计算机专业文献翻译性能测试方法.docx

计算机专业文献翻译性能测试方法.docx

《计算机专业文献翻译性能测试方法.docx》由会员分享，可在线阅读，更多相关《计算机专业文献翻译性能测试方法.docx（17页珍藏版）》请在冰豆网上搜索。

计算机专业文献翻译性能测试方法

届毕业设计（论文）

英文参考文献

英文文献1：

DatabaseSecurity

文献出处，年，Vol.卷（期）NetworkSecurityVolume:

2003,Issue:

6,June,2003,pp.11-12

作者：

PaulMorrison

英文文献2：

ApproachestoPerformanceTesting

文献出处，年，Vol.卷（期）ApproachestoPerformanceTestingVol.18,No.3,pp.312-319,2000

作者：

MattMaccaux

学生院系

专业名称学生班级

学生学号学生姓名

学生层次

ApproachestoPerformanceTesting

byMattMaccaux

09/12/2005

Abstract

Therearemanydifferentwaystogoaboutperformancetestingenterpriseapplications,someofthemmoredifficultthanothers.Thetypeofperformancetestingyouwilldodependsonwhattypeofresultsyouwanttoachieve.Forexample,forrepeatability,benchmarktestingisthebestmethodology.However,totesttheupperlimitsofthesystemfromtheperspectiveofconcurrentuserload,capacityplanningtestsshouldbeused.Thisarticlediscussesthedifferencesandexaminesvariouswaystogoaboutsettingupandrunningtheseperformancetests.

Introduction

PerformancetestingaJ2EEapplicationcanbeadauntingandseeminglyconfusingtaskifyoudon'tapproachitwiththeproperplaninplace.Aswithanysoftwaredevelopmentprocess,youmustgatherrequirements,understandthebusinessneeds,andlayoutaformalschedulewellinadvanceoftheactualtesting.Therequirementsfortheperformancetestingshouldbedrivenbytheneedsofthebusinessandshouldbeexplainedwithasetofusecases.Thesecanbebasedonhistoricaldata（say,whattheloadpatternwasontheserverforaweek）oronapproximationsbasedonanticipatedusage.Onceyouhaveanunderstandingofwhatyouneedtotest,youneedtolookathowyouwanttotestyourapplication.

Earlyoninthedevelopmentcycle,benchmarktestsshouldbeusedtodetermineifanyperformanceregressionsareintheapplication.Benchmarktestsaregreatforgatheringrepeatableresultsinarelativelyshortperiodoftime.Thebestwaytobenchmarkistochangeoneandonlyoneparameterbetweentests.Forexample,ifyouwanttoseeifincreasingtheJVMmemoryhasanyimpactontheperformanceofyourapplication,incrementtheJVMmemoryinstages（forexample,goingfrom1024MBto1224MB,thento1524MB,andfinallyto2024MB）andstopateachstagetogathertheresultsandenvironmentdata,recordthisinformation,andthenmoveontothenexttest.Thiswayyou'llhaveacleartrailtofollowwhenyouareanalyzingtheresultsofthetests.Inthenextsection,Idiscusswhatabenchmarktestlookslikeandthebestparametersforrunningthesetests.

Lateroninthedevelopmentcycle,afterthebugshavebeenworkedoutoftheapplicationandithasreachedastablepoint,youcanrunmorecomplextypesofteststodeterminehowthesystemwillperformunderdifferentloadpatterns.Thesetypesoftestsarecalledcapacityplanning,soaktests,andpeak-resttests,andaredesignedtotest"real-world"-typescenariosbytestingthereliability,robustness,andscalabilityoftheapplication.ThedescriptionsIusebelowshouldbetakenintheabstractsensebecauseeveryapplication'susagepatternwillbedifferent.Forexample,capacity-planningtestsaregenerallyusedwithslowramp-ups（definedbelow）,butifyourapplicationseesquickburstsoftrafficduringaperiodoftheday,thencertainlymodifyyourtesttoreflectthis.Keepinmind,though,thatasyouchangevariablesinthetest（suchastheperiodoframp-upthatItalkabouthereorthe"think-time"oftheusers）theoutcomeofthetestwillvary.Itisalwaysagoodideatorunaseriesofbaselinetestsfirsttoestablishaknown,controlledenvironmenttocompareyourchangeswithlater.

Benchmarking

Thekeytobenchmarktestingistohaveconsistentlyreproducibleresults.Resultsthatarereproducibleallowyoutodotwothings:

reducethenumberoftimesyouhavetorerunthosetests;andgainconfidenceintheproductyouaretestingandthenumbersyouproduce.Theperformance-testingtoolyouusecanhaveagreatimpactonyourtestresults.Assumingtwoofthemetricsyouarebenchmarkingaretheresponsetimeoftheserverandthethroughputoftheserver,theseareaffectedbyhowmuchloadisputontotheserver.Theamountofloadthatisputontotheservercancomefromtwodifferentareas:

thenumberofconnections（orvirtualusers）thatarehittingtheserversimultaneously;andtheamountofthink-timeeachvirtualuserhasbetweenrequeststotheserver.Obviously,themoreusershittingtheserver,themoreloadwillbegenerated.Also,theshorterthethink-timebetweenrequestsfromeachuser,thegreatertheloadwillbeontheserver.Combinethosetwoattributesinvariouswaystocomeupwithdifferentlevelsofserverload.Keepinmindthatasyouputmoreloadontheserver,thethroughputwillclimb,toapoint.

Figure1.Thethroughputofthesysteminpagespersecondasloadincreasesovertime

Notethatthethroughputincreasesataconstantrateandthenatsomepointlevelsoff.

Atsomepoint,theexecutequeuestartsgrowingbecauseallthethreadsontheserverwillbeinuse.Theincomingrequests,insteadofbeingprocessedimmediately,willbeputintoaqueueandprocessedwhenthreadsbecomeavailable.

Figure2.Theexecutequeuelengthofthesystemasloadincreasesovertime

Notethatthequeuelengthiszeroforaperiodoftime,butthenstartstogrowataconstantrate.Thisisbecausethereisasteadyincreaseinloadonthesystem,andalthoughinitiallythesystemhadenoughfreethreadstocopewiththeadditionalload,eventuallyitbecameoverwhelmedandhadtostartqueuingthemup.

Whenthesystemreachesthepointofsaturation,thethroughputoftheserverplateaus,andyouhavereachedthemaximumforthesystemgiventhoseconditions.However,asserverloadcontinuestogrow,theresponsetimeofthesystemalsogrowsevenasthethroughputplateaus.

Figure3.Theresponsetimesoftwotransactionsonthesystemasloadincreasesovertime

Notethatatthesametimeastheexecutequeue（above）startstogrow,theresponsetimealsostartstogrowatanincreasedrate.Thisisbecausetherequestscannotbeservedimmediately.

Tohavetrulyreproducibleresults,thesystemshouldbeputunderahighloadwithnovariability.Toaccomplishthis,thevirtualusershittingtheservershouldhave0secondsofthink-timebetweenrequests.Thisisbecausetheserverisimmediatelyputunderloadandwillstartbuildinganexecutequeue.Ifthenumberofrequests（andvirtualusers）iskeptconsistent,theresultsofthebenchmarkingshouldbehighlyaccurateandveryreproducible.

Onequestionyoushouldraiseis,"Howdoyoumeasuretheresults?

"Anaverageshouldbetakenoftheresponsetimeandthroughputforagiventest.Theonlywaytoaccuratelygetthesenumbersthoughistoloadalltheusersatonce,andthenrunthemforapredeterminedamountoftime.Thisiscalleda"flat"run.

Figure4.Thisiswhataflatrunlookslike.Alltheusersareloadedsimultaneously.

Theoppositeisknownasa"ramp-up"run.

Figure5.Thisiswhataramp-uprunlookslike.Theusersareaddedataconstantrate（xnumberpersecond）throughoutthedurationofthetest.

Theusersinaramp-uprunarestaggered（addingafewnewuserseveryxseconds）.Theramp-uprundoesnotallowforaccurateandreproducibleaveragesbecausetheloadonthesystemisconstantlychangingastheusersarebeingaddedafewatatime.Therefore,theflatrunisidealforgettingbenchmarknumbers.

Thisisnottodiscountthevalueinrunningramp-up-styletests.Infact,ramp-uptestsarevaluableforfindingtheballparkinwhichyouthinkyoulaterwanttorunflatruns.Thebeautyofaramp-uptestisthatyoucanseehowthemeasurementschangeastheloadonthesystemchanges.Thenyoucanpicktherangeyoulaterwanttorunwithflattests.

Theproblemwithflatrunsisthatthesystemwillexperience"wave"effects.

Figure6.Thethroughputofthesysteminpagespersecondasmeasuredduringaflatrun

Notetheappearanceofwavesovertime.Thethroughputisnotsmoothbutratherresemblesawavepattern.

ThisisvisiblefromallaspectsofthesystemincludingtheCPUutilization.

Figure7.TheCPUutilizationofthesystemovertime,asmeasuredduringaflatrun

Notetheappearanceofwavesoveraperiodoftime.TheCPUutilizationisnotsmoothbutratherhasverysharppeaksthatresemblethethroughputgraph'swaves.

Additionally,theexecutequeueexperiencesthisunstableload,andthereforeyouseethequeuegrowingandshrinkingastheloadonthesystemincreasesanddecreasesovertime.

Figure8.Theexecutequeueofthesystemovertimeasmeasuredduringaflatrun

Notetheappearanceofwavesovertime.TheexecutequeueexactlymimicstheCPUutilizationgraphabove.

Finally,theresponsetimeofthetransactionsonthesystemwillalsoresemblethiswavepattern.

Figure9.Theresponsetimeofatransactiononthesystemovertimeasmeasuredduringaflatrun

Notetheappearanceofwavesovertime.Thetransactionresponsetimelinesupwiththeabovegraphs,buttheeffectisdiminishedovertime.

Thisoccurswhenalltheusersaredoingapproximatelythesamethingatthesametimeduringthetest.Thiswillproduceveryunreliableandinaccurateresults,sosomethingmustbedonetocounteractthis.Therearetwowaystogainaccuratemeasurementsfromthesetypesofresults.Ifthetestisallowedtorunforaverylongduration（sometimesseveralhours,dependingonhowlongoneuseriterationtakes）eventuallyanaturalsortofrandomnesswillsetinandthethroughputoftheserverwill"flattenout."Alternatively,measurementscanbetakenonlybetweentwoofthebreaksinthewaves.Th