsthelinkhavingnon-zeroarrival(indexedwiththerstentry).Firstwenotef(G;)=1)]TJ /F6 7.97 Tf 15.41 4.71 Td[(2 Kandf(G)=K.Nextwehave()=1 K,toseethatobservefor=f1)]TJ /F6 7.97 Tf 15.4 4.71 Td[(2 K;1 K;:::;1 Kgwehavef(G;)=1and=+1 Ke.Thus()=1)]TJ /F6 7.97 Tf 6.59 0 Td[((1)]TJ /F20 5.978 Tf 9.22 3.25 Td[(2 K) K=2 K2where()=1 K.Thuswehave()K 2(),whichshowsthedifferencecanscalewithnodecountlinearlyunderthegivencircumstances. 74

PAGE 75

4.3TightnessofDelayParameters 4.3.1DerivingDelayBoundsforaGeneralizedMWSPolicyUsingaQuadraticPotentialFunctionIn[ 15 ]authorsproposedamodiedMWSpolicywherequeue-lengthsarescaledwithwi2R+thusactivationvectorm(t)attime-slottsatisesm(t)Tz=(G;z)wherezi=wiQi(t).Inthissectionwewilljustifytheparameterchoicemadein[ 15 ].Ouranalysisreveals,usingthepotentialfunctionfrom[ 15 ],theweightfunctionchosenoptimizesthederivedqueue-sumbound.Werevealtheservicevectorthatiscomparedwiththeactivationandthequeueweightsaretwosetsofdecisionvariablesareintegraltotheunderlyingoptimization.Inthissectionweperformoptimizationoverbothsets.HoweverinSection 4.3.2 wewillrestricttheweightstotheform~ewhilederivingasingleparameterperformancecharacterizationofdelayintraditionalMWSpolicy.Considerany2Co(ML)suchthat>andthefollowingLyapunovfunctionV:V(Q(t))=1 2NXi=1wiQ2i(t):Potentialdifferencebetweentwotime-slotscanbecalculatedas:V(Q(t))=1 2NXi=1wiQi(t+1))]TJ /F3 11.955 Tf 11.96 0 Td[(Qi(t)Qi(t+1)+Qi(t)=1 2NXi=1wiAi(t))]TJ /F3 11.955 Tf 11.95 0 Td[(Di(t)2Qi(t)+Ai(t))]TJ /F3 11.955 Tf 11.95 0 Td[(Di(t)=NXi=1wii)]TJ /F3 11.955 Tf 11.96 0 Td[(Di(t)Qi(t)+1 2jVjXi=1wiAi(t))]TJ /F3 11.955 Tf 11.95 0 Td[(Di(t)2NXi=1wi(i)]TJ /F3 11.955 Tf 11.95 0 Td[(i)Qi(t)+1 2jVjXi=1wiAi(t))]TJ /F3 11.955 Tf 11.96 0 Td[(Di(t)2: (4)AcommonwaytoboundE[PjVji=1Qi(t)]istonoteE[V(Q(t))]=0atstationarityandrelatetermPNi=1wi(i)]TJ /F3 11.955 Tf 12.11 0 Td[(i)Qi(t)toPNi=1Qi(t).Inthefollowingderivationtheonlyconstraint 75

PAGE 76

onqueuesizesistheyarerequiredtobenon-negativeintegerswhichindicatesthenalboundshouldbefreefromtermsQi(t).ThesecondterminEquation 4 ,wiAi(t))]TJ /F3 11.955 Tf 12.12 0 Td[(Di(t)2,playsapassiveroleinthederivationoftheboundssinceitisgenerallyupper-boundedeasily.LetBi=1 2i+E[A2i(t)])]TJ /F7 11.955 Tf 11.96 0 Td[(22ithenusingEquation 4 ,wehave:E[1 2jVjXi=1wi(Ai(t))]TJ /F3 11.955 Tf 11.96 0 Td[(Di(t))2]=NXi=1wiBi:Denei=wi(i)]TJ /F3 11.955 Tf 12.36 0 Td[(i)thenouraimistondisuchthatatstationarityforallQi(t)2ZN+,whereZN+isthesetofnon-negativeintegers,wehavePNi=1Qi(t)PNi=1iQi(t)PNi=1wiBi.Thatinturnreducestotheoptimizationproblemofmaximizingthequeue-sumundertheimpliedconstraintswhichcanbeposedas:Q=maxE[NXi=1Qi(t)]s.t.E[NXi=1iQi(t)]NXi=1wiBi:Ifwecanndaconstantk1suchthatkQPNi=1iQi(t)wecanconcludeourexpectedqueue-sumboundwillbeatleastasgoodasPNi=1wiBi k.Toachievethesmallestkwestartwiththefollowingsteps:E[NXi=1iQi(t)]NXi=1wiBi(minii)E[NXi=1Qi(t)]NXi=1wiBiE[NXi=1Qi(t)]PNi=1wiBi minii:Tondthebestpossiblestationaryqueueboundweneedtosolvethefollowingoptimizationproblemwithdecisionvariableswand: 76

PAGE 77

min;wPNi=1wiBi miniis.t.>;2Co(ML);w0:Theproblemabovecanberewrittenasatwoleveloptimizationproblem,rstaminimiza-tionoverallweightsandsecondaminimizationoverallservicevectors,asfollows:minminwPNi=1wiBi miniis.t.>;2Co(ML);w0:Wecandothefollowingscalingwithoutaffectingtheoptimalvaluesofthedecisionvariables(thoughtheoptimalvalueoftheoptimizationproblemwillbescaled)minii=miniwi(i)]TJ /F3 11.955 Tf 11.95 0 Td[(i)=1,thenwehave:minminwNXi=1wiBis.t.>;2Co(ML);w0;miniwi(i)]TJ /F3 11.955 Tf 11.95 0 Td[(i)=1:Whenisxedweget:minwNXi=11 2wii+E[A2i(t)])]TJ /F7 11.955 Tf 11.96 0 Td[(22is.t.w0;miniwi(i)]TJ /F3 11.955 Tf 11.95 0 Td[(i)=1:Nowsinceminiwi(i)]TJ /F3 11.955 Tf 12.21 0 Td[(i)=1isaconstraint,ifwesetwi=1 i)]TJ /F5 7.97 Tf 6.59 0 Td[(ithatconstraintwillbesatisedforallindicesianditwouldbeanoptimalsolutiontothegivenoptimizationproblem 77

PAGE 78

sincetheconstraintswillallbesatisedwithequality.NowpluggingthenewdenitionofwtotheoriginalproblemwehaveminNXi=1i+E[A2i(t)])]TJ /F7 11.955 Tf 11.95 0 Td[(22i 2(i)]TJ /F3 11.955 Tf 11.95 0 Td[(i) (4)s.t.>;2Co(ML); (4)asthetightestqueueboundthatisofferedbytheframeworkthatcanbederivedusingthequadraticpotentialfunctionchosen.Notethatthepotentialfunctionisanaturalone,sinceitcorrespondstothesumofintegrationsoftheweightfunctions.Nextwewillstudythisclaimindetail. 4.3.2EstablishinganIntuitiveMeaningfor():AStepTowardsSingleParameterDelayCharacterizationofMWSPolicyThissectionderivesthedelayperformanceofMWSpolicyusingthetechniquesemployedinSection 4.3.1 .MWSpolicyreducestothegeneralizedalgorithmwiththeidenticalqueueweights,wi=wj8i;j.UsingthesamepotentialfunctionfromSection 4.3.1 wewillconductasimilarderivation.Ifwesettheweightstobeidenticali.e.:w=~ewehavethefollowingoptimizationproblem:minPNi=1Bi minii)]TJ /F3 11.955 Tf 11.96 0 Td[(is.t.>;2Co(ML):Byomittingtheconstantfactors,avariationoftheproblemabovecanberewrittenas(argminispreserved): 78

PAGE 79

maxminii)]TJ /F3 11.955 Tf 11.96 0 Td[(is.t.>;2Co(ML):Theoptimizationproblemabovewillbesatisedbyavectorontheboundaryofthecapacityregion.Thuswecanmodifytheconstraintsaccordingly:maxminii)]TJ /F3 11.955 Tf 11.96 0 Td[(is.t.>;ML=;MLXi=1i=1;i0:Let=minii)]TJ /F3 11.955 Tf 11.95 0 Td[(i.Thentheoptimizationproblemcanberewrittenas:maxs.t.ML)]TJ /F3 11.955 Tf 11.95 0 Td[(>e;MLXi=1i=1;i0;whichistheequivalenttotheLPformulationof().Asaresult,whenwerestrictourselvestotheunweightedcaseweobtain()asthetightestsingle-parameterperformancecharacteriza-tion.Anothervariationofthequestionaddressesaboveiswhatifwerestrictthevectorstobeoftheform+efor>0. 79

PAGE 80

min=+e;>0minwPNi=1wiBi miniis.t.>;2Co(ML);w0:min>0minwPNi=1wiBi miniwis.t.+e2Co(ML);w0:ThescalingargumentusedabovefortheminiiThusw=eisoptimalforminwPNi=1wiBi miniwi.Now,againbyomittingtheconstantterm,theoptimizationproblemcanbetranslatedto:min1 s.t.+e2Co(ML);w0;whichwithminorarrangementsisequivalenttooptimizationproblemofnding()fromSection 4.3.1 .Theanalysisconductedaboveestablishesthesignicanceof();withthenaturalchoiceofapotentialfunction,()providesthetightestboundthatcanbeachievedthroughquadraticpotentialfunctionframework.In[ 15 ]authorsprovetheboundsderivedusinggeneralizedMWSpolicywiththeweightselectioninEquation 4 with 4 areatleastasgoodastheboundsthatareachievedbyarandomizedpolicyusingaprobabilitydistributionofactivationoverthesetofschedules.IntuitivelyoptimalinEquation 4 with 4 actsastheaverageservicerateachievedoverthelinks,whentheweightvectorispickedtobe1 )]TJ /F5 7.97 Tf 6.58 0 Td[(.SimilarlyitispossibletoreasonovertheservicerateofMWSpolicywhichis+(). 4.4ResultsandDiscussionInthischapterweattempttoanswerthequestion:WhatisthebestparameterofchoicetocharacterizethedelayperformanceofMWSpolicy?Althoughthereareearlierattemptstobound 80

PAGE 81

delaythroughstationaryqueue-sumsuchasin[ 48 ],[ 26 ],[ 15 ],andourattemptsinChapter 3 ,ouremphasisinthischapterisnotonlyontheboundbutalsoonthetightnessoftheboundsprovided.WiththatendinviewwerstintroduceanewparametertobounddelayofMWSpolicy,thatoutperformspreviousresults.Nextweprovethecomputationofthenewparameterisintractableandprovideasimplemethodtoestimateit.Tocomparetheoldandnewparametersonthedelayboundweprovideanexamplethatdemonstratesanimprovementtoberealized.Theperformancecomparisonbetweentheoldandnewparameterisalsoconductedanditisshownthattheboundswiththenewparametercanbesuperiortoalinearscale,wherelinearityiswithrespecttothelinkcountinthewirelessnetwork.Thedistinctnessoftheoldandthenewparameter,however,isnotrealizedonallgraphclasses.Weidentifynecessaryandsufcientconditionsforagraphtohaveidenticalvaluesforoldandnewparameters.Asaresultweprovefortheclassofvertex-transitivegraphsbothparametersassumethesamevalue.Thesecondpartofthechapterdealswithndingtightdelayboundsusingquadraticpotentialfunctions.Weshowtheweightschosenin[ 15 ]areoptimalwhenthequadraticpotentialfunctionisoptimizedoverallpossibleservicevectors,andallpossibleweightselections.Nextweutilizethesameframeworktoshow,thenewparameterintroducedisalsooptimalthroughthesamemethodology.AsaresultofourstudywebuiltanintuitionontheservicevectorthatisachievedbybothgeneralizedandnativeversionsofMWSpolicy. 81

PAGE 82

CHAPTER5GENERALIZINGMWSPOLICY 5.1BackgroundInthischapterwestudyageneralizationofMWSpolicy.ThetraditionalMWSpolicy[ 52 ]usesthequeuelengthvectortocomputeWMISforactivation.Althoughthatapproachensuresstability,itisnotimmediateifotherperformancemeasures,e.g.delayorfairnessamonglinks,areoptimized.In[ 15 ]authorsproposeageneralizationofMWSpolicyusingpositiveconstantstoscalequeuesizes.Asaresult,delayboundsareimproved.Thuswhenasecondarysetofobjectivesareinquestionageneralizationmightoutperformthenativealgorithm.Thischapterfurtherextendsthegeneralizationin[ 15 ]byreplacingtheconstantweightswithgeneralfunctions.Inotherwordsateachtimeslotthequeueweightiscomputedusingafunction,whichmightbedistinctfordifferentqueues.Ourmethodologyextendstheframeworkpresentedin[ 27 ],whichstudiesthesufcientconditionsonthosefunctions.Thevalueofthisgeneralizationistwofold.FirsttheforemostdifcultyinMWSpolicyliesinsolvingtheWMISproblemandcomputingweightswitharbitraryfunctionsdonotfurthercomplicatethistaskunlessthefunctionsitselfarehardtocompute.Secondanoptimizationforasecondaryparametersetisdoneoverallthroughputoptimalfunctions.Weneedtoidentifythefeasiblesetoftheoptimizationtostartsuchaprocedure.Alsowithageneralenoughframeworkwecandirectlyconcludethealgorithmpresentedin[ 15 ]isimmediatelythroughputoptimal.Althoughthedelayboundspresentedin[ 15 ]isatleastasgoodasanyrandomizedalgorithmthatusesaprobabilitydistributionoverthesetofpossibleschedules,itrequirestomeasurethearrivalrateandsolveanoptimizationproblemusingtheestimatedarrival.Theproblemwiththisapproachisthatmeasuringthearrivalwithprecisionrequiresalongperiodofsamplecollection(thesamplesizecanbeboundedusingChebyshev'sinequality),whichmightevenbelongerthanthetimenetworkreachesstationarityfromtheinitialstate. 82

PAGE 83

5.2ChoosingWeightFunctioninMWSforStabilitySystemmodelinthischapterfollowscloselyfromchapter 3 .Wereferreadertosection 3.2 fordetailedsystemmodelandnotations.Notethatthereareminordifferencesintroducedbuttheyareunderlinedinthediscussions.In[ 52 ]authorsproposedtheMWSpolicyforthroughoutoptimality.Inthissectionwewillstudyageneralization.Wedeneaf()-MWSscheme,foravectorfunctionf:RN!RN,tobeapolicywhoseactivationvectors,m(t),foralltimeslotsttosatisfy:NXi=1fi(Qi(t))mi(t)maxm2MLNXi=1fi(Qi(t))miAlthoughMWSpolicyisthroughputoptimal,itisnotimmediatelyclearifitispossibleoptimizeotherparametersofinterestusingtheMWSschemewithdifferentweightfunctions.In[ 27 ]authorsprovidedasufcientconditionforweightfunctionstobethroughputoptimalunderMWSschemes.Wewillreneandextendtheirstatementinthefollowingdiscussion.In[ 50 ]authorscommentedthat-MWSschemes(i.e.:8ifi(x)=x)provideoptimalsteadystatedelayboundswhen!0.Laterin[ 15 ]authorsprovedthatthefunction8ifi(x)=wixcanprovidestationaryexpectedqueuesumperformancethatsurpassesanyrandomizedpolicyassigningconstantactivationprobabilitiestotheindividualschedules.Inwhatfollowswewillgeneralizetheresultin[ 27 ]toageneralizedinterferencemodelusinggraphs.Alsotheresultin[ 27 ]assumesBernoulliarrivalprocessforwhichA(t)2f0;1gandwegeneralizedtheschemeforA(t)2f0;g.ConsiderthepotentialfunctionF(Q(t)),PNi=1Fi(Qi(t))whereFi(x)=Fi(0)+Zx0fi(u)du:Let(t),A(t))]TJ /F3 11.955 Tf 11.97 0 Td[(D(t),thus)]TJ /F7 11.955 Tf 9.3 0 Td[(1(t).WeassumethesecondmomentofthearrivalprocessisboundedbyCasinChapter 3 .Nowtheexpectedpotentialdifferencebetweentwoconsecutivetimeslotscanbecomputedas: 83

PAGE 84

E[F(Q(t))jQ(t)]=E[F(Q(t+1)))]TJ /F3 11.955 Tf 11.96 0 Td[(F(Q(t))jQ(t)]=E[F(Q(t)+(t)))]TJ /F3 11.955 Tf 11.95 0 Td[(F(Q(t))jQ(t)]=NXi=1E[Fi(Qi(t)+i(t)))]TJ /F3 11.955 Tf 11.95 0 Td[(Fi(Q(t))jQ(t)]=NXi=1E"Zi(t)0fi(Qi(t)+u)dujQ(t)#: (5)GivenRa0(a)]TJ /F3 11.955 Tf 11.96 0 Td[(x)f0(b+x)dxwecandointegrationbyparts:Zbaudv=[uv]ba)]TJ /F12 11.955 Tf 11.96 16.27 Td[(Zbavdu: (5)Leta)]TJ /F3 11.955 Tf 11.95 0 Td[(x=uandf0(b+x)dx=dvinEquation 5 ,thenwehave;Za0(a)]TJ /F3 11.955 Tf 11.95 0 Td[(x)f0(b+x)dx=[(a)]TJ /F3 11.955 Tf 11.95 0 Td[(x)f(b+x)]a0+Za0f(b+x)dx:=[(a)]TJ /F3 11.955 Tf 11.95 0 Td[(x)f(b+x)]a0+Za0f(b+x)dx:=)]TJ /F3 11.955 Tf 11.95 0 Td[(af(b)+Za0f(b+x)dx: (5)UsingEquation 5 inEquation 5 theexpectedpotentialdifferencebecomes:E[F(Q(t))jQ(t)]=NXi=1E"i(t)fi(Qi(t))+Zi(t)0(i(t))]TJ /F3 11.955 Tf 11.96 0 Td[(u)f0i(Qi(t)+u)dujQ(t)#:Letf(Qi):=maxu2fQi)]TJ /F6 7.97 Tf 6.59 0 Td[(1;Qi+gjf0(u)jthen: 84

PAGE 85

E[F(Q(t))jQ(t)]Ef(Q(t))T(t)jQ(t)+NXi=1E2i(t) 2fi(Q(t))jQ(t) (5)Foreachqueuei,functionDi(t)indicateswhetherornotlinkireceivedserviceattimeslott.Notethat:Di(t)=8>><>>:1;ifQi(t)1andiisscheduled.0;otherwise.Althoughthediscrepancybetweenactivationandbeingabletoserveapacketatatimeslotisseeminglyaminorone,nextwe'llshowduetothisdiscrepancythesufcientconditionin[ 27 ]needsamodication.Thediscrepancybetweenactivationanddeparturecanbeboundedasfollows:f(Q(t))TD(t)+(G;f(Q(t)));where=maxm2MLNXi=1fi(0)mi:ConsideringtherstterminEquation 5 :Ef(Q(t))T(t)jQ(t)=Ef(Q(t))TA(t))]TJ /F3 11.955 Tf 11.96 0 Td[(f(Q(t))TD(t)jQ(t))]TJ /F3 11.955 Tf 21.91 0 Td[(()NXi=1fi(Qi(t))+:Notethat: 85

PAGE 86

2i(t)=(Ai(t))]TJ /F3 11.955 Tf 11.95 0 Td[(Di(t))2=A2i(t))]TJ /F7 11.955 Tf 11.96 0 Td[(2Ai(t)Di(t)+D2i(t)C+1:Returningtodrift:E[F(Q(t))jQ(t)])]TJ /F3 11.955 Tf 21.92 0 Td[(()NXi=1fi(Qi(t))++(C+1) 2NXi=1fi(Qi(t)): (5)NowwewillstateFoster'scriteriawhichprovidesaneasyguidancetoestablishpositiverecurrenceoftheunderlyingMarkovchain.NotethatstabilityandpositiverecurrenceoftheunderlyingMarkovchainareequivalentconditions. Theorem5.1. (Foster'scriteria).SupposeMarkovchainfXngisirreducible.Supposethereexistsalower-boundedfunctionV:E!R,some>0,andanitesubsetE0ofEsuchthatE[V(Xn+1))]TJ /F3 11.955 Tf 11.95 0 Td[(V(Xn)jXn=i])]TJ /F3 11.955 Tf 21.92 0 Td[(;ifi=2E0E[V(Xn+1)jXn=i]<1;ifi2E0:Then,theMarkovchainispositivepositiverecurrent.Letkxk1:=Pni=1jxij.Assumef0i()iscontinuousforalli,x2RN+andthefollowingconditionholds:limkxk1!1PNi=1fi(xi)+maxm2MLPNi=1fi(0)mi PNi=1fi(xi)=0: (5)ThentheresultingMarkovchainispositiverecurrentunderpoliciessatisfyingthef()-MWSscheme.TheoutlineoftheproofincludescombiningTheorem 5.1 andEquation 5 86

PAGE 87

Itispossibletomeettheconditionbyrestrictingtheindividualfunctionstothefollowingclass.If8ilimx!1fi(x) fi(x)=0;fi(0)=0;thenconditioninEquation 5 holds.Alargeclassoffunctionssatisestheconditionabove.fi(x)=log(x);fi(x)=xkk2R+arefewexamples.Howevernotallincreasingfunctionsareinthatclass.Forexamplealthoughexponentialfunction,f(x)=exp(x)isincreasing,itdoesnotsatisfytheconditionabove.Alsotheslightdifferencewiththestatementin[ 27 ]rulesoutsomefunctionsthatwerenotruledoutpreviously,e.g.:fi(x)=xkk2fR)]TJ /F13 11.955 Tf 11.96 0 Td[(R+g.Nextwewilldiscussthemaintheoremin[ 27 ]andtheeffectsoftheoverlookedterm: Lemma34. (Theorem1in[ 27 ])Assumethatf:R!Risnon-negativeandcontinuouslydifferentiableonR+,andthatlimx!1f(x) f(x)=0.Then,foranyBernoullii.i.d.admissibletrafcforqueuesi,f(x)-MWSsatises:E[Qi(t)]<1Notethatfunctionssuchasf(x)=1 x;f(x)=)]TJ /F3 11.955 Tf 9.3 0 Td[(xsatisfytheconditionprovidedinLemma 34 wherethef()-MWSschemefailstoprovidethroughputoptimalitywhenthosefunctionsareused.Theformerfunctionfailstoprovidedriftduetotermbeingindenitewhenx=0andlatterfunctionhasacorrespondingpotentialfunction,F(x)=)]TJ /F6 7.97 Tf 10.5 4.71 Td[(1 2x2,whichisnotlowerboundedarequirementimposedbytheFoster'sCriteria.Choosingaconstantfunctionfortheweightsi.e.:fi(x)=kk2R+willnotestablishstability,sinceitignoresarrivalpatterninmakingdecisions.Aslightmodicationwhere:f(x)=8>><>>:0;ifx=0k;ifx0; 87

PAGE 88

willalsofailtoprovidestabilityduetothethedecisionsbeingmadewithouttakingthequeuesizesintoaccountwhentheygrowlarge.ConsiderthefollowingnetworkconsistingofthreelinkshavingBernoulliarrivalpatternswiththeexpectedarrivalratesof=f1 2;1 3;1 2g. Figure5-1. Conictgraphofathreelinknetworkillustratingthelinetopology. Whentherearepacketstoserve,theMWSalgorithmwiththeabovementionedfunctionactivatesthemaximumcardinalityindependentsetonthegraphthatisinducedbynon-emptyqueues.ForndingthemaximumindependentsetMWSalgorithmisfreetobreaktiesarbitrar-ily,andstillensurestability.Thusthefollowingchoiceonfavoringlink1and3over2isduetothesamefreedom.Giventhatiftherearearrivalstoeithernode1or3(orboth);thealgorithmwillactivatetheschedulef1;3g.Iftherearenoarrivalsinthecorrespondingtime-slotto1and3,thenlink2willbescheduledfortransmission.SincearrivalprocessisBernoulli,theprobabilityofactivationforlink2is1 4.Howeverlink2hasanaveragearrivalrateof1 3thusitsqueuewillgrowindenitely.Anupperboundedincreasingfunctionsatisfyingf(0)=0ontheotherhandwillensurestability.Intuitivelythereisadifferencebetweentreatingqueuesizesequivalentlyafterathresholdandfavoringthelongerones,eventhoughthefavoringprocesslosesitsstrengthasthequeuesizesincrease. 5.3ResultsandDiscussionInthischapterweconsiderageneralizationofMWSpolicywherefunctionsofqueuelengthsareusedtoobtainweights,beforeWMISiscomputed.Weavoidmakingconstrainingassumptionsonthenatureoffunctions,onlyrequirethemtobeintegrableanddifferentiable.Werenethesufcientconditionsprovidedin[ 27 ]andshowthatpreviouslyoverlookedconditionssuchaspotentialfunctionbeinglower-boundedmightplayanimportantroleinfunctionprovidingstability.Aconnectionbetweenthefunctionvaluewhenqueuesareemptyandstabilityisalsoestablished.Accordingly,forachoiceoffunctiontoprovidestability,not 88

PAGE 89

onlyitisrequiredtodominateitsderivativeasymptoticallybutalsothefunctionvalueatpoint0.Alsothestatementshouldbetrueforsumsofentities,whenanon-uniformfunctionsetisbeinganalyzed.Therearestillopenquestionsfollowingourdiscussion.Thewiderangeoffunctionssatisfyingthroughputoptimalitybringsupthequestion:Isispossibletoachieveboundsin[ 15 ]usingareactivealgorithm(respondingtoonlycurrentqueuesizes)asopposedtomeasuringthearrivalpatterntoobtaininformationabouttheaveragerates?Thatresultwouldprovidesatisfactorydelaycharacteristicswhileeliminatingtherequirementtomeasureandestimatearrivalpatternaveragewhichisadifculttasktoaccomplishwithprecisionwithinoperationalconstraintsofawirelesscommunicationnetwork. 89

PAGE 90

CHAPTER6CONCLUSIONS 6.1SummaryofResultsWehaveprovidedanegativityresultonthecomputationalcomplexityofthroughputoptimalalgorithmsthathaspolynomiallyboundedtransientqueuesizeswithpertime-slotpolynomialcomputations.Ourresultsarederivedwithinadeterministicsettingwheretherandomizednatureofthearrivalsarenotfullycaptured.Howevercombiningwiththeresultsprovidedin[ 49 ],ourresultscanalsobeusedtoconcludeaboutsimplerandomizedcaseswithminormodication.Wefoundtheframeworkprovidedin[ 49 ]andourframeworktoposesdifferentstrengthsandeachcapturingdifferentsetofparametersbetter.Ourframeworkincorporatesthearrivalvector'sdistancetotheboundaryandaccordinglyshowsthereexistsapositivecounterparttoourargument:ifthereisWMISsolveravailable,itispossibletoachieveefcientqueuesizeswithpertime-slotpolynomialcomputations.However[ 49 ]eliminatedthedistancetotheboundarybyembeddingittotheapproximationratioandnallyachievingatightboundthatisderivedoutofinapproximabilityofmaximumindependentsetproblem.Asofnowwearenotawareofanymethodstoextendeitherframeworktocontaintheotheronewithin.WealsoprovidedboundsofapproximateMWSschedulingatsystemstationarity.Ourresultprimarilybuildsonagraphtheoreticalinequality,whichcapturesthedriftconditionsonthequadratic,andhigherorder,potentialfunctions.Inashrinkingportionofthecapacityregionitispossibletoderiveboundsforhigherordermomentsaswell.Weprovidedtheformulafortheshrinkingfactorintermsoftopologicalcharacteristicsoftheinterferencegraphandtheapproximationratiok.WehavedisplayedtheefciencyofLQFpolicywhenitiscomparedwithagenericapproximateMWSpolicy.OurresultfurtherjustiesusingLQFinpracticalsettings.AlthoughMWSpolicyprovidesthroughputoptimalityitsdelayperformanceisnotoptimal.In[ 15 ]ageneralizationofMWSisproposedtoachievetighterdelaybounds.WestartwithimprovingdelayboundsforMWSpolicyandprovideresultspertainingtoanewparameterthatdenestheimprovement.Theweightparameterchoicein[ 15 ]providesattractivequeue-sum 90

PAGE 91

bounds,weprovetheweightselectionisdoneoptimallyandusingthesameframeworkweshowournewboundistight.FinallywestudyageneralizationofMWSpolicywhereafunctionofqueuesizereplacesthequeuesizeincalculationofthequeueweights.Werenedpreviouslyprovidedresultsin[ 27 ],andexplainwhyaclassoffunctionssatisfyingtheconditionsthereinfailtoprovidestability.Wealsoconcentrateonthemaximumcardinalityschedulingalgorithmandprovideanexamplewhereisfailstostabilizeanetwork.Thatunderlineswithinthespectrumofaconstantfunctiontoanupper-boundedincreasingfunctionstabilityisachievedsomewhereinthetransition. 6.2FutureResearchGiventherevealedlimitationsofschedulingpolicieswebelievefutureresearchshoulddirecttowardsunderstandingthewirelessinterferencebettersothateitherthegraphbasedK-hopinterferencemodelwillbereplacedwithanewmodelthatadmitseasierper-time-slotsub-problemsortherealistictopologycharacteristicswillbeunderstoodtothepointwherethehardnessresultswillnolongerbeapplicable.Forexampleactionaldiscoveryoninterferencegraphsrevealingtheircontainmentwithinthefamilyofperfectgraphswillrulethehardnessresultsoutimmediatelyduetofractionalcoloringandindependentsetproblemsbeingpolynomiallysolvableontheperfectgraphs. 91

PAGE 92

REFERENCES [1] BaruchAwerbuchandFrankThomsonLeighton.Asimplelocal-controlapproximationalgorithmformulticommodityow.InIEEESymposiumonFoundationsofComputerScience,pages459,1993. [2] H.Balakrishnan,C.Barrett,V.Kumar,M.Marathe,andS.Thite.Thedistance-2matchingproblemanditsrelationshiptotheMAC-layercapacityofadhocnetworks.IEEEJournalOnSelectedAreasInCommunications,22(6):1069,2004. [3] BerkBirand,MariaChudnovsky,BernardRies,PaulSeymour,GilZussman,andYoriZwols.Analyzingtheperformanceofgreedymaximalschedulingvialocalpoolingandgraphtheory.InProc.ofIEEEINFOCOM,pages2213,March2010. [4] CemBoyaciandYeXia.Delayanalysisoftheapproximatemaximumweightschedulinginwirelessnetworks.2011. [5] A.Brzezinski,G.Zussman,andE.Modiano.Localpoolingconditionsforjointroutingandscheduling.InInformationTheoryandApplicationsWorkshop,pages499,2008. [6] AndrewBrzezinski,GilZussman,andEytanModiano.Enablingdistributedthroughputmaximizationinwirelessmeshnetworks:apartitioningapproach.InProceedingsofACMMobicom,pages26,2006. [7] PrasannaChaporkar,KoushikKar,andSaswatiSarkar.Throughputguaranteesthroughmaximalschedulinginwirelessnetworks.InProc.ofAllertonConferenceonCommunica-tion,ControlandComputing,pages28,2005. [8] AntonisDimakisandJeanWalrand.Sufcientconditionsforstabilityoflongest-queue-rstscheduling:Second-orderpropertiesusinguidlimits.AdvancesinAppliedProbability,38:505,2006. [9] A.Eryilmaz,O.Asuman,andE.Modiano.Polynomialcomplexityalgorithmsforfullutilizationofmulti-hopwirelessnetworks.InProceedingsofINFOCOM2007,pages499,May2007. [10] UrielFeigeandJoeKilian.Zeroknowledgeandthechromaticnumber.InProceedingsofCCC,page278,Washington,DC,USA,1996.IEEEComputerSociety. [11] MichaelR.GareyandDavidS.Johnson.ComputersandIntractability:AGuidetotheTheoryofNP-Completeness.W.H.Freeman&Co.,NewYork,NY,USA,1979. [12] ChrisGodsilandGordonRoyle.AlgebraicGraphTheory.Springer,April2001. [13] M.Grotschel,L.Lovasz,andA.Schrijver.Theellipsoidmethodanditsconsequencesincombinatorialoptimization.Combinatorica,1:169,1981. 92

PAGE 93

[14] MartinGrotschel,LaszloLovasz,andAlexanderSchrijver.GeometricAlgorithmsandCombinatorialOptimization,volume2ofAlgorithmsandCombinatorics.Springer,secondcorrectededitionedition,1993. [15] GaganRajGuptaandNessB.Shroff.Delayanalysisforwirelessnetworkswithsinglehoptrafcandgeneralinterferenceconstraints.IEEE/ACMTransactionsonNetworking,18:393,April2010. [16] PiyushGuptaandP.R.Kumar.Thecapacityofwirelessnetworks.IEEETransactionsonInformationTheory,46(2):388,2000. [17] B.HajekandG.Sasaki.Linkschedulinginpolynomialtime.IEEETransactionsonInformationTheory,34(5):910,Sep1988. [18] KlausJansen.Approximatestrongseparationwithapplicationinfractionalgraphcoloringandpreemptivescheduling.Theor.Comput.Sci.,302(1-3):239,2003. [19] KlausJansenandLorantPorkolab.Onpreemptiveresourceconstrainedscheduling:Polynomial-timeapproximationschemes.SIAMJ.Discret.Math.,20(3):545,2006. [20] L.JiangandJ.Walrand.ACSMAdistributedalgorithmforthroughputandutilitymaximizationinwirelessnetworks.Technicalreport,UniversityofCalifornia,Berkeley,2008. [21] LibinJiangandJeanWalrand.Adistributedcsmaalgorithmforthroughputandutilitymaximizationinwirelessnetworks.IEEE/ACMTrans.Netw.,18:960,June2010. [22] ChangheeJoo.Alocalgreedyschedulingschemewithprovableperformanceguarantee.InProc.ofACMMOBIHOC,pages111,2008. [23] ChangheeJoo,XiaojunLin,andN.B.Shroff.Performancelimitsofgreedymaximalmatchinginmulti-hopwirelessnetworks.InDecisionandControl,200746thIEEEConferenceon,pages1128,dec.2007. [24] ChangheeJoo,XiaojunLin,andN.B.Shroff.Understandingthecapacityregionofthegreedymaximalschedulingalgorithminmulti-hopwirelessnetworks.InProc.ofIEEEINFOCOM,pages1103,2008. [25] ChangheeJooandNessShroff.Performanceofrandomaccessschedulingschemesinmulti-hopwirelessnetworks.InProceedingsofIEEEINFOCOM,2007. [26] KoushikKar,XiangLuo,andSaswatiSarkar.Delayguaranteesforthroughput-optimalwirelesslinkscheduling.InProc.ofIEEEINFOCOM,pages2331,2009. [27] IsaacKeslassyandNickMcKeown.Analysisofschedulingalgorithmsthatprovide100%throughputininput-queuedswitches.InProc.ofAllertonConferenceonCommunication,ControlandComputing,2001. 93

PAGE 94

[28] S.KhannaandK.Kumaran.Onwirelessspectrumestimationandgeneralizedgraphcoloring.InProceedingsofIEEEINFOCOM,volume3,pages1273vol.3,Mar-2Apr1998. [29] SandiKlavzarandHong-GwaYeh.Onthefractionalchromaticnumber,thechromaticnumber,andgraphproducts.DiscreteMathematics,247:235,March2002. [30] P.R.KumarandS.P.Meyn.Stabilityofqueueingnetworksandschedulingpolicies.IEEETransactionsonAutomaticControl,40(2):251,February1995. [31] E.Leonardi,M.Mellia,F.Neri,andM.AjmoneMarsan.Boundsonaveragedelaysandqueuesizeaveragesandvariancesininput-queuedcell-basedswitches.InProc.ofIEEEINFOCOM2001,volume2,pages1095,2001. [32] BoLi,CemBoyaci,andYeXia.Arenedperformancecharacterizationoflongest-queue-rstpolicyinwirelessnetworks.InProc.ofACMMOBIHOC,pages65,NewYork,NY,USA,2009.ACM. [33] BoLi,CemBoyaci,andYeXia.Arenedperformancecharacterizationoflongest-queue-rstpolicyinwirelessnetworks.IEEE/ACMTrans.Netw.,19(5):1382,October2011. [34] XiaojunLinandShahzadaB.Rasool.Constant-timedistributedschedulingpoliciesforadhocwirelessnetworks.InProceedingsoftheIEEECDC,2006. [35] XiaojunLin,NessB.Shroff,andR.Srikant.Theimpactofimperfectschedulingoncross-layerratecontrolinwirelessnetworks.IEEE/ACMTransactiononNetworking,14(2):302,April2006. [36] J.D.C.Little.Aproofforthequeuingformula:L=W.OperationsResearch,9(3):383,1961. [37] JiapingLiu,YungYi,AlexandreProutiere,MungChiang,andH.VincentPoor.Maxi-mizingutilityviarandomaccesswithoutmessagepassing.Technicalreport,PrincetonUniversity,2008. [38] MarilynnLivingstonandQuentinF.Stout.Perfectdominatingsets.InCongressusNumerantium79,pages187,1990. [39] L.Lovasz.Ontheratioofoptimalintegralandfractionalcovers.DiscreteMathematics,13(4):383390,1975. [40] CarstenLundandMihalisYannakakis.Onthehardnessofapproximatingminimizationproblems.J.ACM,41(5):960,1994. [41] G.MergenandL.Tong.Stabilityandcapacityofregularwirelessnetworks.IEEETransactionsonInformationTheory,51(6):1938,June2005. 94

PAGE 95

[42] EytanModiano,DevavratShah,andGilZussman.Maximizingthroughputinwirelessnetworksviagossiping.SIGMETRICSPerform.Eval.Rev.,34(1):27,2006. [43] MichaelJ.Neely,EytanModiano,andCharlesE.Rohrs.Tradeoffsindelayguaranteesandcomputationcomplexityfornxnpacketswitches.InProc.ofCISS,2002. [44] M.J.Neely.Delayanalysisformaximalschedulinginwirelessnetworkswithburstytrafc.InProc.ofIEEEINFOCOM,pages6,April2008. [45] S.RajagopalanandD.Shah.Distributedalgorithmandreversiblenetwork.InProc.ofCISS,pages498,March2008. [46] SujaySanghavi,LocBui,andR.Srikant.Distributedlinkschedulingwithconstantoverhead.SIGMETRICSPerform.Eval.Rev.,35(1):313,2007. [47] S.SarkarandS.Ray.Arbitrarythroughputversuscomplexitytradeoffsinwirelessnetworksusinggraphpartitioning.IEEETransactionsonAutomaticControl,53(10):2307,November2008. [48] DevavratShahandMilindKopikare.Delayboundsforapproximatemaximumweightmatchingalgorithmsforinputqueuedswitches.InProc.ofIEEEINFOCOM,pages1024,2002. [49] DevavratShah,DavidN.C.Tse,andJohnN.Tsitsiklis.Hardnessoflowdelaynetworkscheduling,2009. [50] DevavratShahandDamonWischik.Optimalschedulingalgorithmsforinput-queuedswitches.InProc.ofIEEEINFOCOM,pages1,2006. [51] G.Sharma,R.R.Mazumdar,andN.B.Shroff.Onthecomplexityofschedulinginwirelessnetworks.InProc.ofACMMOBICOM,pages227,2006. [52] L.TassiulasandA.Ephremides.Stabilitypropertiesofconstrainedqueueingsystemsandschedulingpoliciesformaximumthroughputinmultihopradionetworks.IEEETransactionsonAutomaticControl,37(12):1936,Dec1992. [53] LeandrosTassiulas.Linearcomplexityalgorithmsformaximumthroughputinradionetworksandinputqueuedswitches.InProc.ofIEEEINFOCOM,pages533,1998. [54] LucaTrevisan.Non-approximabilityresultsforoptimizationproblemsonboundeddegreeinstances.InProceedingsofthethirty-thirdannualACMsymposiumonTheoryofcomputing,STOC'01,pages453,NewYork,NY,USA,2001.ACM. [55] DouglasB.West.IntroductiontoGraphTheory(2ndEdition).PrenticeHall,2000. [56] XinzhouWu,R.Srikant,andJamesR.Perkings.Schedulingefciencyofdistributedgreedyschedulingalgorithmsinwirelessnetworks.InProceedingsofIEEEINFOCOM,2006. 95

PAGE 96

[57] YungYi,AlexandreProutiere,andMungChiang.Complexityinwirelessscheduling:impactandtradeoffs.InProc.ofACMMOBIHOC,pages33,2008. [58] G.Zussman,A.Brzezinski,andE.Modiano.Multihoplocalpoolingfordistributedthroughputmaximizationinwirelessnetworks.InProc.ofIEEEINFOCOM,pages1139,April2008. 96

PAGE 97

BIOGRAPHICALSKETCH CemBoyacireceivedhisB.S.incomputerengineeringfromBilkentUniversity,Ankara,Turkeyin2005,andhisM.S.incomputerengineeringfromUniversityofFlorida,Gainesville,Floridain2011.Hisresearchinterestsareinperformanceandcomplexitycharacterizationofdelayinwirelesscommunicationnetworksviagraphtheoreticalmodels. 97