Storage device performance prediction with CART models(6)

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

FeatureCART

high(505%)

Neuralnetworks

fair(66%)

Poor

Fair

Poor

Poor

fast(seconds)

slow(hours)

fast

(milliseconds)

low(60B)

low(2MB)

Fair

k-nearestneighbors

InterpretabilityAbilitytohandleirrelevantinput

GoodGood

PoorPoor

PredictiontimeEaseofuse

fast

(milliseconds)

Good

slow(minutes)Fair

Table1:Comparisonofregressiontoolsinpredictingper-requestresponsetime.(ThesamedatasetisusedinFigure5.)Thecomparisononrow2,3,4andthelastoneistakenfrom[16].Werankthefeaturesintheorderoftheirimportance.Interpretabilityisthemodel’sabilitytoinfertheimportanceofinputvariables.Robustnessistheabilitytofunctionwellundernoisydataset.Irrelevantinputreferstofeaturesthathavelittlepredictivepowers.

buildtherequest-leveldevicemodelasdescribedinSection4.2.Themodelswereconstructedonthe rstdayofcello99aandtestsrunonthesecondofthesametrace.TheinformaiononthetracesweusedmaybefoundinSection5.

Themodel[29]usesalinearfunctionofXtoapproximatefX.Duetonon-linearstoragedevicebehavior,linearmodelshavepooraccuracy.

Themodel[26]consistsofasetofhighlyinterconnectedprocessingelementsworkinginunisontoapproximatethetargetfunction.Weuseasinglehiddenlayerof20nodes(bestamong20and40)andalearningrateof0.05.Halfofthetrainingsetisusedinbuildingthemodelandtheotherhalfforvalidation.Suchamodeltakesalongtimetoconverge.

The[6]mapstheinputdataintoahighdimensionalspaceandperformsalinearregressionthere.Ourmodelusestheradialbasisfunction

Kxix

expγx

xi

2

asthekernelfunction,andγissettobe2(bestamong1,3,4,6).Weuseanef cientimplementation,SVMlight[18],inourexperiment.Selectingtheparametervaluesrequiresexpertiseandmultipleroundsoftrials.

Themodel[9]ismemory-basedbecausethemodelremembersallthetrain-ingdatapointsandpredictionisdonethroughaveragingtheoutputoftheknearestneighborsofthedatapointbeingpredicted.WeusetheEuclideandistancefunctionandakvalueof5(bestamong5,10,15,and20).Themodelisaccurate,butisinef cientinstorageandcomputation.

Thelastthreetoolsrequirethatallthefeaturesandoutputbenormalizedtotheunitlength.Forfeaturesoflargevaluerange,wetakelogarithmsbeforenormalization.Overall,CARTisthebestatpredictingper-requestresponsetimes,withtheonlydownsidebeingslightlyloweraccuracycomparedtothemuchmorespace-andtime-consumingapproach.