Storage device performance prediction with CART models(11)

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

Relative importanc

e

60%40%20%0%

TimeDiffTimeDiffTimeDiffTimeDiffTimeDiffTimeDiffTimeDiffTimeDiffTimeDiffTimeDiff1LBNLBNDiffLBNDiffLBNDiffLBNDiffLBNDiffSize(a)relativeimportancemeasuredonthe9GBAtlas10KBdiskusingcello99aand

cello99c

SeR(b)relativeimportancemeasuredontheRAID5diskarrayusingcello99aandcello99c

Figure4:Relativeimportanceoftherequestdescriptionfeatures.

5.1CalibratingRequest-LevelModels

Thissectiondescribeshowweselectparametervaluesforkandlfortherequest-leveldevicemodels.

Figure4showstherelativeimportanceoftherequestdescriptionfeaturesindeterminingper-requestresponsetimebysettingkto10andlto5.Thefeature’srelativeimportanceismeasuredbyitscontributioninerrorreduction.Thegraphsshowtheimportanceofrequestdescriptionfeaturesmeasuredonbothdevices,trainedontwotraces(cello99aandcello99c).Weuseonlythe rstdayofthetracesandreducethedatasetsizeby90%withuniformsampling.

First,weobservethattherelativeimportanceisworkloaddependent.Asweexpected,forbusytraf- csuchasthatwhichoccurredinthecello99atrace,thequeuingtimedominatestheresponsetime,andthereby,theTimeDifffeaturesaremoreimportant.Ontheotherhand,cello99chassmallresponsetimes,andfeaturesthatcharacterizethedatatransfertime,suchasSizeandRW,havegoodpredictivepowerinmodelingthesingledisk.

Second,weobservethatthemostimporantfeatureshiftsfromTimeDiff8toTimeDiff7wherecom-paringthesingledisktothediskarrayforcello99abecausethequeuingtimebecomeslesssigni cantforthediskarray.Thedistinctionbetweenthetwotraces,however,persists.

Wesetkto10forTimeDiffandlto3forLBNDiffinthesubsequentexperimentssothatwecanmodeldevicebehaviorunderbothtypesofworkloads.

Weshowthemodelaccuracyinpredictingper-requestresponsetimesinFigure5.ThemodelisbuiltfortheAtlas10Kdisk.Thetrainingtraceisthe rstdayofcello99a,andthetestingtraceistheseconddayofthesametrace.Figure5(a)isascatterplot,showingthepredictedresponsetimesagainsttheactualonesforthe rst5,000requests.Mostofthepointsstayclosetothediagonalline,suggestingaccuratepredictionoftherequest-leveldevicemodel.Figure5(b)furthercomparestheresponsetimedistributions.Thelongtailofthedistributioniswellcapturedbytherequest-levelmodel,indicatingthattherequestdescriptionis