Traffic Engineering in Software Defined Networks

2024年3月7日发(作者：我可爱的小狗)

2013 Proceedings IEEE INFOCOM1TrafficEngineeringinSoftwareDefinedNetworksSugamAgarwal1MuraliKodialamBellLabs,Alcatel-LucentHolmdel,NJ,USA{muralik,lakshman@}anAbstract—SoftwareDefinedNetworkingisanewnetworkingparadigmthatparatesthenetworkcontrolplanefromthepacketforwardingplaneandprovidesapplicatallycentralizedcontrollerthathasaglobalnetworkviewisresponsibleforallthecontroldecisionsanditcommunicateswiththenetwrecentlyannounced[5]thatitisusingaSoftwareDefinedNetwork(SDN)tointerconnectitsdatacentersduetotheea,efficiencyandflexibilityinperformingtraffictstheSDNarchitecturetoresultinbettributionofthispaperisontheeffectiveuofSDNsfortrafficengineeringesicular,weshowhowtoleveragethecentralizedcontrollertogetsignificantimprovthattheimprovementsarepossibleeveninculatetheSDNcontroller’soptimizationproblemfortrafficengineeringwithpartialdeploymentanddevelopfastFullyPolynomialTimeApproximationSchemes(FPTAS),bybothanalysisandns-2simulations,theperformancegainsthatareaUCTIONSoftwareDefinedNetworking(SDN)isanewnetworkingparadigm[1],[13],[14]nctionalparationandtheim-plementationofcontrolplanefunctionsonparatecentralizedplatformshasbeenofmuchrearchinterestduetovariouxpectedoperationalbenefitinginterdomainroutingfromindividualroutersandusingalogicallycentralizedRout-ingControlSystemwaspropodin[2],[4]asameanstomakeroutingsystemsmoremanageable,tRouterarchitecture[12]propodthedisaggregationofroutersintopacketforwardingelements,withopenstandardizedinterfaces,spropodasameansforquickerintroductionofnetworkfunctionssuchastrafficengineering,newVPNfeatures,otoachievevariousothercost,operationalbenefi-architectingthecontrolplaneintoadisminationplaneandadecisionplanewaspropodin[10].Thedisminatonplanereliablydisributesinformationtothenetworkelementsandthedecisionplanemakesalldecisonsthataffectthenetwork.1WorkdonewhileatBellLabsThisre-architecturemakesiteasierforthedecisionplanetohaveaglobalviewofnetworksttoalltheaboveapproachesisthatanSDNcomprisoftwomaincomponents:•SDNController(SDN-C):Thecontrollerisalogicallycentralizedfunction[3],[8].trollersdeterminestheforwardingpathforeachflowinthenetwork.•SDNForwardingElement(SDN-FE):icforforward-ingthepacketsisdeterminedbyflow[14]isastandardizedinterfacethflowisinitiatedinthenetworkthefollowingactionsaretaken:(i)ThefirstpacketoftheflowisntbytheSDN-FEtotheSDNcontroller,(ii)TheforwardingpathfortheflowiscomputedbytheSDNcontroller(SDN-C),(iii)TheSDN-CndstheappropriateentriestoinstallintheforwardingtablesateachSDN-FEonthepathfromthesourcetothedestination,(iv)Allsubquentpacketsintheflcontrollerisresponsibleforpathsopossibletoimplementcentralizedorpartiallycentral-izedtraffitance,GooglehasbuiltaSDNwithOpenflowrouterstointerconnectitsdatacenters(G-Scale)[5].Googleixpectinganimprovementinnetworkutilizationof20-30%aswellasimproveddelayandlossperformance[5]fficengineeringproblemthatweconsiderismoti-vatedanetwork,notallthetraffiaybemultiplecontrollersfordifferentpartsofthequestiontheniswhetheritisstillpossibletodoeffectivetrafficengineeringwhenallthetraffipaper,weconsidertrafficengineeringinthecawheretofthenetworkdoeshop-by-hoproutingusingastandardroutingprotocollikeOSPF.978-1-4673-5946-7/13/$31.00 ©2013 IEEE2211

2013 Proceedings IEEE INFOCOM2TheobjectiveofthepaperistodevelopaSDNdeploymentschemethatcanadaptivelyanddynamicallymanagetrafficinanetworktoaccommodatedifferenttraffincontributionsinthispaperarethefollowing:•Flex Node2512116•••Toourknowledge,thisisthefirstpapulatetheSDNcontroller’soptimizationproblemandoutlineaFullyPolynomialTimeApproximationScheme(FPTAS)byanalysisandsimulatonstheconsiderablegainsindelayandlossperformanceevenwfixednumberofSDN-FEsweoutlineanalgo-rithmfordeterminingtheplacementoftheforwardingelements.317131014Flex Node4Controller9815Flex wardingElementTheSDN-FEsperformthefollowingfunctions:Forwarding:earemultiplenexthopsforagivendestination,thentheSDN-FEscansplittraffictothedestinationinapre-specifilativelyeasyforthecontrollertocomputemultiplenexthopsandloadtheroutingtabletotheSDN-FEs[12].Thereareveralwaysofsplittingtrafficonmultiplenexthops[15]whileensuringthatagivenfltheapproachesneedextrameasurementsanditisrelaoreintherestofthepaper,weassumethetheSDN-FEscansplittraffiassumethattheSDN-FEscanperformtrafficmeasurementsusingtechniquessuchasthoin[15].Measurement:TheroutingtableattheSDN-FEsismodifiedslightly,comparedtoastandardroutingtable,inordertoaidtraffiaticreprenta-tionshowingthedifferencebetweenastaatthereisanextraacketisprocesdbytheSDN-FEitdoesalongestprefiincredoneinordertodeterminetheamountoftraffithatnode15(IPaddress45.67.2.5)e11withinterfaceIPaddress43.2.34.7,umncorrespondingtothetraffictracksthenumberofbytesroutedfromnode2tonode15forthedestinationprefisyfortheSDN-FEtoalsocomputethetotaltraffipacelimitations,wedonotshowextensionstotheapproachinthispapertothepDESCRIPTIONWeconsideranetworkwhereacentralizedSDNcontofthenom-FEsforwardpacketsandthelogicforctiontoforwardingpackets,theSDN-FEsdosomesimpletraffitrollerusthistrafficinformationalongwithinformationdisminatedinthenetworkbyOSPF-TEtodynamicallychangetheroutingtablesattheSDN-FEsinordertoadapttochangingtraffitrolleralsoexploitsthefactthatitcanmakeco-ordinatedchangesacrossthedifferentSDN-FEsthatitcontrolstomanagechangingtraffiularnodesinthenetwork,alsoreferredtoasthenon-SDN-FEs,bridnetworkscenariowithtraditionalnetworksintermixedwithSDNsisalsoconsideredin[11].Ourobjectivesareconsiderablydifferentfromtheobjectivesin[11].Weassumethatnochangesaremadetothenon-SDN-FEs-FEs2,9,uthisnetwomethatalltualalgorithmthatarerunatthecontrollerisoutlinedinSection3.2212

2013 Proceedings IEEE INFOCOM3PrefixNodeNext

HopTraffic135.23/trollerTheSDN-Chasalltheroutinglogicanditcoordinatestrollerdoesthefollowingfunctions:Peering:TheSDN-Cpeerswiththeothernodesinthenet-workexchanginglinkweightsandothertopologyinformationusingOSPF-TE.(See[9]foranexample.)NotethatinOSPF-TE,thenodesorethecontrollerknowsthecurrentOSPFweightsaswellastheamountoftrafficflowoneachlink(averagedoversometimeperiod).RouteComputation:Thecontrollerisresputestheroutingtablestakingintoconsiderationtheroutingdonebynon-SDN-FEs(badonOSPFlinkweights),thetrafficatthelinks(derivedfromOSPF-TEinformationor)andthecurrenttrafficpattern(inferredfromthemeasurementsattheSDN-FEs).ThealgorithmforcomputingtheroutingtablefortheSDN-FEshastoensurethatroutingwidescribetheSDNcontroller’CONTROLLER’SPROBLthatthenet⊆NdenotethetofSDN-FEsandD=(e)andc(e)denotetheOSPFlinkweightandcapacityrespectivelyofalinke∈(e)toreprentthetrafficflflowonalllink∈sdtoreprentthetrafficratefromnodes∈Ntosomeothernoded∈NandWudtoreprentthetotalamountoftrafficfordestinationd∈NthateitheroriginatesorpassthroughSDN-FEu∈atingeneralWud≥-FEucanmeasureWudforueofTsdforallnodepairs(s,d)tingtableatnodeu∈H(u,d)rwords,NH(u,d)isthefiemainderofthispaper,weassumethatthenexthopisuniqueforallthenon-SDN-FEs,i.e,NH(u,d)hasonlyoneelementforallu∈hniquesinthispaperextenddirectlytothecawheretherealternateshortestpathsbetweentwonodesandtraffiatwhileNH(u,d)iscomputedbadonshortestpathsforallnodesu∈D,NH(u,d)canbetarbitrarilywhenu∈methatalllinkwethetreethatthatnodes2,9,atNH(6,13)=10,NH(1,13)=temple,node2cansplitthetraffictonostPathTreetoNode13Definition1:GivenatofSDN-FEnodesC,apaths=u0,u1,u2,...uk=dfromasourcenodestoades-tinationnodedwillbetermedfeasibleifforj=1,2,...,k,(uj−1,uj)∈Eanduj=NH(uj−1,d)ifuj−1∈blepathwhereu0,u1,...,edefinition,notethatapathisfeasibleifthenexthoptoagivendorewehavetoensurethatallthetrafficbetweensanddhastoberoutedonP∈mple,inFigure3,3−2−5−12−atthisisnottheshortestpathwhichis3−2−11−h3−6−11−13isnotadmissiblesincethenexthopfornode3whichisanon-SDN-FEhastobethenexthopontheshortestpath,whichisnode2.2213

2013 Proceedings IEEE INFOCOM4Definition2:Givenshortestpathroutingatthenon-SDN-FEs,trafficthatgoesfromsourcetodestinationwithouttran-sitingthroughaSDN-FEwillbereferredtoasuncontrollabletraffiourceofapacketisaSDN-FE,orifitpassthroughatleastoneSDN-FEbeforeitreachesitsdestinationthenthistrafficwillbecalledcontrollabletraffirwords,controllabletrafficcomprisofpacketsthatpasatleastanopportunityattheSDN-FEstomanipulatethepathofcontrollabletraffimplethetrafficfrom6to13isroutedbyOSPFalong6−10−13,andsinceneither6nor10areSDN-FEs,traffirast,trafficfromnode8to13passthroughnode9whichisaSDN-FEandhencethistraffifinition3:WesaythataSDN-FEu∈Cinjectsapacketif••ationoftheSDN-C’sProblemSincetheonlytrafficthatwecanmanipulateisthetrafficthatpassthroughtheSDN-FEs,wejustfocusonthetraffificIudisinjectedbySDN-FEu∈nlydosoalongoneoftheadmissiblepathsP∈(e)denotetheuncontrollableflowonlinke.(Notethatg(e)iasytocomputedifthesource-destinationtraffiedearlier,r,westillgivetheformulationbelowinordertomotivatetheactualdynamicroutingproblemsolvedbytheSDN-C).TheobjectiveoftheSDN-Cistoroutethecontrollabletraffiayandpacketlossatthelinksareincreasingfunctionsofthelinkutilizationandthereforeweuthelinkutilizationasasurrogateforthedelay/uralobjectiveforormulation,thevariablesarex(P),whichistheflhenumberofpathsinthenetworkcanbeexponentialinthenumberofnodesandarcs,erthispathtothemorecompactnode-arcformulationsinceitlend-Csolvesthefollowingoptimizationproblem:fficthatisinjectedbySDN-FEu∈Ctosomedestinationnoded∈ore,forallcontrollabletrafficthereisauniqueSDN-FEthatinjectsthistraffiattheSDN-FEmayormaynotbethesourceofthetraffifigure,thenumbernexttothenodereprentsthetraffimple,thetrafficfromnode1tonode13(T1,13)finition3,notethatthetraffialuesofTsdareknownforallsource-destinationpairs(s,d),thenthevalueofIudcanbecomputedasfollows:RemovethelinksgoingoutoftheSDN-FEsanfficthataccumulatesattheSDN-FEsisthetraffimpleinFigure3,I2,13=9,I9,13=13andI14,13=edearlier,ymeasurementsavailableattheSDN-CarethevaluesofWudwhichisthetrafficfordestinationdthatpassthroughnodeu∈C.12251minimizeθsubjecttog(e)+󰀄P:P󰀅ex(P)x(P)x(P)≤≥≥θc(e)∀e∈EIud∀u∈Cd∈N0∀P(1)(2)(3)󰀄P∈ndentlyRoutableTrafficattheSDN-FEsThefirsttofinequalitiensurethatthetotalflowonthelinkwhichisthesumoftheuncontrollableflow(reprentedbyg(e))andthecontrollableflow(whichisthecondsumtermontherighthandside)islessthantheproductofthemaximumlinkutilization(θ)andthecapacityofthelink(c(e)).•Thecondtofinequalitiensuresthatthetotalinjectedtrafficisroutedinthenetwork.•Thethirdtofinequalitiensuresthattheflimumvalueofθatiftheoptimumvalueofθ<1,eSDN-Csolvesthisoptimizationproblem,itiasytocomputethenexthoormulationabove,weassumedthatvaluesIudandg(e)ityboththequantitiesIudaswellasg(e)havetobecomputedbytheSDN-Cbadon•2214

2013 Proceedings IEEE INFOCOM5themeasingIudForeachdestinationd∈etheroutingorderR(d)firstnodeuinR(d)tIud=neunitofflowfromutodandtβv(u,d)bethefractionofthisunitflowthatreachesnodev∈hsuccessivenode󰀁winR(d)SetIwd=Wwd−u≺dwβw(u,d)neunitofflowfromwtodandcomputeβv(w,d)forallv∈evaluesofIudareknownforallu∈Cforalld∈N,weuthistocomputethevaluesoftheg(e)whichistheuncontrollabletrafficthatflowsonlinke∈doneasfollowsWeinjectoneunitofflowatnodeu∈Cfordestinationd∈Nandcomputesαe(u,d)whichisthefractionofthisunitfleknowIudforu∈C,wecancompute󰀄󰀄αe(u,d)Iud∀e∈E.g(e)=f(e)−u∈ingIudandg(e)DynamicallyTheonlymeasuredquantitiesthatareavailabletotheSDN-Care••Thelinkloadf(e)foralllink∈ntitiesWudforallu∈Cforalld∈hetwoquantities,theSDN-Chastocomputethevaluesofg(e)foralle∈EandIudforallu∈Cforalld∈fierafisallthenexthopsforallnodesinDandatalltheSDN-FEsitnotknowsallthenexthopsforthedestinationandthetraffifinition4:Givenadestinationdandthecurrentroutinginthenetwork,theroutingorderofthenodesinCwithrespecttothisdestinationdisdefinedasanorderingofthenodesinCdsuchthatifu∈Cappearsbeforev∈CinthislistthenthereisnotraffitetheroutingorderfordestinationnodedasR(d)andthefactthatuappearsbeforevinR(d)asu≺utingorderiswelldefinedforanydestinationnodedsincetherecannotbeanyroutingloopsinthetrafficflowingtodestinationd.(InfactitispossibletoorderallthenodesinthenetworknotjustthenodesinC,butweareinterestedonlyintheorderingofthenodesinC.)canotethatthereistraffiore9≺tingorderis(2,9,14).Otherorderingsarepossiblebutnode14shouldappearafternode9inanyordering.521211TheSDN-CnowknowsthevaluesofIudforallnodesu∈Cforalld∈Naswellasthevaluesofg(e)foralle∈atingtheDynamicRoutingProblemTheSDN-Croutestraffivalentproblemthatismoreconvenienttosolveistokeepthecapacitiesofthelinkfixedbutscaletheinjectedtrafficsothatitstillfioblemisthefollowing:maximizeλsubjectto󰀄P:P󰀅ex(P)x(P)x(P)≤≥≥c(e)−g(e)=b(e)∀e∈EλIud∀u∈Cd∈N0∀P(4)(5)(6)󰀄P∈raptimalλ>1thenthecurrenttrafficcanberouteattheoptimalsolutiontothisscalingprobleeofthefactthattheproblemhasanexponentialnumberofvariables,wecansolvertowritetheduallinearprogramtothedynamicroutingproblemshownabove,weassociatedualvariablesl(e)witheachlinkcapacityconstraint(4)andzudforthe2215

2013 Proceedings IEEE INFOCOM6demandconstraints(5).Thedualcannowbewrittenas󰀄e∈Eminimizesubjectto󰀄e∈Pb(e)l(e)l(e)≥≥≥zud∀P∈Pud∀u∈C∀d(7)10∀e∈E.(8)(9)inphaswhereineachprimalphaflowisroutedtoagivendestinationfromallSDN-FEsalongthelightestpermissiblepath(usingthedualvectorl(e)asthelinkweight).OnceeachSDN-FEuhasshippedaflowofIudtodestinationd,the(dual)weightsofthearcsonwhichflocessofaugmentingflowaorithmisgiveninmoredetailbelow:AlgorithmCOMPUTETHROUGHPUT:DL←0l(e)←δ/b(e)e∈ERsd←0∀(s,d)whileDL<1doforeachdestinationd∈Nd󰀈(u)=Iud∀u∈CwhileDL<1andd󰀈(u)>0forsomeu∈CdoPud:Shortestadmissiblepathusingl,∀uwithd󰀈(u)>0c=mine∈∪sPsdb(e)ρ(e)istheutilizationofe∈Eρ=max{1,maxe∈∪uPudρ(e)}f(u)=min{d󰀈(u),c}∀uu)Routef(ρflowfromeachutod.u)d󰀈(u)=d󰀈(u)−f(ρu)Rud=Rud+f(ρl(e)=l(e)(1+󰀨ρ(e))󰀁RecomputeDL=e∈Eb(e)l(e)endwhileendforendwhile󰀄󰀄u∈Cd∈NIudzudl(e)Assumethatwetl(e)tobetheweightoflinke∈E.(WeuthetermweightinsteadofcostinordertoavoidconfusionwiththeOSPFlinkcosts).Notefromthefirsttofconstraintszudisthelightestpathfromutod.(AgainweuthetermlightestpathtoavoidconfusionwiththeshortestpathusingOSPFcosts).LetLuddenotethelightestpathfromutodusingthelinkweightsl(e)lcannowbere-writtenas󰀄e∈Eminimizesubjectto󰀄󰀄u∈Cd∈Nb(e)l(e)IudLudl(e)≥≥10∀e∈E.(10)(11)Inotherwords,givenanynon-negativetoflinkweights󰀁bE(e)l(e)󰀁isanupperboundonthel(e),notethat󰀁e∈u∈Cd∈utlinethesolutionofthedynamictraffigtheDynamicRoutingProblemWeuaFullyPolynomialTimeApproximationScheme(FPTAS)sonforsolvingtheproblemasanFPTASinsteadofastandardlinearprogrammingproblemisthattheFPTASisverysimpletoimplementandrunssignificantlyfasterthanageneraSprovidesthefollowingperformanceguarantees:forany󰀨>0,thesolutionhasobjectivefunctionvaluewithin(1+󰀨)-factoroftheoptimal,andtherunningtimeisatmostapolynomialfunctionofthenetworksizeand1/󰀨.maldualalgorithmforourproblemworksasfollows:Thealgorithmfirstcomputesavalueδthatisafunctionofthedesiredaccuracylevel󰀨,lweightofeachedgee∈Eisinitializedtol(e)=b(δe).Theprimaldualalgorithmoperatesudλ=minRTudOutputλThenextresultgivestherunningtimeofthealgorithmandtheproofofthisresultisalmostidenticaltotheoneinKarakosatas[7].Theorem1:Set󰀂󰀃111−󰀨󰀈δ=1−󰀈m(1+n󰀨)󰀈thentherunningtimeoftheprimal-dualalgorithmisO(󰀨−2m2logO(1)m)marks:orithmfollowsinthesameveinasKarakostas[7].Thecorrectnessofthealgorithmaswellastherunningtimrehoweversomekeydifferencesintheimplementationofthealgorithm.2216

2013 Proceedings IEEE theKarakostas[7]paper,criticallyimportantforussincewecancomputetheroutingfrheroutingatthenon-SDN-FEsisbadonthedestination,iteration,wehavetocore6,wegivetheOSbersnexttothelinksreprentsthedualweights(notOSPFcosts).thatwearenowcomputingthelightestadmissiblepathstonode13fromalltheSDN-FEs.(2,9,14).SincetheadmissiblepathshavetoutheOSPFshortestpathatallthenon-SDN-FEs,odatthearcwith0.8fromnode2tonode13reprentsthepath2−11−13andthearcwithweight0.4fromnode2tonode13isthepath2−5−12−isnewgraphisformed,theligucedgraphisshownexplicitlyonlyforillustrativeearemaximizingthethroughputofthenetwork,thduetothefacr,mal-dualalgorithmtypicallyfidagainstloopsintheoptimalsolution,wepostprocesstheoptimalsolutionusingabreadth-finingtimeofdGraphforLightestPathtoNode130.10.511120.330.560.3130.2FElocationshavebeendetermined,theSDN-CsolvesthedynamicroutingproblemperiodicallytodeterminetheroutingoftrafficattheSDN-FEsbadonthetraffiticallyanysystemwithSDN-FEswilloutperformasysteticetheimprove-mentinimumlectionofnodeswilldependonthetraffifirstistopickthenodesindependentofthetrafficmatrixandthecondapproachistousomeestimateofthetraffimptedmethatweknowthenumberofSDN-FEsinthenetworkandwearegivenatentativetrafficmatrixTwhereTsdisthetrafficbetweennodess∈Nandd∈ualtrafficcan,andingeneralwill,deviatefromthistraffifinition5:ThethroughputofatrafficmatrixToveratofSDN-FEsCisdefinedasthelargestscalarλsuchthatλtethethroughputvaluebyλ(T,C).NotethatifC=∅correspondstotewhereC=edefinition,itiasytoeλ(T,∅)≤λ(T,C)≤λ(T,N)∀T,tethatλ(T,N)canbecomputedbysolvingastandardmaximumconcurrentfloregivenTandf,theobjectivethenistodetermineC:|C|=h0.5170.20.3100.20.31440.580.190.3150.50.3maxλ(T,C).INGTHELOCATIONOFSDN-FESGivenanetworktopology,thefistepofthealgorithm,we2217

2013 Proceedings IEEE INFOCOM8115nodesExodusAbovenet0.9thevaluesofTsdareknown,thethroughputproblemcanbewrittenasNormalized

Throughput0.8maximizeλsubjectto󰀄P:P󰀅e0.70.60.5x(P)x(P)x(P)≤≥≥c(e)∀e∈EλTsd∀s∈Nd∈N0∀P(12)0.4󰀄(13)(14)Fig.8.0.312345678P∈Psd9122Number of Flex NodesEffectofNumberofSDN-FEsSincetheproblemisstructurallythesameastheSDN-Csproblem,MENTALRESULTSWerantwogroupsofexperimentstochecktheeffectivenessofthealgorithmusingthefollowingthreetopologies:(i)The15nodetopologyshowninFigure1,(ii)TheExodus(Europe)pologyhas22nodesand74links,(iii)15nodetopologyallthelinkcaROCKETFUELtopologies,thelinkweightsaregivenandthelinkcapacitiesareassumedtobetheinverofthelinkcosts.(Wehaveconsolidatedmultiplelinksbetweentwonodesintoasinglelinkintheexperiments).firstisthestaticperformancePerformanceMeasurementTheexperimentswereperformedtocomputetheexpectedperformanceimprovementduetothedynamicroutingalgo-rithmifwearegivenatraffiheplotsforstaticperformancemeasurement,malizedthroughputforagiventofSDN-FEsCisdefinedasλ(T,C).λ(T,N)atforOSPFroutingwithnoSDN-FEsthevalueofC=∅.Weperformedtwotsofexperiments.•NormalizedThroughputversusNumberofSDN-FEsForallthreetopologies,emstobethereasosubquentexperiments,the•numberofSDN-FEsforthe15nnessoftheChoiceofSDN-FEsInthecondtofexperiments,wetestedthensitivityoftheperformancewithrespecttotherealtrafficmatrix(asoppodtothetrafficmatrixthatisudtopicktheSDN-FEs).Inotherwords,sincethelocationoftheSDN-FEsisfixedassumingsometrafficmatrix,wewantedtoehowwellthischoiceperformedifthetraffiore,inthecondtofexperiments,wefixedtheSDN-FEsfortheExodustopologytofourandfixedtheirlocationbadonsomeestimatedtrafficho20randomtrafficmatricestochecktheperformanceimprovementthatwegetwiththedifferenttraffieplotthenormalizedatthenormalizedthroughputofSDNroutingissignificantlybetterthanOSPFforalltheexperiments.0.9OSPF RoutingFlex Routing0.850.8Normalized

Throughput0.750.70.650.60.550.5111213Experiment manceforDifferentTraffi-2SimulationExperimentsHeretheobjectiveistomeasurelinkdelaysandkstateroutingprotocolwasmodifiedtoallowSDN-FEswhhofthethreenetworksv-eraltrafficpatternsweregeneratedwitheachsourcending2218

2013 Proceedings IEEE INFOCOM915nodeEXP115nodeEXP215nodeEXP3ExodusEXP1ExodusEXP2ExodusEXP3AbovenetEXP1AbovenetEXP2AbovenetEXP3MaxLossSDNRouting0.0000.0000.0000.0000.0002.0001.00052.0000.000MaxLossOSPF415.000391.000320.000140.000105.000106.000449.000555.000552.00015nodeEXP115nodeEXP215nodeEXP3ExodusEXP1ExodusEXP2ExodusEXP3AbovenetEXP1AbovenetEXP2AbovenetEXP3MeanDelaySDNRouting0.0030.0050.0020.0700.0600.0530.0100.0140.013MeanDelayOSPF0.0110.0100.0150.1060.1060.1000.0210.0290.025TABLEICOMPARISONOFMAXIMUMLOSSOVERALLLINKSTABLEIVCOMPARISONOFMEANDELAYOVERALLLINKS15nodeEXP115nodeEXP215nodeEXP3ExodusEXP1ExodusEXP2ExodusEXP3AbovenetEXP1AbovenetEXP2AbovenetEXP3MeanLossSDNRouting0.0000.0000.0000.0000.0000.0400.0120.6120.000MeanLossOSPF13.54311.89413.2284.9064.6934.1068.20011.5526.670TABLEIICOMPARISONOFMEANLOSSOVERALLLINKSdiffishownhowimprovednetworkperformancecaingevenafewstrategicaemedoesnotinvolvemakinganypro-tocolchangesattheremainingnodesinthenetworkwhichroutetraffilperformancemeasurementsaswellasns-2simulationsshowthatthemethodcansignificantlyimproveoverallnetworkthNCES[1],an,,,n,r,”Ethane:TakingControloftheEnterpri”,ACMSIGCOMMCCR,37(4):1-12,2007[2],ll,er,d,,Merwe,”DesignandImplementationofaRoutingControlPlatform”,NetworkedSystemsDesignandImplementation,May2005.[3],n,,,,n,”Nox:TowardsaNetworkOperatingSystem”,ACMSIGCOMMCCR,July,2008[4]er,ishnan,d,,Merwe,”TheCaforSeparatingRoutingfromRouters”,FDNA2004.[5],”OpeningAddress:2012OpenNetworkSummit”,April2012.[6]”NetworkDevelopmentandDeploymentInitiative(NDDI)”,/network/o/.[7]stas,”FasterApproximationSchemesforFractionalMulti-commodityFlowProblems”,PrentedatACM-SIAMSODA2002.[8].,”Onix:ADistributedControlPlatformforLargeScaleProductionNetworks”,OSDI2010,October,2010.[9]ento,berg,or,,na,aes”VirtualRoutersasaService:theRouteFlowapproachleveragingSoftware-DefinedNetworks”,CFI2011.[10].,”Network-WideDecisionMaking:TowardaWafer-ThinControlPlane”,HotNets-III,November2004.[11]nberg,,”RevisitingRoutingControlPlatformswiththeEyesandMusclesofSoftware-DefinedNetworking”,ACM-SIGCOMMHotSDNWorkshop,2012[12]an,opal,,i,,”TheSoftRouterArchitecture”,ProceedingofHotnets2004,November2004.[13]n,on,ishnan,ar,on,d,r,,”OpenFlow:EnablingInnovationinCampusNetworks”,ACMSIGCOMMCCR,April,2008.[14]TheOpenflowSwitch,openfl[15]ran,,,”AchievingNear-OptimalTrafficEngineeringSolutionsforCurrentOSPF/IS-ISNetworks”,IEEE/ACMTransactionsonNetworking,V.13,No.2,April2005.[16],,”OptimizingOSPF/IS-ISWeightsinaChangingWorld”,IEEEJournalonSelectedAreasinCommunications,traffientedswereudtogeneraterandomtraffi15noIandIIshowthemaximumnumberofpacketslostoverallthelinksandthemeannumberatSDNRoutingdoessignifieistrueforthemaximumdelay(TableIII)andmeandelay(TableIV).ThedifferenceinpSIONIncrementalSDNdeploymentwhereSDNsco-existwithtraditionaofparticularimportanceforlargenetworkswherecompletegreenfielddeploymentisMaxDelaySDNRouting0.1190.1530.0600.9130.7570.7170.2280.2880.297MaxDelayOSPF0.2830.2850.2771.6451.7321.7120.2420.3960.49615nodeEXP115nodeEXP215nodeEXP3ExodusEXP1ExodusEXP2ExodusEXP3AbovenetEXP1AbovenetEXP2AbovenetEXP3TABLEIIICOMPARISONOFMAXIMUMDELAYOVERALLLINKS2219