Storage device performance prediction with CART models(7)

Storage device performance prediction is a key element of self-managed storage systems and application planning tasks, such as data assignment. This work explores the application of a machine learning tool, CART models, to storage device modeling. Our appr

Figure2:Modelconstructionthroughtraining.RTiistheresponsetimeofrequestri.

4PredictingPerformancewithCART

ThissectionpresentstwowaysofconstructingdevicemodelsbasedonCARTmodels.

4.1Overview

OurgoalistobuildamodelforagivenstoragedevicewhichpredictsdeviceperformanceasafunctionofI/Oworkload.Thedevicemodelreceivesaworkloadasinputandpredictsitsaggregateperformance.Wede neaworkloadasasequenceofdiskrequests,witheachrequest,ri,uniquelydescribedbyfourattributes:arrivaltime(ArrivalTimei),logicalblocknumber(LBNi),requestsizeinnumberofdiskblocks(Sizei),andread/writetype(RWi).Thestoragedevicecouldbeasingledisk,adiskarray,orsomeotherlike-interfacedcomponent.Theaggregateperformancecanbeeithertheaverageorthe90-thpercentileresponsetime.

OurapproachusesCARTtoapproximatethefunction.Weassumethatthemodelconstructionalgorithmcanfeedanyworkloadintothedevicetoobserveitsbehaviorforacertainperiodoftime,alsoknownas“training.”Thealgorithmthenbuildsthedevicemodelbasedontheobservedresponsetimes,asillustratedinFigure2.Modelconstructiondoesnotrequireanyinformationabouttheinternalsofthemodeleddevice.Therefore,itisgeneralenoughtomodelanydevice.

Regressiontoolsareanaturalchoicetomodeldevicebehavior.Suchtoolsaredesignedtomodelfunc-tionsonmulti-dimensionalspacegivenasetofsampleswithknownoutput.Thedif cultyistotransformworkloadsintodatapointsinamulti-dimensionalfeaturespace.Weexploretwowaystoachievethetrans-formationasillustratedinFigure3.Arequest-levelmodelrepresentsarequestriasavectorRi,alsoknownasthe“requestdescription,”andusesCARTmodelstopredictper-requestresponsetimes.Theaggregateperformanceisthencalculatedbyaggregatingtheresponsetimes.Aworkload-levelmodel,ontheotherhand,representstheentireworkloadasasinglevectorW,orthe“workloaddescription,”andpredictstheaggregateperformancedirectlyfromW.Inbothapproaches,thequalityoftheinputvectorsiscriticaltothemodelaccuracy.Thenexttwosectionspresenttherequestandworkloaddescriptionsindetail.

4.2Request-LevelDeviceModels

ThissectiondescribestheCART-basedrequest-leveldevicemodel.ThismodelusesaCARTmodeltopredicttheresponsetimesofindividualrequestsbasedonrequestdescriptions.Themodel,therefore,isabletogeneratetheentireresponsetimedistributionandoutputanyaggregateperformancemeasures.

Weadoptthefollowingtwoconstraintsindesigningtherequestdescription.1.Ridoesnotincludeanyactualresponsetimes.Onecouldrelaxthisconstraintbyallowingthein-clusionoftheresponsetimeinformationforalltherequeststhathavealreadybeenservedwhenthecurrentrequestarrives.Thisrelaxation,however,isfeasibleonlyforonlineresponsetimepredictions;itwouldnotbeappropriateforapplicationplanningtasksbecausetheplannerdoesnotrunworkloadsondevices.