A scalable and highly available system for serving

AScalableandHighlyAvailableSystemforServingDynamicDataatFrequentlyAccessedWebSitesJimChallenger,PaulDantzig,andArunIyengarIBMResearchT.J.WatsonResearchCenterP.O.Box704YorktownHeights,NY10598Abstract:ThispaperdescribesthesystemandkeytechniquesusedforachievingperformanceandhighavailabilityattheofficialWebsiteforthe1998OlympicWinterGameswhichwasoneofthemostpopularWebsitesforthedurationoftheOlympicGames.TheWebsiteutilizedthirteenSP2systemsscatteredaroundtheglobecontainingatotalof143processors.AkeyfeatureoftheWebsitewasthatthedatabeingpresentedtoclientswasconstantlychanging.Whenevernewresultswereenteredintothesystem,updatedWebpagesreflectingthechangesweremadeavailabletotherestoftheworldwithinseconds.Onetechniqueweusedtoservedynamicdataefficientlytoclientswastocachedynamicpagessothattheyonlyhadtobegeneratedonce.WedevelopedandimplementedanewalgorithmwecallDataUpdatePropagation(DUP)whichidentifiesthecachedpagesthathavebecomestaleasaresultofchangestounderlyingdataonwhichthecachedpagesdepend,suchasdatabases.FortheOlympicGamesWebsite,wewereabletoupdatestalepagesdirectlyinthecachewhichobviatedtheneedtoinvalidatethem.Thisallowedustoachievecachehitratesofcloseto100%.OursystemwasabletoservepagestoclientsquicklyduringtheentireOlympicGamesevenduringpeakperiods.Inaddition,thesitewasavailable100%ofthetime.Wedescribethekeyfeaturesemployedbyoursiteforhighavailability.WealsodescribehowtheWebsitewasstructuredtoprovideusefulinformationwhilerequiringclientstoexamineonlyasmallnumberofpages.1.IntroductionThispaperdescribesthesystemandkeytechniquesusedforachievingperformanceandhighavailabilityattheofficialWebsiteforthe1998OlympicWinterGamesinNagano,1of308/6/9810:36AMAScalableandHighlyAvailableSystemf...icDataatFrequentlyAccessedWebSites(DUP)whichidentifiesthecachedpagesthathavebecomestaleasaresultofchangestounderlyingdataonwhichthecachedpagesdepend,suchasdatabases.Astalepagecaneitherbeinvalidatedorupdateddirectlyinthecache.FortheOlympicGamesWebsite,wewereabletoupdatestalepagesdirectlyinthecache.Thisallowedustoachievecachehitratesofcloseto100%.Bycontrast,anearlierversionofoursystemwhichdidnotuseDUPachievedcachehitratesofaround80%attheofficialWebsiteforthe1996OlympicGames.Thetechniquesdeployedatthe1998OlympicGamesWebsitecanbeusedatotherWebsiteswhereitisdesirabletoprovidesignificantdynamiccontent.Websitedesignersoftenshyawayfromprovidingtoomanydynamicpagesbecauseoftheoverheadthatthisentails.Wehavedemonstratedthatthisdoesnothavetobethecase.Usingourtechniques,asingleservercanserveseveralhundreddynamicpagespersecondifthepagesarecacheable.ThereisconsiderabledemandforprovidingdynamiccontentefficientlyontheWeb.Consequently,ourtechnologyisbeingincorporatedintoIBMproductssothatitcanbeusedatmanyWebsites.The1998OlympicGamesWebsiteachievedhighavailabilitybyusingredundanthardwareandbyservingpagesfromfourdifferentcomplexesindifferentgeographiclocationscontainingreplicatedinformation.Ifaserverfailed,requestswereautomaticallyroutedtootherservers.Ifanentirecomplexfailed,requestscouldberoutedtotheotherones.Thenetworkcontainedredundantpathstoeliminatesinglepointsoffailure.Itwasdesignedtohandleatleasttwotothreetimestheexpectedbandwidthinordertoaccommodatethehighvolumesofdatashouldportionsofthenetworkfail.Therestofthispaperisorganizedasfollows.Section2describestheDUPalgorithmfor2of308/6/9810:36AMAScalableandHighlyAvailableSystemf...icDataatFrequentlyAccessedWebSites:staticanddynamic.Staticpagesareservedfromthefilesystem.Bycontrast,dynamicpagesarecreate