Storage device performance prediction with CART models(10)

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

Tracenamecello99acello99bcello99c

Length4weeks4weeks4weeks

Tracedescription

AverageSize

7.8Million

7.1KB118.0KB8.5KB

1.1Million

singledisk

35.4%

115.71ms113.61ms5.04ms

99.9%

7.40ms59.28ms

Table2:Tracesummary.WemodelanAtlas10K9GBandaRAID5diskarrayconsistingof8Atlas10Kdisks.TheresponsetimeiscollectedbyreplayingthetracesonDiskSim3.0[5].

Traces.Weusethreesetsofreal-worldtracesinthisstudy.Table2liststhesummarystatisticsoftheeditedtraces.The rsttwo,cello92andcello99capturetypicalcomputersystemresearchI/Oworkloads,collectedatHPLabsin1992and1999respectively[27,14].Wepreprocesscello92toconcatenatetheLBNsofthethreemostactivedevicesfromthetraceto llthemodeleddevice.Forcello99,wepickthethreemostactivedevices,amongthe23devices,andlabelthemcello99a,cello99b,andcello99c.Thecello99traces tina9GBdiskperfectly,sonotraceeditingisnecessary.Asthesetracesarelong(twomonthsforcello92andoneyearforcello99),wereportdataforafour-weeksnapshot(5/1/92to5/28/92and2/1/99to2/28/99).

TheSAPtracewascollectedfromanOracledatabaseserverrunningSAPISUCCS2.5Binapowerutilitycompany.Theserverhasmorethan3,000usersanddiskaccessesre ecttheretrievalofcustomerinvoicesforupdatingandreviewing.SequentialreadsdominatetheSAPtrace.

Evaluationmethodology.Theevaluationusesthedevicemodelstopredicttheaverageand90thper-centileresponsetimeforone-minuteworkloadfragments.Wereportthepredictionerrorsusingtwometrics:

Y,andrelativeerrorabsoluteerrorde nedasthedifferencebetweenthepredictedandtheactualvalue,Y

de nedasYY