A Web Service for evaluation of load balancing str

AWebServiceforevaluationofloadbalancingstrategiesfordistributedwebserversystemsDigvijaySinghLambaSeniorUndergraduate,Dept.ofComputerSc.andEngg.,IITKanpurDr.PankajJaloteDept.ofComputerSc.andEngg.,IITKanpurDr.DheerajSanghiDept.ofComputerSc.andEngg.,IITKanpurAbstractManylargewebsitesgetmillionsofhitseveryday.Theyneedascalablewebserversystemthatcanprovidebetterperformancetoalltheclientsthatmaybeindifferentgeographicalregions.Commonapproachtoimproveperformanceistohavefullyreplicatedwebserverclustersindifferentgeographicalregionswithreplicatedservers.Insuchanenvironment,oneofthemostimportantissuesisselectingaserverforservicingarequest.Clientrequestsshouldbedirectedtoaserversuchthatthetimetakenforservicingtherequestcanbeminimised.Differentpoliciesarepossibleforserverselectionanditisdifficulttodeterminetheimpactofdifferentpolicies.Inthispaper,wedescribeawebservicethatcanbeusedtoevaluatetheperformanceofdifferentschemes.Weprovideawebbasedtestbedtoperformthistask.ThisserviceutilizesadedicatedtestbedthatemulatestheWorldWideWebandonwhichtheperformanceofvariousstrategiescanbestudied.Ausergivesthevaluesofdifferentnetworkandloadparametersandthespecificsofthepolicy.Thetestbedisconfiguredtoemulateawebserversystemusingthespecifiedpolicyandperformancetestingisdone.Datafromthetestsiscollected,analyzedandgraphicallydisplayedtotheuser.Theservicealsoallowsnewapproachestobetested.IntroductionNumberofusersaccessingtheInternetisincreasingquiterapidlyanditiscommontohavemorethan100millionhitsadayforpopularwebsites.Forexample,netscape.comwebsitereceivesmorethan120millionhitsaday.Thenumberofusersisexpectedtocontinueincreasingatafastrateandhenceanywebsitethatispopular,facesthechallengeofservingverylargenumberofclientswithgoodperformance.Fullmirroringofwebserversorreplicationofwebsitesisonewaytodealwithincreasingnumberofrequests.Manytechniquesexistfortheselectionofthenearestwebserverfromtheclient’spointofview.Ideally,selectionofbestservershouldbedonetransparentlywithouttheinterventionoftheuser.Manyoftheexistingschemesdoonlyload-balancing.Theseschemesassumethatthereplicatedsitehasallthewebserversinonecluster.Thisisalrightformediumsizedsites,butbeyondacertainamountoftraffic,theconnectivitytothisoneclusterbecomesabottleneck.Solargewebsiteshavemultipleclusters,anditisbesttohavetheseclustersgeographicallydistributed.Thischangestheproblemtofirstselectthenearestclusterandthendoloadbalancingwithintheserversofthatcluster.Ofcourse,ifallserversinaclusterareheavilyloadedthenanotherclustershouldhavebeenchosen.Sotheproblemismorecomplexinsuchanenvironment.Designingsuchasysteminvolvesmakingdecisionsabouthowthebestservercanbeselectedforarequestsuchthattheusergetsaresponseinminimumtimeandhowthisrequestisdirectedtothatserver.Insomestrategies,aserverisselectedwithouttakingintoaccountanysystemstateinformation,e.g.random,roundrobinetc[6].Somepoliciesuseweightedcapacityalgorithmstodirectmorepercentageofrequeststomorecapableservers[7].Somestrategiesselectaserverbasedontheserverstate[7]andsomeotherstakeclientstateinformationintoaccount[7].Thereisalwaysatradeoffbetweentheoverheadduetocollectionofsystemstateinformationandperformancegainbyuseofavailablestateinformation.Iftoomuchstateinformation(ofserverorclients)iscollected,itmayresultinhighoverheadsforcollectionofinformationandperformancegainmaynotbecomparabletooverheads.Theperformanceofanyloadbalancingapproachdependsonahostoffeatureslikenetworkdelays,packetlosses,transmissionerrors,rateofrequests,serverloadetc.Itisusuallyveryhardtoanalyticallydeterminetheperformanceofapolicygivensomeconditions.Simulationsalsohavelimitationsinthattheycanonlytakelimitedvariablesintoaccountandwiththecomplexitiesinvolvedinthiscaseeffectofallthevariablescannotbeanalyticallydeterminedandsimulated.Hencetestingforperformancebysettingupatestbedisareasonablewaytoevaluatedifferentstrategies.Thistestingwillingeneralrequireadedicatedtestbedonwhichperformancestudiescanbedone.Suchatestbedshouldbeconfigurablefordifferentstrategiesandnetworkcharacteristics.Assuchatestbedislikelytobeadedicatedsetupwithafocussedpurpose,itwillbeusefulifawebservicecanbecreatedforevaluatingapolicy.Inthis,allparameterswillbesetinthewebservice,whichwillthendrivethetestbedtoobtaintheresults.Suchaservicewillmakeadedicatedtestbedaccessibleacrosstheworld.Suchaserviceshouldfulfilanumberofrequirements.Itshouldbepossibletostudytheimpactofdifferentparametersontheperformanceofvariousstrategiesandtocomparethem.Itshouldbepossibletosubmitnewstrategiesandcomparetheirperformancewithexistingonesortoidentifyconditionswherethepolicyperformsbest.Inthispaper,wedescribeawebserviceforevaluatingloadbalancingstrategiesfordistributedwebserversystems.Thewebserviceinteractswithatestbedtoautomatetheprocessofsubmissionofparameters,testingandresultgenerationtoallowtheusertotestandcompareloadbalancingpoliciesthroughinahighlyconfigurablemanner.Wehavepre-definedsomepopularstrategieswhichcanbeevaluatedfordifferentparametersettings.Theservicealsoallowsuserstotestnewpoliciesforloadbalancingandcomparetheirp