Storage device performance prediction with CART models(13)

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

averageresponsetime

(b)Predictionerrorfor90thpercentileresponsetime

Figure6:Comparisonofpredictorsforasingle9GBAtlas10Kdisk.

Forexample,constantlyover-predictsforcello99cbecausethemodelwasnevertrainedwththesmallsequentialaccessesthatareparticulartocello99c.Section5.4givesaninformalerroranalysisandidenti esinadequatetrainingbeingthemostsigni canterrorsource.

Fourth,highquantileresponsetimesaremoredif culttopredict.Weobservelargerpredictionerrorsfromallthepredictorsfor90thpercentileresponsetimepredictionsthanforaverageresponsetimepredic-tions.TheaccuracyadvantageofthetwoCART-basedmodelsishigherfor90thpercentilepredictions.

Insummary,thetwoCART-basedmodelsgiveaccuratepredictionswhenthetrainingandtestingwork-loadssharethesamecharacteristicsandinterpolatewellotherwise.Thegoodaccuracysuggeststheeffec-tivenessoftherequestandworkloaddescriptionsincapturingimportantworkloadcharacteristics.

5.3ModelingADiskArray

Figure7comparestheaccuracyofthefourpredictorsinmodelingthediskarray.ThepredictorisnotpresentedbecausetheSAPtracedoesnotprovideenoughinformationonarrivaltimeforustoknowtheoffsetwithinaweek.Theoverallresultsaresimilartothoseforthesingledisk.ThetwoCART-basedmodelsarethemostaccuratepredictors.Theabsoluteerrorsbecomesmallerduetothedecreasedresponsetimefromthesingledisktothediskarray.Therelativeaccuracyamongthepredictors,however,staysthesame.Overall,theCART-baseddevicemodelingapproachworkswellforthediskarray.

5.4ErrorAnalysis

Thissectionpresentsaninformalerroranalysistoidentifythemostsigni canterrorsourcefortheCART-baseddevicemodels.