Storage device performance prediction with CART models(15)

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

200

# of intervals

1501005000%

25%50%75%100%% of sequential requests

(a)Histogramsonsequentiality(b)Averageabsoluteerror(c)Absoluteerrordistribution

Figure8:Effectsofdifferenttrainingworkloads.

6Conclusions

Storagedeviceperformancemodelingisanimportantelementinself-managedstoragesystemsandotherapplicationplanningtasks.Ourtargetmodeltakesaworkloadasinputandpredictsitsaggregateperfor-manceonthemodeleddeviceef cientlyandaccurately.Thispaperpresentsourinitialresultsinexploringmachinelearningtoolstobuilddevicemodels.Ablackboxpredictivetool,CART,makesdevicemodelsindependentofthestoragedevicesbeingmodeled,andthus,generalenoughtohandleanytypeofdevices.Themodelconstruction,alsoknownastraining,consistsoftwophases:replayingtracesonthedevicesandbuildingaCARTmodelbasedontheobservedresponsetimes.Modelinganewdeviceinvolvesonlytrainingonthetargetdevice.

CART-basedmodelstakeinputintheformofvectors,soworkloadsmustbetransformedintovectorsinordertouseCARTasthebasisfordevicemodels.Thispaperpresentstwowaystoaccomplishsuchatransformation,yieldingtwotypesofdevicemodels.Therequest-leveldevicemodelsrepresenteachrequestasavectorandpredictitsresponsetime.Asaresult,themodelsareabletopredicttheentireresponsetimedistribution.Theexperimentsshowthatthepredictedresponsetimehasademerit gureof33%foramodernUNIX leservertrace,leadingtoamedianrelativeerroraslowas16%foraggregateperformancepredictions.Theworkload-leveldevicemodels,ontheotherhand,transformaworkloadfragmentintoavectorandpredictitsaggregateperformancedirectly.Thevectortakesadvantageoftheef ciententropyplotmetrictocapturethetemporalandspatialburstinessaswellasthecorrelationswithinI/Oworkloads.Themedianrelativeerrorcanbeaslowas29%fortheworkload-leveldevicemodels.

Theerroranalysissuggeststhatthequalityofthetrainingworkloadsplaysacriticalroleinthemodelaccuracy.Themodelsareunabletopredictworkloadsthataredifferentfromthetrainingworkloads.Toaccuratelypredictarbitraryworkloads,itisimportantforthetrainingworkloadstobeasdiverseaspossibletocoverawiderangeofworkloads.Ourfutureworkwillexploretheeffectivenessofexistingsyntheticworkloadgeneratorsinproducinghigh-qualitytrainingworkloads.

Continuingresearchcanimprovethemodelpredictionaccuracy.First,ourexperimentsshowtherele-vanceoftrainingtraces.Generatingrulestoassistintrainingsuchmodelsbroadlyenoughwillbeimportant.Second,theworkloadcharacterizationproblempersists,affectingtheworkload-levelmodels.Webelieve,however,thatthecontextofferedbythemodelscanhelpproduceinsightintothislong-standingproblem.Third,thetwotypesofdevicemodelsshowdesirablepropertiesintrainingandpredicting,respectively.Itshouldbevaluabletohaveamodelthatcombinesthebestofbothapproaches.