外文翻译基于哈希模式的负载均衡性能研究.docx

外文翻译基于哈希模式的负载均衡性能研究.docx

《外文翻译基于哈希模式的负载均衡性能研究.docx》由会员分享，可在线阅读，更多相关《外文翻译基于哈希模式的负载均衡性能研究.docx（21页珍藏版）》请在冰豆网上搜索。

外文翻译基于哈希模式的负载均衡性能研究

中文5790字

毕业设计外文翻译

专业网络工程

班级

学生姓名xx

学号xx

指导教师

PerformanceofHashing-BasedSchemesforInternetLoadBalancing

ZhiruoCao,ZhengWang,EllenZegura

CollegeofComputing

GeorgiaInstituteofTechnology

Atlanta,GA30332-0280

BellLabs

LucentTechnologiesHolmdel,NJ07733

Abstract—LoadbalancingisakeytechniqueforimprovingInternetperformance.Effectiveuseofloadbalancingrequiresgoodtrafficdistributionschemes.Westudytheperformanceofseveralhashingschemesfordistributingtrafficovermultiplelinkswhilepreservingtheorderofpacketswithinaflow.Althoughhashing-basedloadbalancingschemeshavebeenproposedinthepast,thisisthefirstcomprehensivestudyoftheirperformanceusingrealtraffictraces.

Weevaluatefivedirecthashingmethodsandonetable-basedhashingmethod.Wefindthathashingusinga16-bitCRCovertheFivetuplegivesexcellentloadbalancingperformance.Further,load-adaptivetable-basedhashingusingtheexclusiveORofthesourceanddestinationIPaddressesachievescomparableperformancetothe16-bitCRC.Table-basedhashingcanalsodistributetrafficloadaccordingtounequalweights.Wealsoreportonfourotherschemeswithpoortomoderateperformance.

Keywords—Loadsharing,hashing.

I.INTRODUCTION

Loadbalancing（alsoknownasloadsharing）isakeytechniqueforimprovingtheperformanceandscalabilityoftheInternet.Forexample,manylargeenterprisenetworksareconnectedtomultipleInternetServiceProviders（ISPs）toachieveredundantconnectivityandtodistributetrafficloading.InsidetheInternet,thebackbonesareoftenengineeredtohavemultipleparalleltrunksbetweenmajorPointsofPresencetoensurehighavailability.Typically,theseparalleltrunksareconfiguredasequal-costpathsandallowloadbalancingoverthem.

TheparalleltrunksmaybecomeevenmoreubiquitouswhenthepromisingDenseWavelengthDivisionMultiplexing（DWDM）technologyisdeployedinthefutureInternetback-bone.DWDMexpandsthecapacityofcommunicationtrunksbyallowingagreaternumberofchannelstobecarriedonasingleopticalfiber.WithpotentiallytensorevenhundredsofDWDMchannelsbetweenmajorpoints,loadbalancingisessentialinbestutilizingthemultipleparallelchannels.

ParallelarchitectureshavebeenusedforpacketprocessingforcopingwithexponentialgrowthinInternettraffic,Insteadofoneprocessingengine,packetsaredispatchedtomultipleparallelenginesinsidearoutertoincreasetheoverallprocessingthroughput.Thesametechniqueisalsousedinscalingwebservers.PopularwebserversoftenoperateafarmofmachinesandtheroutersconnectedtothemsplittheHTTPrequeststodifferentmachines.

Foralloftheseexamples,effectiveuseofloadbalancingrequiresgoodschemesforsplittingtrafficovermultiplelinks.Inaddition,sincethemajorityofthetrafficontheInternetisTCP-based[1],trafficsplittingschemesneedtoavoidpacketmisorderingwithinaTCPflow,whichcanfalselytriggercongestioncontrolmechanismsandcauseunnecessarythroughputdegradation[2],[3].

Inthispaper,weproposeandevaluateaclassofhashingbasedtrafficsplittingalgorithmswhichpreserveper-flowpacketordering.Weconsiderfivehashfunctionsthatare“direct,”meaningthatthehashfunctionproducesavalueintherangeof0...N-1,whereNisthenumberofoutgoinglinks.Wealsoconsideratable-basedgeneralizationthatinvolveshashingtoMbins,thenassigningtheMbinstotheNoutgoinglinks.Tablebasedhashingrequiresmorestatethandirecthashing,buthastheflexibilitytosupportunequalloaddistributionanddynamicadaptation.

Ourresultsareobtainedbysimulatingtheperformanceofatrafficsplitter,usingpackettracestakenfromtwotrunksofamajorInternetbackboneprovider.WefindthatdirecthashingwiththedestinationIPaddresscausessignificantimbalanceacrosstwolinks.UsingtheInternetchecksumortheexclusiveORofboththesourceIPaddressanddestinationIPaddressimprovestheperformanceconsiderably,thoughmoderateimbalancepersists.Themorecomputationallycomplex16-bitCRCofthefive-tuple（sourceaddress,destinationaddress,sourceport,destinationportandprotocolid）givesexcellentloadbalancingperformance,keepingtheloadandqueuelengthsverysimilarontwolinks.Equallygoodloadbalancingcanbeachievedusingtable-basedhashingwithadaptation,whichrequireslesscomputationthantheCRCbutnecessitatesmonitoringthelinkloadsandstoring（andadjusting）themappingfromtablebinstolinks.

Table-basedhashinghastheadditionaladvantagethatitcandistributetheloadaccordingtounequalweights.Further,anindex-basedversionofthisschemecanaltertheweightdistributionwithminimaldisruptiontoexistingflows.Ourresultsconfirmthattheindex-basedhashingcanaccuratelyachieveaweighteddistributionwhenadaptationisalsoused.

Therestofthispaperisorganizedasfollows.InSectionIIwediscussrelatedworkintrafficsplittingandloadbalancing.SectionIIIdescribesthebehaviorofanidealtrafficsplitter,explainstherequirementsforapracticalsystem,anddefinestheperformancemetricsthatwillbeusedtoassessvarioushashing-basedschemes.ThesetofschemesthatweconsideraredescribedinSectionIV.TheresultsofourstudyaredescribedinSectionV,andincludeanalysisoftherandomnessinherentinthetracedata（SectionV-A）.WeconcludeandmentionareasforfutureworkinSectionVI.

II.RELATEDWORK

Loadbalancinghasbeenusedintelecommunicationnetworksintheformofinversemultiplexing[4].Inversemultiplexingallowsserviceproviderstoofferwidebandchannelsbycombiningmultiplenarrowband56kbpsand64kbpstrunks[5].Theloadbalancingininversemultiplexingistypicallybasedonroundrobindistributionofpacketsorbytes[6],[7].

Ourworkdiffersfrominversemultiplexingintwoimportantdimensions.First,inversemultiplexingisdesignedforuseoverpoint-to-pointlinks;itstechniquesarenottypicallyapplicablefornetworklayerloadbalancing.Internetloadbalancing,however,makesuseofthenaturalredundancyinthenetworktopology.Thepathsforloadbalancing,forexample,equal-costmulti-paths,arediscovereddynamicallybyroutingprotocols,suchasOSPF[8],ratherthanthroughconfiguration.Second,inordertomaintainsynchronizationandper-flowFIFOpacketorderingininversemultiplexing,itisnecessarytoaddextrapacketheaderswithsequencenumbersortokeepstateatbothendsofthechannel.But,implementingtheseadditionalmechanismsfornetworkloadbalancingrequiresanewnetworkprotocol.Incomparison,thehashing-basedschemescanmaintainper-flowpacketorderingandcanbeimplementedwithoutrequiringanyadditionalprotocolsupport.

Hashinghasbeenwidelyusedinindexingandsearching[9].Inthenetworkingcontext,hashing-basedalgorithmsforaddresslookup[10],flowidentification[11]andpacketdemultiplexing[12]havebeenproposedinthepast.Theuseofhashingfornetworkloadbalancingisnotnew.SomecommercialrouterproductshaveimplementedsimplehashingovertheIPdestinationaddresstodistributetraffic[13].IntheOSPFOptimizedMultipathprotocol（OSPF-OMP）[14],anumberofpossibleapproachesforloadbalancingovermultiplepathsarementioned,includingper-packetroundrobin,dividingdestinationprefixesamongavailablenexthopsintheforwardingtable,anddividingtrafficaccordingtoahashfunctionappliedtothesourceanddestinationpair.However,theproposedschemesarenotevaluatedwithsimulationorrealnetworkmeasurement.InthestudyofloadbalancingwithOSPF-OMP,perfecthashingisassumed[15].

Atrafficsplittingschemeusingrandomnumbersisproposedin[16].Itappliesthename-basedmappingsapproachtoloadbalancing[17].Inthisscheme,eachnext-hopisassignedwithaweightbasedonasimplepseudo-randomnumberfunctionseededwiththeflowidentifierandthenext-hopidentifier.Whenapacketarrives,theweightsaregenerated,andthenext-hopreceivingthehighestweightisusedforforwarding.Theschemeisapproximatelytimesasexpensiveasahashing-basedscheme,whereisthenumberofoutgoinglinks.Again,noperformancestudyontheproposedschemeispresented.

Itisclearthatalthoughhashing-basedschemesfortrafficsplittinghavebeenproposedinthepast,andsomesimpleschemeshaveevenbeenimplementedincommercialproducts,theperformanceofsuchschemeshasnotbeenadequatelyevaluated.Thispaperpresentsthefirstcomprehensiveperformancestudyonawiderangeofhashing-basedschemes,usingrealpackettracesfrombackbonenetworks.

III.FRAMEWORK

Inthissection,wedescribethebehaviorofanidealtrafficsplitter,explaintherequirementsforapracticalsystem,anddefinetheperformancemetricsforassessingvariousschemes.

A.ReferenceModel

AloadbalancingsystemtypicallycomprisesatrafficsplitterandmultipleoutgoinglinksasshowninFigure1.Insuchasystem,thetrafficsplitterreceivesanincomingpacketfromahigher-speedlinkandforwardsittooneofthelower-speedoutgoinglinks.Agoodloadbalancingsystemshouldbeabletosplitthetraffictothemultipleoutgoinglinksevenlyorbysomepre-definedproportion.

In[7],ithasbeenobservedthatthereisacloserelationshipbetweenfairqueuingandloadbalancing.Wenowextendtheirobservationtoamathematicalmodeltoobtaintheconstraintsforidealtrafficsplitting.

Letusfirstlookatanidealfluidmodelwherethetrafficisinfinitelydivisible.Supposethatthereareoutgoinglinksintheloadbalancingsystem,andthecapacityoflinkIisui.LetSi（T,t）be