Specification and management of QoS in real-time databases s(7)

MÞÞGOgLgAgOOÀ1

DYN;mðzÞ

¼

mTðTmÞ

ðzÀ1ÞðzÀ1þTðTm

OÞÞGgLgAgateTTÀ1DYN;ateðzÞ¼ate

ðzÀ1ÞðzÀ1þTTate

ÞrelatethecontrolinputÁLERðzÞtothecontrolledvariables

MMðzÞ,MOðzÞ,andATEðzÞ,respectively.

Aspresentedabove,wehaveextendedtheSTAmodelandthenewmodel,DYN,capturesadditionaldynamicsofthecontrolledsystem,bygivingmoreaccuratetime-domainrelationsMbetweenthecontrolinput lERandtheoutputsu,m,mO,andate.WehavealsodescribedmodelsparametersgL,gA,gMm,gOhowm,gate,Tu,TmMtocompute,TmO

the

,andTate.

5PERFORMANCEEVALUATION

Inthissection,adetaileddescriptionoftheperformedexperimentsisgiven.Thegoalandthebackgroundoftheexperimentsarediscussedand,finally,theresultsarepresented.

5.1PerformanceEvaluationGoals

Consideringthegoalofourwork,statedinSection2,thetwoobjectivesoftheperformanceevaluationare1)todetermineifthepresentedalgorithmscanprovideQoSguaranteesaccordingtoaQoSspecificationand2)todeterminethesuitabilityoftheproposedmodelDYNfordescribingtheperformanceofRTDBs.Consideringourfirstobjective,wehavestudiedandevaluatedthebehaviorofthealgorithmsundervariousconditions,whereasetofparametershavebeenvaried.Theyare:1)Load(load)ascomputationalsystemsmayshowdifferentbehaviorsfordifferentloads,especiallywhenthesystemisoverloaded.Forthisreason,wemeasuretheperformancewhen

applyingdifferentloadstothesystem.2)Executiontime

estimationerror(esterr)as,often,exactexecutiontimeestimatesoftransactionsarenotknown.Tostudyhowruntimeerroraffectsthealgorithms,wemeasuretheperformanceconsideringdifferentexecutiontimeestima-tionerrors.ThesecondobjectiveisinvestigatedbycomparingthecontrollertunedusingDYNwiththecontrollertunedusingSTAwithregardtoperformancereliabilityandperformanceadaptation.

5.2SimulationSetup

Thesimulatedworkloadconsistsofupdateandusertransactions,whichaccessdataandperformvirtualarithmetic/logicaloperationsonthedata.WehaveusedaworkloadbasedontheanalysisoftheNYSEstocktradesandcommercialdatabases[23].Updatetransactionsoccupyapproximately50percentoftheworkload.Inourexperi-ments,onesimulationrunlastsfor10minutesofsimulatedtime.Foralltheperformancedata,wepresenttheaverageof10simulationruns.Wehavederived95percentconfidenceintervalsbasedonthesamplesobtainedfromeachrunandusingthet-distribution[24].Theworkloadmodeloftheupdateandusertransactionsisdescribedasfollows:Weusethefollowingnotation,wheretheattributexireferstothetransactionTiandxi½ti isassociatedwiththesubtransactiontiofTi.

DataandUpdateTransactions.TheDBholds1,000tem-poraldataobjects(di),whereeachdataobjectisupdatedbyastream(Streami,1 i 1;000).4Theperiod(pi)isuniformlydistributedintherange(100ms,50s),i.e.,U:ð100ms;50sÞ,andestimatedexecutiontime(eeti)isgivenbyU:ð1ms;8msÞ.TheactualexecutiontimeofanupdateisgivenbythenormaldistributionN:ðeetpi;

eetiÞ.Theaverageupdatevalue(avi)ofeachStreamiisgivenbyU:ð0;100Þ.Theactualvalue(vi)ofanupdateissetaccordingtoN:ðavi;aviÂvarfactorÞ,wherevarfactorisuniformlydistributedin(0,1).Thedeadlineissettoarrivaltimeiþpi.Wedefinedataerrorastherelativedeviationbetweencviand

viasgivenbydei¼100ÂjcviÀvij

ið%Þ.

UserTransactions.EachSourceigeneratesatransactionTi,consistingofonemandatorysubtransaction,mi,andjOij(1 jOij 10)optionalsubtransaction(s),oi;j(1 j jOij).jOijisuniformlydistributedbetween1and10.Theestimated(average)executiontime(eeti½ti )ofthemandatoryandtheoptionalsubtransactionsisgivenbyU:ð5ms;15msÞ.Theexecutiontimeestimationerrorfactoresterrisusedtointroduceexecutiontimeestimationerrorintheaverageexecutiontimegivenbyaeti½ti ¼ðesterrþ1ÞÂeeti½ti .Further,upongenerationofatransaction,SourceiassociatesanactualexecutiontimeeachsubtransactiontigivenbyN:ðaeti½ti ;pto

,whichisaeti½ti Þ.ThedeadlineissettoarrivaltimeiþeetiÂslackfactor.Theslackfactorisuni-formlydistributedaccordingtoU:ð20;40Þ.Transactionsareevenlydistributedinfourclassesrepresentingerrorfunctionordersof0.5,1,2,and5(e.g.,25percentofthetransactionshaveanerrororderof1).WeusetheQoSspecifications

4.Incurrentautomotiveenginecontrolunits(ECUs),atypicalRTDBconsistsofapproximately500dataobjectswithtemporalconstraints.FutureECUswithmorefunctionalitywillincreasethedatavolumefurther.

Abstract—Real-time applications such as e-commerce, flight control, chemical and nuclear control, and telecommunication are becoming increasingly sophisticated in their data needs, resulting in greater demands for real-time data services that are provided

314IEEETRANSACTIONSONCOMPUTERS,VOL.55,NO.3,MARCH

2006

Fig.4.Experiment1:Averageperformancewhenvaryingload.(a)PC-MPUandPC-MP.(b)PC-ATE.

O

QoSSpecA¼fmMr¼1%;mr¼10%;mder¼2%;

Ts 60s;Mp 30%g

forPC-MPUandPC-MP,and

QoSSpecB¼fater¼20%;mder¼5%;Ts 60s;Mp 30%gforPC-ATEEDF,PC-ATEHEF,andPC-ATEHEDF.

5.3Baselines

Tothebestofourknowledge,therehasbeennoearlierworkontechniquesformanagingdataimprecisionandtransactionimprecisionaimedatsatisfyingspecificQoSorQoDrequirements.Forthisreason,wehavedevelopedtwobaselinealgorithms,Baseline-1andBaseline-2,tostudytheimpactoftheworkloadonthesystem.Inourperformanceevaluation,wealsoincludePC-ATEEDF,whichisworkingasareferencealgorithminadditiontothebaselines.WechooseEDFsinceitisoptimalinminimizingdeadlinemissesandhaswell-knownbehavior.ThealgorithmoutlineofBaseline-1andBaseline-2isgivenbelow.DependingonthegivenQoSspecificationtype,let beeithermOorate.Baseline-1.If (i.e.,mOorate)isgreaterthanitsreference,theutilizationisloweredbydiscardingmoreupdatetransactions,i.e.,increasingmde.Consequently,theprecise- …… 此处隐藏：5526字，全部文档内容请下载后查看。喜欢就下载吧 ……