First Product Shipped In 1991 F Acquired By Stratus .

Transcription

Corporate ProfileIsis Distributed SystemsFFFFFDr. Kenneth P. BirmanFounder and Chief ScientistFUS Govt Funded Research ProjectFirst product shipped in 1991Acquired by Stratus Computer in December, 1993Currently 100 employees, 150 customersIndustries include Financial Services,Telecommunications, Manufacturing andGovernment/ScientificProducts run on most major UNIX systems as well asWindows NT, Windows 3.1, VMS, VOS & FTXNovember 21th 1995Forbis, London November 21th 1995Forbis, London November 21th 1995Reliable Distributed ComputingFRequired in critical applications: finance,telecommunications, air traffic control, powersystems management.FTraditionally addressed late in the game: firstbuild the system, then harden it.FMust be compatible with standard tools forbuilding distributed applications.What makes it hard?FNeed to replicate critical data and servers, butlack tools for doing this easily.FNeed a way to detect failures, but failurereporting is often erratic, inconsistent.FNeed a way to restart failed system componentsso they can rejoin the system, but withoutdisrupting service.Forbis, London November 21th 1995Forbis, London November 21th 1995Distributed Trading SystemDistributed Trading System1Pricing DB’sTrader ClientsHistorical DataMarketDataFeedsTrader ClientsPricing DB’sHistorical DataMarketDataFeedsCurrent PricingAnalyticsCurrent Pricing Availability for historical dataAnalyticsTokyo, London, Zurich, .Long-Haul WAN SpoolerForbis, London November 21th 1995Tokyo, London, Zurich, .Long-Haul WAN SpoolerForbis, London November 21th 1995

Distributed Trading SystemPricing DB’s1Trader ClientsDistributed Trading SystemPricing DB’s1Historical DataMarketDataFeedsHistorical DataTrader ClientsMarketDataFeeds22Current Pricing Availability for historical data- Load balancing and consistent Availability for historical data- Load balancing and consistenAnalyticsTokyo, London, Zurich, .message delivery for pricedistributionForbis, London November 21th 1995Trader ClientsTokyo, London, Zurich, .Long-Haul WAN SpoolerForbis, London November 21th 1995Distributed Trading SystemInconsistent failure reportingA simple Client serverPricing DB’s1Analyticsmessage delivery for pricedistribution Parallel execution for analyticsLong-Haul WAN SpoolerCurrent Pricing3Historical t Pricing3 Availability for historical data- Load balancing and consistentAnalyticsBackupmessage delivery for pricedistribution Parallel execution for analyticsTokyo, London, Zurich, .Long-Haul WAN SpoolerForbis, London November 21th 1995Inconsistent failure reportingClient and primary consider one-another faulty because areliable connection broke yet neither actually crashedClientApplicationPrimaryForbis, London November 21th 1995Inconsistent failure reportingThe problem gets even worse with more ckupBackupClientApplicationForbis, London November 21th 1995Forbis, London November 21th 1995

Inconsistent failure reportingInconsistent event reportingWho considers whom broken and icereplicated foravailabilityClientApplicationForbis, London November 21th 1995Forbis, London November 21th 1995Inconsistency causes problemsFReplicated data out of syncFLocks not acquired or released, or deadlockFUnable to balance load or subdivide workFCan't implement primary/backup fault-tolerancealgorithmsFCan’t automate system management. limiting application reliability!FInconsistency is built into communicationprimitives in Unix, VMS, NT, Windows 95FApplications built on such systems must livewith inconsistency: they are unreliable and hardto manageThe key to Isis is a consistency mechanism:Virtually Synchronous Process GroupsForbis, London November 21th 1995Forbis, London November 21th 1995Virtual Synchrony in Isisc0m0s0c1s1Consistency in Isiss2FBuilt over UDP using proprietary protocolsFConsistent failure reportingFGroup members see events in same orderFState transfer provided to joining memberm1m3m2m4s3CRASHjoinsstate transferForbis, London November 21th 1995Forbis, London November 21th 1995

Isis Software Developer's ToolkitIsis SDK turns model into "tools"HOST AFAPI for group join, multicast, state transferFReplicated data and synchronizationFCoordinated, load-balanced request executionFFault-tolerance through softwareFSelf-management interfacesFHardware multicast for big fanouts, point topoint for small groupsHOST BSERVERPROCESSForbis, London November 21th 1995Forbis, London November 21th 1995Isis Software Developer's ToolkitHOST AIsis Software Developer's ToolkitHOST BSERVERPROCESSHOST AProcess GrouptechnologySERVERSERVERPROCESSPROCESSHOST BSERVERPROCESSCLIENT 1FServer processes join a group to form a serviceFFForbis, London November 21th 1995SERVERSERVERPROCESSPROCESSCLIENT 2.CLIENT nServer processes join a group to form a serviceClient requests are delivered to the group with Isisreliable group messagingIsis Product SetHOST BSERVERPROCESSSERVERSERVERPROCESSPROCESSProcess GrouptechnologyApplicationsEnd-users, VARs, ToolsCLIENT 1FFFFCLIENT 2.CLIENT nServer processes join a group to form a serviceClient requests are delivered to the group with Isisreliable group messagingExecution can be replicated, load-balancedIsis message and failure consistency ensures correctnessof coordinated actionsForbis, London November 21th 1995Process GrouptechnologyForbis, London November 21th 1995Isis Software Developer's ToolkitHOST ASERVERSERVERPROCESSPROCESSIsis Core TechnologySDK, Message Distribution ServiceForbis, London November 21th 1995

Isis Product SetIsis Product SetApplicationsApplicationsEnd-users, VARs, ToolsEnd-users, VARs, NFSNFSOrbixOrbix IsisIsisIsis Core TechnologyIsis Core TechnologySDK, Message Distribution ServiceSDK, Message Distribution ServiceForbis, London November 21th 1995Forbis, London November 21th 1995Isis Product SetIsis Product SetApplicationsApplicationsEnd-users, VARs, ToolsEnd-users, VARs, sisIsisIsis NFSNFScorbanfsOrbixOrbixforforforfor is Core TechnologyIsis Core TechnologySDK, Message Distribution ServiceSDK, Message Distribution ServiceForbis, London November 21th 1995Forbis, London November 21th 1995Isis Product SetApplicationsEnd-users, VARs, ToolscorbanfsIsisIsisforforNFSNFSodbcOrbixOrbix s Core TechnologySDK, Message Distribution ServiceForbis, London November 21th 1995RADIO: Putting It All TogetherThree dual 133 MHzPentium computenodesThree 8 gigabytestorage nodesTwo network nodes- 100 mbit Enet hubor- 100 mbit Enet switchForbis, London November 21th 1995

Radio and Isis TechnologyFFFFHighest availability PC cluster server hosting Isisavailability technology and products.Isis for y serviced, hot-pluggable, self-managed.Many levels of availability: shrink-wrap NT orUnixware application benefit, if desired use Isisactive replication for continuous availability.Seamlessly integrated with Isis on your networkand Isis applications on your workstations.databases arereplicated for loadbalancing; availabilityForbis, London November 21th 1995Isis for DBIsis for DBOracle,Sybase,etc.Oracle,Sybase,etc.Process GroupForbis, London November 21th 1995Isis Message Distribution ServiceOrbix Isis - The Big PictureFFFFClientClient usesuses singlesingle logicallogical interfaceinterface (CORBA(CORBA IDL)IDL)ThatThat interfaceinterface isis implementedimplemented byby multiplemultiple replicatedreplicated objectsobjects(“object(“object group”)group”)FF MappingMapping ofof singlesingle interfaceinterface toto objectobject groupgroup isis providedprovided bybypackagedpackaged “templates”“templates”PublisherMDS ServerPublisherSubscriberMDS templates”Forbis, London November 21th 1995FPublishers send messages by subjectFSubscribers register for subjects of interestFIsis MDS provides guaranteed message deliveryFMDS maintains persistent message streams forsubscribers that leave and returnForbis, London November 21th 1995Isis Reliable Network File SystemSERVER ASubscriberIsis Availability ManagerSERVER BRNFSSERVERRNFSSERVERClientFFault-tolerant infrastructure for supervisingapplications and critical serversFSimple graphical interface is easy to useFFully customizable with rule-based scriptsFSensor/action model instruments your applicationin natural ways that you defineFAutomatically restarts critical serversClientFReplicates NFS files across multiple serversFRemoves single points of failureFAutomatically recovers from server failuresFApplication program, NFS server unchangedForbis, London November 21th 1995Forbis, London November 21th 1995

Profile of Customer ApplicationsIsis Availability ManagerFProcess Group of AgentsPrimaryAgent(node a)EventNotificationsPrimaryAgent(node b)SecondaryAgentAction Request(node c)FNodeInformation SecondaryAgent(node d)FManagementRequestsGUI(node f)GUI(node g)SecondaryAgent(node e)FProcess Group of GUI’sFinancial Services– NYSE, Swiss Stock Exchange, World Bank, Lehman Bros.– Market data distribution, replicated databases, wide areacommunications, bulletin boards, management of analyticsTelecommunications– Northern Telecom, Southwestern Bell– Intelligent switching systems and cellular servicesFactory Automation– Advanced Micro Devices, Texas Instruments, SEMAtech– Semiconductor factory automation, distributed logicsimulationGovernment/Scientific– Martin Marietta, Los Alamos Labs, Sandia Labs– Command and control systems, signal processingForbis, London November 21th 1995Forbis, London November 21th 1995Distributed Service NodeDistributed Trading SystemOne phone number per personPricing DB’s1.Historical DataMarketDataFeedsTrader ClientsTelephony Data/Digitized Voice Pathx86/UNIXRISC/UNIXTelephoneTrunk LinesRISC/UNIXDumbSwitch2.AnalyticsCurrent ity for historical dataLoad balancing and consistentmessage delivery for price distributionParallel execution for analyticsLong-Haul WAN SpoolerTokyo, London, Zurich, .Forbis, London November 21th 1995Forbis, London November 21th 1995Distributed Service NodeDistributed Service NodeOne phone number per personOne phone number per personTelephony Data/Digitized Voice Pathx86/UNIXRISC/UNIXTelephoneTrunk LinesRISC/UNIXDumbSwitch1.Ethernet Isis Replicated files for digitized voice storeForbis, London November 21th 1995Telephony Data/Digitized Voice elephoneTrunk LinesRISC/UNIXDumbSwitch2.Ethernet Isis Replicated files for digitized voice store- Redundancy for database availabilityForbis, London November 21th 1995Calls,Changes,Adds,Deletes

Distributed Service NodeOne phone number per personTelephony Data/Digitized Voice Pathx86/UNIX1.RISC/UNIXTelephoneTrunk LinesRISC/UNIX2.DumbSwitch3.Ethernet Isis Replicated files for digitized voice store- Redundancy for database availability Load balancing for call handling & routingShop Floor Process Control ––––––SUN OSSOLARISHP/UXIBM AIXWINDOWS NTWIN 3.1ALPHA OSF/1FTXVMSSGIPYRAMIDSEQUENTForbis, London November 21th trollerEnetHPStationControllerFactory equipmentForbis, London November 21th 1995Supported PlatformsOtherHPHPForbis, London November 21th rIsis Market PositioningFAddresses application-level problems ofdistributed computing:Reliability Scaling ConsistencyFActive replication versus restart-recoveryFApplication availability versus server availabilityFVendor independent and portableFCommitted to integration and standardsForbis, London November 21th 1995

Manager Isis Core Technology SDK, Message Distribution Service nfs corba odbc snmp Forbis, London November 21th 1995 RADIO: Putting It All Together . Isis for Database Client Client Client Client Oracle, Sybase, etc. Oracle, Sybase, etc. Process Group Isis for DB Isis