Storage device performance prediction with CART models(8)

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

Figure3:TwotypesofCART-baseddevicemodels.

2.Ricanbecalculatedfromrj,ji.Thisconstraintsimpli estherequestdescription.Inmostcases,theresponsetimeofacurrentrequestdependsonlyonpreviousrequestsandtherequestitself.

OurrequestdescriptionRiforrequestricontainsthefollowingvariables:

Ri

TimeDiffi1

TimeDiffik

LBNi

LBNDiffi1

LBNDiffil

Sizei

RWi

Seqi

whereTimeDiffikArrivalTimeiArrivalTimei2k1andLBNDiffilLBNiLBNil.The rstthreegroupsoffeaturescapturethreecomponentsoftheresponsetime,andSeqiindicateswhethertherequestisasequentialaccess.The rstk1featuresmeasurethetemporalburstinessoftheworkloadwhenriarrives,andsupportpredictionofthequeuingtime.WeallowtheTimeDifffeaturestoexponentiallygrowthedistancefromthecurrentrequesttohistoryrequesttoaccommodatelargebursts.Thenextl1featuresmeasurethespatiallocality,supportingpredictionoftheseektimeoftherequest.SizeiandRWisupportpredictionofthedatatransfertime.

Thetwoparameters,kandl,determinehowfarwelookbackforrequestburstsandlocality.Smallvaluesdonotadequatelycapturethesecharacteristics,rgevalues,ontheotherhand,leadstoahigherdimensionality,meaningtheneedforalargertrainingsetandalongertrainingtime.Theoptimalvaluesfortheseparametersarehighlydevicespeci c,andSection5.1showshowweselecttheparametervaluesinourexperiments.

4.3Workload-LevelDeviceModels

Theworkload-levelmodelrepresentstheentireworkloadasasingleworkloaddescriptionandpredictsaggregatedeviceperformancedirectly.TheworkloaddescriptionWcontainsthefollowingfeatures.

W

Averagearrivalrate

Readratio

Averagerequestsize

Percentageofsequentialrequests

Temporalandspatialburstiness

Correlationsbetweenpairsofattributes