1英文翻譯：本文出自ComputerNetwork第四版AndrewS.Tanenbaum著NetworkPerformanceMeasurementWhenanetworkperformspoorly,itsusersoftencomplaintothefolksrunningit,demandingimprovements.Toimprovetheperformance,theoperatorsmustfirstdetermineexactlywhatisgoingon.Tofindoutwhatisreallyhappening,theoperatorsmustmakemeasurements.Inthissectionwewilllookatnetworkperformancemeasurements.ThediscussionbelowisbasedontheworkofMogul(1993).Thebasicloopusedtoimprovenetworkperformancecontainsthefollowingsteps:1.Measuretherelevantnetworkparametersandperformance.2.Trytounderstandwhatisgoingon.3.Changeoneparameter.Thesestepsarerepeateduntiltheperformanceisgoodenoughoritisclearthatthelastdropofimprovementhasbeensqueezedout.Measurementscanbemadeinmanywaysandatmanylocations(bothphysicallyandintheprotocolstack).Themostbasickindofmeasurementistostartatimerwhenbeginningsomeactivityandseehowlongthatactivitytakes.Forexample,knowinghowlongittakesforaTPDUtobeacknowledgedisakeymeasurement.Othermeasurementsaremadewithcountersthatrecordhowoftensomeeventhashappened(e.g.,numberoflostTPDUS).Finally,oneisofteninterestedinknowingtheamountofsomething,suchasthenumberofbytesprocessedinacertaintimeinterval.Measuringnetworkperformanceandparametershasmanypotentialpitfalls.Belowwelistafewofthem.Anysystematicattempttomeasurenetworkperformanceshouldbecarefultoavoidthese.MakeSureThattheSampleSizeIsLargeEnoughDonotmeasurethetimetosendoneTPDU,butrepeatthemeasurement,say,onemilliontimesandtaketheaverage.Havingalargesamplewillreducetheuncertaintyinthemeasuredmeanandstandarddeviation.Thisuncertaintycanbecomputedusingstandardstatisticalformulas.MakeSureThattheSamplesAreRepresentative2Ideally,thewholesequenceofonemillionmeasurementsshouldberepeatedatdifferenttimesofthedayandtheweektoseetheeffectofdifferentsystemloadsonthemeasuredquantity.Measurementsofcongestion,forexample,areoflittleuseiftheyaremadeatamomentwhenthereisnocongestion.Sometimestheresultsmaybecounterintuitiveatfirst,suchasheavycongestionat10,11,1,and2oclock,butnocongestionatnoon(whenalltheusersareawayatlunch).BeCarefulWhenUsingaCoarse-GrainedClockComputerclocksworkbyincrementingsomecounteratregularintervals.Forexample,amillisecondtimeradds1toacounterevery1msec.Usingsuchatimertomeasureaneventthattakeslessthan1msecispossible,butrequiressomecare.(Somecomputershavemoreaccurateclocks,ofcourse.)TomeasurethetimetosendaTPDU,forexample,thesystemclock(say,inmilliseconds)shouldbereadoutwhenthetransportlayercodeisenteredandagainwhenitisexited.IfthetrueTPDUsendtimeis300sec,thedifferencebetweenthetworeadingswillbeeither0or1,bothwrong.However,ifthemeasurementisrepeatedonemilliontimesandthetotalofallmeasurementsaddedupanddividedbyonemillion,themeantimewillbeaccuratetobetterthan1sec.BeSureThatNothingUnexpectedIsGoingOnduringYourTestsMakingmeasurementsonauniversitysystemthedaysomemajorlabprojecthastobeturnedinmaygivedifferentresultsthanifmadethenextday.Likewise,ifsomeresearcherhasdecidedtorunavideoconferenceoveryournetworkduringyourtests,youmaygetabiasedresult.Itisbesttoruntestsonanidlesystemandcreatetheentireworkloadyourself.Eventhisapproachhaspitfallsthough.Whileyoumightthinknobodywillbeusingthenetworkat3A.M.,thatmightbepreciselywhentheautomaticbackupprogrambeginscopyingallthediskstotape.Furthermore,theremightbeheavytrafficforyourwonderfulWorldWideWebpagesfromdistanttimezones.CachingCanWreakHavocwithMeasurementsTheobviouswaytomeasurefiletransfertimesistoopenalargefile,readthewholething,closeit,andseehowlongittakes.Thenrepeatthemeasurementmanymoretimestogetagoodaverage.Thetroubleis,thesystemmaycachethefile,soonlythefirstmeasurementactuallyinvolvesnetworktraffic.Therestarejustreadsfromthe3localcache.Theresultsfromsuchameasurementareessentiallyworthless(unlessyouwanttomeasurecacheperformance).Oftenyoucangetaroundcachingbysimplyoverflowingthecache.Forexample,ifthecacheis10MB,thetestloopcouldopen,read,andclosetwo10-MBfilesoneachpass,inanattempttoforcethecachehitrateto0.Still,cautionisadvisedunlessyouareabsolutelysureyouunderstandthecachingalgorithm.Bufferingcanhaveasimilareffect.OnepopularTCP/IPperformanceutilityprogramhasbeenknowntoreportthatUDPcanachieveaperformancesubstantiallyhigherthanthephysicallineallows.Howdoesthisoccur?AcalltoUDPnormallyreturnscontrolassoonasthemessagehasbeenacceptedbythekernelandaddedtothetransmissionqueue.Ifthereissufficientbufferspace,timing1000UDPcallsdoesnotmeanthatallthedatahavebeensent.Mostofthemmaystillbeinthekernel,buttheperformanceutilitythinkstheyhaveallbeentransmitted.UnderstandWhatYouAreMeasuringWhenyoumeasurethetimetoreadaremotefile,yourmeasurementsdependonthenetwork,theoperatingsystemsonboththeclientandserver,theparticularhardwareinterfaceboardsused,theirdrivers,andotherfactors.Ifthemeasurementsaredonecarefully,youwillultimatelydiscoverthefiletransfertimefortheconfigurationyouareusing.Ifyourgoalistotunethisparticularconfiguration,thesemeasurementsarefine.However,ifyouaremakingsimilarmeasurementsonthreedifferentsystemsinordertochoosewhichnetworkinterfaceboardtobuy,yourresultscouldbethrownoffcompletelybythefactthatoneofthenetworkdriversistrulyawfulandisonlygetting10percentoftheperformanceoftheboard.4網(wǎng)絡(luò)性能的測量當(dāng)一個網(wǎng)絡(luò)的運(yùn)行效果很差的時候，它的用戶通常會向網(wǎng)絡(luò)運(yùn)行商抱怨并要求提高網(wǎng)絡(luò)的質(zhì)量。為了改善網(wǎng)絡(luò)的性能，網(wǎng)絡(luò)操作人員首先必須確定發(fā)生了什么問題。為了找出真正的問題所在，操作人員必須進(jìn)行測量工作。在這一小節(jié)中，我們來看一看網(wǎng)絡(luò)性能的測量問題。下面的討論以Mogul(1993)的工作為基礎(chǔ)。用來改善網(wǎng)絡(luò)性能的基本循環(huán)過程包括以下步驟：（1）測量有關(guān)的網(wǎng)絡(luò)參數(shù)和性能。（2）試圖理解當(dāng)前的網(wǎng)絡(luò)狀況。（3）改變一個參數(shù)。這些步驟不斷重復(fù)，直到網(wǎng)絡(luò)的性能已經(jīng)足夠好，或者改善性能的全部空間都已經(jīng)被發(fā)掘出來了。測量工作可以有許多做法，也可以在許多地點(diǎn)或場所進(jìn)行（既指物理位置，也指協(xié)議棧中的位置）。最基本的一種測量手段是：在開始某一個動作的時候啟動一個定時器，然后確定該需要多長時間。例如，知道一個TPDU需要多長時間才被確認(rèn)是一個很關(guān)鍵的測量指標(biāo)。其他有一些測量指標(biāo)可以通過計數(shù)器來完成，即記錄某種事件發(fā)生的次數(shù)，比如丟失的TPDU的數(shù)量。最后，人們通常對于某些事物的數(shù)量比較感興趣，比如在特定的時間間隔內(nèi)所處理的字節(jié)數(shù)。測量網(wǎng)絡(luò)的性能和參數(shù)有許多潛在的陷阱。以下我們列出其中一部分。任何一種系統(tǒng)化的網(wǎng)絡(luò)性能測量