Packaging-WebResponseTimeReductionForSlowandBusyWebServers HiroshiFujinoki DepartmentofComputerScienceSouthernIllinoisUniversityEdwardsvilleEdwardsville,Illinois62026-1656,USA E-mail:hfujino@siue.edu KiranK.Gollamudi DepartmentofElectricalEngineeringSouthernIllinoisUniversityEdwardsvilleEdwardsville,Illinois62026-1656,USA E-mail:kgoramudi@siue.edu Abstract Inthispaper,anongoingresearchactivitytoreduceresponsetimeinwebfiletransmissionsispresented.Anewtransmissiontechnique,objectpackaging,isproposedforreducedresponsetimeforwebbrowsing.Objectpackagingaimstoreduceoverheadnotonlyatroutersonatransmissionpathbutatatransmittingwebserveforbetterresponsetime.Theobjectpackagingintendstoaddresstwodifferentissuesinwebfiletransmissions;1)diskessoverheadsand2)protocolprocessingoverheads.Experimentalstudyshowedthatthefiletransmissionsbyobjectpackagingreducedresponsetime,numberoftransferredbytesandthepacketsby34.7,40.1and7.1%,respectively,forfileswiththeiraveragefilesizebeing10Kbytes,whichistheaveragefilesizeinthewebtrafficinthecurrent.
1.Introduction Inthecurrent,webtrafficupiesthelargestportionofthetraffic.SincetheadventofMosaicin1993,thetrafficonthehasbeenincreasingatanexponentialpaceandthistrendstillexistseventhesedays[1].Asaresult,oneofthemostseriousproblemswebusersareexperiencingislongresponsetimeandtransmissiondelayduetohighwebtrafficload[2]. SincetheWorldWideWeb(calledwwwhereafter)isbasedontheclient-servermodel,delays,whicharethecausesoflongerresponsetime,areinsertedatvariousplaces.However,majorsourcesofdelayareclassifiedinthefollowingtwogroups:1)delayduetooverheadattheserversideand2)delayduringpropagationinthework.Thedelayduetoserver-sideoverheadconsistsofoperatingsystemoverhead,suchasfileesslatency,memorycopiesandpacketprocessing overheadbyworkprotocol.Theoperatingsystemoverheadwillbeaproblem,1)ifaserverisoverwhelmedbyexcessrequests(i.e.,abusyserver)or2)ifaserverisdrivenbyaslowCPUordoesnothaveenoughmemory(i.e.,aslowserver).OverheadformemorycopiesandpacketprocessingsuchasCRCchecksumcalculationwillberelativelyhighforsmallpackets.Thesecondgroup,thedelayduringpropagationinthework,consistsofqueuingdelayandroutingtablelook-upsatintermediaterouters,anddelayduetopacketretransmissionsduetopacketlossesworkcongestion.Inthisstudy,thedelayduetopacketlossesduringtransmissionisnotconsidered.
2.RelatedWork Inordertoreduceresponsetimeinwebfiletransmission,varioussolutionshavebeenproposed.Fortheoverheadinoperatingsystem,Ousterhout[3]andDruschel[4]analyzedmajorfactorsintheoverheadinoperatingsystemsandfoundthatthememorycopyanddiskI/Oarethelargestoverheadsthatlimitthroughputforfileserverprocesses.MarkatossuggestedthatreducingthenumberofmemorycopieswouldimproveTCP/IPperformanceforHTTP[5].Busarisuggestedthatthemajorityofwebtrafficconsistsofrequeststomanysmallfilesandonlyafewlargefilesexist[6](averagefilesizeisabout10Kbytesintheweb[1]).Sincediskessoverheaddominatestheoperatingsystemoverhead,server-sidecachingisoneofthetechniquestoreducethedelayduetodiskess.However,manyoftherequestedfiles(50to70%)arereferencedonlyonce[6],implyingthelimitationofserver-sidecachingtoimproveresponsetimeduetooperatingsystemoverhead[7,8,9,10].
3.FileTransmissionbyObjectPackaging Themajorpremiseofwebfiletransmissionsbyobjectpackagesisbasedonthefactthatmosthomepagesconsistofmanysmallfiles(fileslessthan16Kbytes).Sincethemajorityoftheserver-sideesfromoperatingsystemoverheadtoretrieverequestedfilesfromalocaldisk,theserver-sideoverheadcanbereducedifthenumberofrequestedfilesisreduced.Moreover,ifwecanreducethenumberofrequestedfiles,thenumberofpacketstobetransferredinworkwillbereduced,byavoidingpacketfragmentation. Basedonthefactsdescribedinthepreviousparagraph,thewebfiletransmissionbyobjectpackagingisproposed.Anobjectpackageisacollectionofwebfilesinahomepagepackedinanpressedarchivefile.Multiplefiles,suchasHTMLtextandimagefiles,willbepackedinanobjectpackagefileforefficienttransmission.Thosefilesinahomepagearesequentiallypackedintoanobjectpackagefilepressiontopreventpressionoverheadatthereceiverside.Atthereceivingside,awebbrowsershouldunpackthefilesinanobjectpackagefile.Objectpackagesshouldberecognizedbyaspecificfileextension.TheformatofanobjectpackageisshowninFigure1. ObjectPackage ObjectInformationField Datafortherequestedfiles File#1Data File#nData Fileheaderforanobjectpackage Figure1—ponentsoftheobjectpackage Thefirstelementinanobjectpackageistheobjectinformationfield.Theobjectinformationfieldcontainstheinformationofthepackedfiles.Thedatafieldcontainsthecontentsoftherequestedfiles.Areceiverreconstructstherequestedfilesusingtheinformationintheobjectinformationfieldandthecontents(fromdatafield)oftherequestedfiles.Theobjectinformationfieldconsistsoffoursubfields(Figure2). ObjectInformationField Filenamesub-fieldFilesizesub-fieldFileattributesub-field Numberofobjects(files)sub-field Figure2—Elementsinobjectinformationfield Thefirstsubfieldisthenumberofpackedfiles.Thesecondisthefilenamesubfield,whichisacollectionofthenamesofthecontainedfiles.Thethirdsubfieldcontainsthefilesizesofthepackedfiles.Thelastsubfieldindicatestheattributesofeachfilesuchasbinary,text,orexecutable.Thefirstfile(File#1)willbereconstructedbasedonthefirstelementinthefilename,thefilesizeandthefileattributesubfields.
4.SimulationExperiments Simulationexperimentswereperformedtoevaluatetheperformanceofwebfiletransmissionbyobjectpackaging.Thewebservermachineandaclientmachineareconnectedthroughahub.Transmissionrateis10Mbpsforallofthelinks.Alloftherequestedfilesareinitiallystoredinthelocalharddriveattheservermachine.Tocaptureworktrafficbetweentheserverandtheclient,anothermachine,calledthepacketmonitor,isconnectedtothehub.ThetestbedisshowninFigure3. RequestedFiles Requests WebServer HubClient LocalDisk TrafficMonitor Figure3-Thetestbedforsimulationexperiments Usingthetestbed,wemeasuredthefollowingfactorsforfilestransferredbytheexistingmethodandobjectpackages:1)numberofactuallytransferredbytes(includingpacketheader),2)numberoftransferredpackets,and3)theresponsetime.Theresponsetimewasmeasuredbythetrafficmonitorandisdefinedtobethetimebetweenthefirstpackettransferredfromtheclienttothewebserver(i.e.,theTCPSYNCmessage)andthelastpackettransferredfromthewebservertotheclient(i.e.,thelastpacketfortherequestedfile). Tables1through3showtheresultsoftheexperiments.Table1showsparisonsoftheexistingfiletransmissionandtheobjectpackaginginthenumberofbytesactuallytransferred.Table2showsthenumberofpacketsactuallytransferredandTable3showstheaverageresponsetimeinseconds.Figure4showstheperformanceoftheobjectpackagingrelativetotheexistingtransmissionmethod.Forsmallfiles(averagefilesizeof10Kbytesorless),theobjectpackagingdemonstratedasignificantreductioninthenumberoftransferredbytesandpackets,andresponsetime.Whentheaveragefilesizeis1Kbytes,43.7%,87.8%and80.7%reductionweremadeforthetransferredbytes,transferredpacketsandtheresponsetime. Table1—Averagetransferredbytes ExistingMethodObjectPackage 1K 188362.3 106057.1 4K 491916.1 409423.9 10K 1094350.5 1016420.7 Table2—Averagenumberofpackets ExistingMethodObjectPackage 1K 944.2 115.3 4K 1189.2 420.2 10K 1720.4 1015.6 Table3—Averageresponsetime(inseconds) ExistingMethodObjectPackage 1K 0.696 0.134 4K 0.927 0.403 10K 1.531 0.999 Percentagetotheexistingmethod 100%80%60% Transferredbytes Responsetime 40% Transferredpackets20% 0%1K Figure4— 4K 10K Averagefilesize(inbytes) Resultsfordifferentfilesizes Figure5showstheresultsfordifferentnumbersoftransferredfiles.Theresultsshowedthattherelativeperformanceoftheobjectpackagingwouldnotbedependentonthenumberoftransferredfilesaslongasatleastmorethantenfilesaretransferred,meaningthattheadvantagesoftheobjectpackagingwillbeeffectiveevenfortransferringasmallnumberoffiles. Percentagetotheexistingmethod 100%80%60% Filesize=10K Filesize=4KFilesize=1K 40% 20% 0%
0 Figure5— 25 50 75 100 Numberoftransferredfiles Resultsfordifferentnumberoffiles
5.ConclusionsandFutureWork Inthispaper,anongoingresearchactivityforefficientwebfiletransmissions,calledmultiplefiletransmissionbyobjectpackage,isproposed.Thetechniquedoesnotrequireanymodificationofanoperatingsystemattheserversidenortransmissionprotocolatrouters.Experimentsshowthatobjectpackagingisefficientinreducingresponsetime.Itwasobservedthattheobjectpackagingreducedresponsetime,numberoftransferredbytesandthepacketsby34.7,40.1and7.1%,respectively,forfileswiththeaveragefilesizebeing10Kbytes,whichistheaveragefilesizeinthewebinthecurrent.Currently,toobservethescalabilityoftheobjectpackagingtransmission,experimentsarebeingperformed.
6.References [1]
M.ArlittandC.Williamson,“WebServerWorkloadCharacterization:TheSearchforInvariants,”Proceedingsofthe1996ACMSIGMETRICSConferenceontheMeasurementandModelingofComputerSystems,May1996,pp.126-137. [2]GVU’sWWWUserSurveys,iaInstituteofTechnologyURL:http://www.gvu.gatech.edu/user_surveys [3]
J.Ousterhout,“WhyAren'tOperatingSystemsGettingFasterAsHardware?
,”ProceedingsofSummer1990USENIXConference,June1990,pp.247-256. [4]
P.Druschel,“OperatingSystemSupportforHigh-SpeedNetworking,”CommunicationsoftheACM,vol.39,no.2,September1996,pp.41-51. [5]
P.Markatos,“Speeding-upTCP/IP:FasterProcessorsArenotEnough,”Proceedingsofthe21stIEEEInternationalPerformance,Computing,andCommunicationsConference,April2002,pp.341-345. [6]
M.BusariandC.Williamson,“OntheSensitivityofWebProxyCachePerformancetoWorkloadCharacteristics,”ProceedingsofIEEEINFOCOM,April2001,pp.1225-1234. [7]
J.Dilley,“TheEffectofConsistencyonCacheResponseTime,”IEEENetwork,vol.14,no.3,May/June2000,pp.24-28. [8]
S.Glassman,“AcachingrelayfortheWorld-WideWeb,”ComputerNetworksandISDNSystems,vol.27,no.2,October1994,pp.165-173. [9]
D.Lee,“Pre-FetchDocumentCachingtoImproveWorld-WideWebUserResponseTime,”Master'sThesis.VirginiaPolytechnicInstituteandStateUniversity,March1996. [10]
J.Mogul,“SqueezingMoreBitsOutofHTTPCaches,”IEEENetwork,vol.14,no.3,May/June2000,pp.6-14.