Specification and management of QoS in real-time databases s(3)

4APPROACH

Next,wedescribeourapproachformanagingtheperfor-manceofanRTDBintermsofQoTandQoD.First,westartbydefiningperformancemetricsinSection4.1.TheQoSspecificationmodelsaredescribedinSection4.2.AnoverviewofthefeedbackcontrolschedulingarchitectureisgiveninSection4.3,followedbythedescriptionofQoScontrollersinSection4.4.InSection4.5,wepresentthealgorithmsPC-MPU(precisioncontrolmisspercentageutilization),PC-MP(precisioncontrolmisspercentage),PC-ATEHEFandPC-ATEHEDFThesealgorithmsdeterminehowQoDisadjusted,i.e.,towhatextenttheprecisionofthedataobjectsismodifiedbasedonthecurrentsystemperformance.InSection4.6,wepresenttwomodels,describingthedynamicsofRTDBs,whichareusedtotunetheQoScontrollers.

4.1PerformanceMetrics

Inourapproach,thedatabaseoperator1canexplicitlyspecifytherequireddatabaseQoS,definingthedesiredbehaviorofthedatabase.Long-termperformancemetricssuchasaveragedeadlinemissratioarenotsufficienttospecifythedesiredperformanceofreal-timesystemsthatrequirestringentQoSenforcement[11].Therefore,inthiswork,weadaptbothlong-termperformancemetrics,referredtoassteady-stateperformancemetrics,andtransient-stateperformancemetrics.Weadaptthefollowingnotationofdescribingdiscretevariablesinthetime-domain:aðkÞreferstothevalueofthevariableaatthetimekT,whereTisthesamplingperiodandkisthesamplinginstant.2

QoTMetrics.WeconsiderthefollowingmetricsformeasuringQoTofadmittedtransactions:.

DeadlinemisspercentageofmandatoryusersubtransactionsjdeadlmissMðkÞjisgivenbymMðkÞ¼100Â%ÞwherejdeadlmissM

ððkÞjdenotesthenumberofmandatorysubtransactionsthathavemissedtheirdeadlineandjtermMðkÞjisthenumberofterminatedmandatorysubtransactions.

1.Byadatabaseoperator,wemeananagent,humanorcomputer,thatoperatesthedatabase,includingsettingtheQoS.

2.Fortherestofthispaper,wesometimesdropk,wherethenotionoftimeisnotofprimaryinterest.

.

Deadlinemisspercentageofoptionalusersubtran-sactionsisgivenbymOðkÞ¼100Âjdeadlmiss

O

ðkÞj

wherejdeadlmissO

ðkÞjdenotesthenumberð%Þ,ofop-tionalsubtransactionsthathavemissedtheirdead-lineandjtermOðkÞjisthenumberofterminatedoptionalsubtransactions.Note,jdeadlmissOðkÞjandjtermOðkÞjincludetheoptionalsubtransactionsthatarenotcompleted.

.

Averagetransactionerrorisdefinedas

P

ateðkÞ¼100Â

i2termðkÞtei

jtermðkÞj

ð%Þ;wheretermðkÞdenotesthesetofterminatedtransactions.

QoDMetric.ThemaximumdataerrormdeðkÞgivesthemaximumdataerrortoleratedforthedataobjects(asdescribedinSection2).

QoSFairnessMetric.Forsomeapplications,itisdesirabletomeasureQoSfairnessamongtransactionsand,therefore,weintroducethestandarddeviationoftransactionerror,

s P2

sdteðkÞ¼

i2termðkÞð100ÂteiÀateðkÞÞ

jtermðkÞjÀ1whichisameasureofhowmuchthetransactionerrorofterminatedtransactionsdeviatesfromtheaveragetransac-tionerror.

SystemUtilization.WemeasuresystemutilizationuðkÞingtheutilizationofthesystem,wecanshowwhetherouralgorithmsprovidehighthroughput.

Steady-StateandTransient-StatePerformanceMetrics.Thedesiredperformanceofthesystemisgivenbyasetofreferencesspecifyingthedesiredlevelofthecontrolledvariables,whichrepresenttheactualsystemperformance.Weconsiderthefollowingtransient-stateperformancemetrics(seeFig.1a).OvershootMpistheworst-casesystemperformanceinthetransientsystemstateanditisgiveninthepercentagebywhichacontrolledvariableovershootsitsreference.SettlingtimeTsisthetimeforthetransientovershoottodecayandsettlearoundthesteadystateperformanceanditisameasureofsystemadaptability,i.e.,howfastthesystemconvergestowardthedesiredperfor-mance.Hence,theperformanceofthecontrollersisdistinguishedbyhowwelltheyforceacontrolledvariableyðkÞtofollowortrackadesiredlevelgivenbyareferenceyrðkÞ,despitethepresenceofdisturbancesinthecontrolledsystem.ItisthereforeinterestingtomeasurethedifferencebetweenyrðkÞandyðkÞoveraperiodfunctionsJa¼PNoftime,whichisobtainedusingthe1k¼is1jyrðkÞÀyðkJs¼1

PN

Þjandk¼1ðyrðkÞÀyðkÞÞ2,whereNthenumberofsamplestaken.ThelowerJsandJaare,thebetteracontrollerisabletokeepynearyrand,also,thefasteryconvergestowardyr.

4.2QoSSpecificationModels

Themaximumdataerrorprovidesadirectmeasureoftheprecisionofthedataobjectsand,hence,weexpressQoDintermsofmde.AnincreaseinQoDreferstoadecreasein

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

308IEEETRANSACTIONSONCOMPUTERS,VOL.55,NO.3,MARCH

2006

Fig.1.Performancespecificationandsystemarchitecture.(a)Settlingtimeandovershoot.(b)Feedbackcontrolschedulingarchitecture.

mde,i.e.,anincreaseindataprecision.Incontrast,adecreaseinQoDreferstoanincreaseinmde.WeconsidertwoalternativewaysofdefiningQoS,belowreferredtoasQoSspecificationtypeAandtypeB,wheretheydifferinthewayQoTisexpressed.

QoSSpecificationTypeA.Inthecasewhenitisnotpossibletomodeltransactionerrorusingerrorfunctions,wehavetoexpressQoTbyothermeans.Weknowthatthemoreoptionalsubtransactionswecompletebeforethedeadline,thesmallerthetransactionerrorwillbe.Therefore,thedeadlinemissratioofoptionalsubtransac-tionsqualifiesasanapproximatemeasureofthetruetransactionerrorand,hence,wedefineQoTintermsofmO.QoTdecreasesasmOincreases(similarly,QoTincreasesasmOdecreases).Thedatabaseoperatorcanspecifysteady-stateandtransient-statebehaviorformM,mO,u,andmde.Thespecificationforuisgive …… 此处隐藏：5547字，全部文档内容请下载后查看。喜欢就下载吧 ……