数据库系统概念全套配套课件PPT ch17.ppt 49页

'Chapter17:DatabaseSystemArchitectures Chapter17:DatabaseSystemArchitecturesCentralizedandClient-ServerSystemsServerSystemArchitecturesParallelSystemsDistributedSystemsNetworkTypes CentralizedSystemsRunonasinglecomputersystemanddonotinteractwithothercomputersystems.General-purposecomputersystem:onetoafewCPUsandanumberofdevicecontrollersthatareconnectedthroughacommonbusthatprovidesaccesstosharedmemory.Single-usersystem(e.g.,personalcomputerorworkstation):desk-topunit,singleuser,usuallyhasonlyoneCPUandoneortwoharddisks;theOSmaysupportonlyoneuser.Multi-usersystem:moredisks,morememory,multipleCPUs,andamulti-userOS.Servealargenumberofuserswhoareconnectedtothesystemvieterminals.Oftencalledserversystems. ACentralizedComputerSystem Client-ServerSystemsServersystemssatisfyrequestsgeneratedatmclientsystems,whosegeneralstructureisshownbelow: Client-ServerSystems(Cont.)Databasefunctionalitycanbedividedinto:Back-end:managesaccessstructures,queryevaluationandoptimization,concurrencycontrolandrecovery.Front-end:consistsoftoolssuchasforms,report-writers,andgraphicaluserinterfacefacilities.Theinterfacebetweenthefront-endandtheback-endisthroughSQLorthroughanapplicationprograminterface. Client-ServerSystems(Cont.)Advantagesofreplacingmainframeswithnetworksofworkstationsorpersonalcomputersconnectedtoback-endservermachines:betterfunctionalityforthecostflexibilityinlocatingresourcesandexpandingfacilitiesbetteruserinterfaceseasiermaintenance ServerSystemArchitectureServersystemscanbebroadlycategorizedintotwokinds:transactionserverswhicharewidelyusedinrelationaldatabasesystems,anddataservers,usedinobject-orienteddatabasesystems TransactionServersAlsocalledqueryserversystemsorSQLserversystemsClientssendrequeststotheserverTransactionsareexecutedattheserverResultsareshippedbacktotheclient.RequestsarespecifiedinSQL,andcommunicatedtotheserverthrougharemoteprocedurecall(RPC)mechanism.TransactionalRPCallowsmanyRPCcallstoformatransaction.OpenDatabaseConnectivity(ODBC)isaClanguageapplicationprograminterfacestandardfromMicrosoftforconnectingtoaserver,sendingSQLrequests,andreceivingresults.JDBCstandardissimilartoODBC,forJava TransactionServerProcessStructureAtypicaltransactionserverconsistsofmultipleprocessesaccessingdatainsharedmemory.ServerprocessesThesereceiveuserqueries(transactions),executethemandsendresultsbackProcessesmaybemultithreaded,allowingasingleprocesstoexecuteseveraluserqueriesconcurrentlyTypicallymultiplemultithreadedserverprocessesLockmanagerprocessMoreonthislaterDatabasewriterprocessOutputmodifiedbufferblockstodiskscontinually TransactionServerProcesses(Cont.)LogwriterprocessServerprocessessimplyaddlogrecordstologrecordbufferLogwriterprocessoutputslogrecordstostablestorage.CheckpointprocessPerformsperiodiccheckpointsProcessmonitorprocessMonitorsotherprocesses,andtakesrecoveryactionsifanyoftheotherprocessesfailE.g.,abortinganytransactionsbeingexecutedbyaserverprocessandrestartingit TransactionSystemProcesses(Cont.) TransactionSystemProcesses(Cont.)SharedmemorycontainsshareddataBufferpoolLocktableLogbufferCachedqueryplans(reusedifsamequerysubmittedagain)AlldatabaseprocessescanaccesssharedmemoryToensurethatnotwoprocessesareaccessingthesamedatastructureatthesametime,databasessystemsimplementmutualexclusionusingeitherOperatingsystemsemaphoresAtomicinstructionssuchastest-and-setToavoidoverheadofinterprocesscommunicationforlockrequest/grant,eachdatabaseprocessoperatesdirectlyonthelocktableinsteadofsendingrequeststolockmanagerprocessLockmanagerprocessstillusedfordeadlockdetection DataServersUsedinhigh-speedLANs,incaseswhereTheclientsarecomparableinprocessingpowertotheserverThetaskstobeexecutedarecomputeintensive.Dataareshippedtoclientswhereprocessingisperformed,andthenshippedresultsbacktotheserver.Thisarchitecturerequiresfullback-endfunctionalityattheclients.Usedinmanyobject-orienteddatabasesystemsIssues:Page-ShippingversusItem-ShippingLockingDataCachingLockCaching DataServers(Cont.)Page-shippingversusitem-shippingSmallerunitofshippingmoremessagesWorthprefetchingrelateditemsalongwithrequesteditemPageshippingcanbethoughtofasaformofprefetchingLockingOverheadofrequestingandgettinglocksfromserverishighduetomessagedelaysCangrantlocksonrequestedandprefetcheditems;withpageshipping,transactionisgrantedlockonwholepage.LocksonaprefetcheditemcanbeP{calledback}bytheserver,andreturnedbyclienttransactioniftheprefetcheditemhasnotbeenused.Locksonthepagecanbedeescalatedtolocksonitemsinthepagewhentherearelockconflicts.Locksonunuseditemscanthenbereturnedtoserver. DataServers(Cont.)DataCachingDatacanbecachedatclienteveninbetweentransactionsButcheckthatdataisup-to-datebeforeitisused(cachecoherency)CheckcanbedonewhenrequestinglockondataitemLockCachingLockscanberetainedbyclientsystemeveninbetweentransactionsTransactionscanacquirecachedlockslocally,withoutcontactingserverServercallsbacklocksfromclientswhenitreceivesconflictinglockrequest.Clientreturnslockoncenolocaltransactionisusingit.Similartodeescalation,butacrosstransactions. ParallelSystemsParalleldatabasesystemsconsistofmultipleprocessorsandmultipledisksconnectedbyafastinterconnectionnetwork.Acoarse-grainparallelmachineconsistsofasmallnumberofpowerfulprocessorsAmassivelyparallelorfinegrainparallelmachineutilizesthousandsofsmallerprocessors.Twomainperformancemeasures:throughput---thenumberoftasksthatcanbecompletedinagiventimeintervalresponsetime---theamountoftimeittakestocompleteasingletaskfromthetimeitissubmitted Speed-UpandScale-UpSpeedup:afixed-sizedproblemexecutingonasmallsystemisgiventoasystemwhichisN-timeslarger.Measuredby:speedup=smallsystemelapsedtimelargesystemelapsedtimeSpeedupislinearifequationequalsN.Scaleup:increasethesizeofboththeproblemandthesystemN-timeslargersystemusedtoperformN-timeslargerjobMeasuredby:scaleup=smallsystemsmallproblemelapsedtimebigsystembigproblemelapsedtimeScaleupislinearifequationequals1. Speedup Scaleup BatchandTransactionScaleupBatchscaleup:Asinglelargejob;typicalofmostdecisionsupportqueriesandscientificsimulation.UseanN-timeslargercomputeronN-timeslargerproblem.Transactionscaleup:Numeroussmallqueriessubmittedbyindependentuserstoashareddatabase;typicaltransactionprocessingandtimesharingsystems.N-timesasmanyuserssubmittingrequests(hence,N-timesasmanyrequests)toanN-timeslargerdatabase,onanN-timeslargercomputer.Well-suitedtoparallelexecution. FactorsLimitingSpeedupandScaleupSpeedupandscaleupareoftensublineardueto:Startupcosts:Costofstartingupmultipleprocessesmaydominatecomputationtime,ifthedegreeofparallelismishigh.Interference:Processesaccessingsharedresources(e.g.,systembus,disks,orlocks)competewitheachother,thusspendingtimewaitingonotherprocesses,ratherthanperformingusefulwork.Skew:Increasingthedegreeofparallelismincreasesthevarianceinservicetimesofparallelyexecutingtasks.Overallexecutiontimedeterminedbyslowestofparallelyexecutingtasks. InterconnectionNetworkArchitecturesBus.Systemcomponentssenddataonandreceivedatafromasinglecommunicationbus;Doesnotscalewellwithincreasingparallelism.Mesh.Componentsarearrangedasnodesinagrid,andeachcomponentisconnectedtoalladjacentcomponentsCommunicationlinksgrowwithgrowingnumberofcomponents,andsoscalesbetter.Butmayrequire2nhopstosendmessagetoanode(ornwithwraparoundconnectionsatedgeofgrid).Hypercube.Componentsarenumberedinbinary;componentsareconnectedtooneanotheriftheirbinaryrepresentationsdifferinexactlyonebit.ncomponentsareconnectedtolog(n)othercomponentsandcanreacheachotherviaatmostlog(n)links;reducescommunicationdelays. InterconnectionArchitectures ParallelDatabaseArchitecturesSharedmemory--processorsshareacommonmemoryShareddisk--processorsshareacommondiskSharednothing--processorsshareneitheracommonmemorynorcommondiskHierarchical--hybridoftheabovearchitectures ParallelDatabaseArchitectures SharedMemoryProcessorsanddiskshaveaccesstoacommonmemory,typicallyviaabusorthroughaninterconnectionnetwork.Extremelyefficientcommunicationbetweenprocessors—datainsharedmemorycanbeaccessedbyanyprocessorwithouthavingtomoveitusingsoftware.Downside–architectureisnotscalablebeyond32or64processorssincethebusortheinterconnectionnetworkbecomesabottleneckWidelyusedforlowerdegreesofparallelism(4to8). SharedDiskAllprocessorscandirectlyaccessalldisksviaaninterconnectionnetwork,buttheprocessorshaveprivatememories.ThememorybusisnotabottleneckArchitectureprovidesadegreeoffault-tolerance—ifaprocessorfails,theotherprocessorscantakeoveritstaskssincethedatabaseisresidentondisksthatareaccessiblefromallprocessors.Examples:IBMSysplexandDECclusters(nowpartofCompaq)runningRdb(nowOracleRdb)wereearlycommercialusersDownside:bottlenecknowoccursatinterconnectiontothedisksubsystem.Shared-disksystemscanscaletoasomewhatlargernumberofprocessors,butcommunicationbetweenprocessorsisslower. SharedNothingNodeconsistsofaprocessor,memory,andoneormoredisks.Processorsatonenodecommunicatewithanotherprocessoratanothernodeusinganinterconnectionnetwork.Anodefunctionsastheserverforthedataonthediskordisksthenodeowns.Examples:Teradata,Tandem,Oracle-nCUBEDataaccessedfromlocaldisks(andlocalmemoryaccesses)donotpassthroughinterconnectionnetwork,therebyminimizingtheinterferenceofresourcesharing.Shared-nothingmultiprocessorscanbescaleduptothousandsofprocessorswithoutinterference.Maindrawback:costofcommunicationandnon-localdiskaccess;sendingdatainvolvessoftwareinteractionatbothends. HierarchicalCombinescharacteristicsofshared-memory,shared-disk,andshared-nothingarchitectures.Toplevelisashared-nothingarchitecture–nodesconnectedbyaninterconnectionnetwork,anddonotsharedisksormemorywitheachother.Eachnodeofthesystemcouldbeashared-memorysystemwithafewprocessors.Alternatively,eachnodecouldbeashared-disksystem,andeachofthesystemssharingasetofdiskscouldbeashared-memorysystem.Reducethecomplexityofprogrammingsuchsystemsbydistributedvirtual-memoryarchitecturesAlsocallednon-uniformmemoryarchitecture(NUMA) DistributedSystemsDataspreadovermultiplemachines(alsoreferredtoassitesornodes).NetworkinterconnectsthemachinesDatasharedbyusersonmultiplemachines DistributedDatabasesHomogeneousdistributeddatabasesSamesoftware/schemaonallsites,datamaybepartitionedamongsitesGoal:provideaviewofasingledatabase,hidingdetailsofdistributionHeterogeneousdistributeddatabasesDifferentsoftware/schemaondifferentsitesGoal:integrateexistingdatabasestoprovideusefulfunctionalityDifferentiatebetweenlocalandglobaltransactionsAlocaltransactionaccessesdatainthesinglesiteatwhichthetransactionwasinitiated.Aglobaltransactioneitheraccessesdatainasitedifferentfromtheoneatwhichthetransactionwasinitiatedoraccessesdatainseveraldifferentsites. Trade-offsinDistributedSystemsSharingdata–usersatonesiteabletoaccessthedataresidingatsomeothersites.Autonomy–eachsiteisabletoretainadegreeofcontroloverdatastoredlocally.Highersystemavailabilitythroughredundancy—datacanbereplicatedatremotesites,andsystemcanfunctionevenifasitefails.Disadvantage:addedcomplexityrequiredtoensurepropercoordinationamongsites.Softwaredevelopmentcost.Greaterpotentialforbugs.Increasedprocessingoverhead. ImplementationIssuesforDistributedDatabasesAtomicityneededevenfortransactionsthatupdatedataatmultiplesitesThetwo-phasecommitprotocol(2PC)isusedtoensureatomicityBasicidea:eachsiteexecutestransactionuntiljustbeforecommit,andtheleavesfinaldecisiontoacoordinatorEachsitemustfollowdecisionofcoordinator,evenifthereisafailurewhilewaitingforcoordinatorsdecision2PCisnotalwaysappropriate:othertransactionmodelsbasedonpersistentmessaging,andworkflows,arealsousedDistributedconcurrencycontrol(anddeadlockdetection)requiredDataitemsmaybereplicatedtoimprovedataavailabilityDetailsofaboveinChapter22 NetworkTypesLocal-areanetworks(LANs)–composedofprocessorsthataredistributedoversmallgeographicalareas,suchasasinglebuildingorafewadjacentbuildings.Wide-areanetworks(WANs)–composedofprocessorsdistributedoveralargegeographicalarea. Local-areaNetwork NetworksTypes(Cont.)WANswithcontinuousconnection(e.g.,theInternet)areneededforimplementingdistributeddatabasesystemsGroupwareapplicationssuchasLotusnotescanworkonWANswithdiscontinuousconnection:Dataisreplicated.Updatesarepropagatedtoreplicasperiodically.Copiesofdatamaybeupdatedindependently.Non-serializableexecutionscanthusresult.Resolutionisapplicationdependent. EndofChapter17 Figure17.01 Figure17.02 Figure17.03 Figure17.04 Figure17.05 Figure17.06 Figure17.07 Figure17.08 Figure17.09 Figure17.10 Figure17.11'