Computing Prof Dr Andreas Polze HassoPlattnerInstitute for Software Engineering at University Potsdam CloudFutures Workshop Redmond April 89 2010 Agenda Towards Predictable Cloud Computing ID: 1046851
Download Presentation The PPT/PDF document "Towards Predictable Cloud" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
1. Towards Predictable CloudComputingProf. Dr. Andreas PolzeHasso-Plattner-Institute for Software Engineeringat University PotsdamCloudFutures Workshop, Redmond, April 8-9, 2010
2. AgendaTowards Predictable Cloud ComputingHasso Plattner Institute and Operating Systems and Middleware GroupWRK, KStruct and NTraceResource Management for Service ComputingCloud Computing status quoResource Management for Cloud Services
3. Hasso-Plattner-Institute for Software Engineering (HPI)privately funded and independent research institute, founded in 1999associated with the University of Potsdam, GermanyB.Sc. and M.Sc. curriculum in IT-Systems EngineeringPh.D. programmerich experience in research projects that are typically conducted with industrial partners, both on a national and international level
4.
5. MiddlewareService ComputingEmbedded SystemsOperating SystemsWindows Research KernelHP/UX, OpenVMS, SolarisWorkload ManagementAdaptive Services GridSOA Runtime GovernanceDistributed Control LabIntelligent BuildingRowing ComputerFactory AutomationAnalytic Redundancy andDynamic System UpdateRealtime .NET andECMA 335 Realtime OSScheduling ServerGrid OccamBeckhoff, ABB, Siemens PTDSoftware AG, DaimlerChrysler ResearchMicrosoft Research,MS Windows Group,HP Bristol LabsVET-Trend Project,tele-lab, eLearningFontane ProjecteHealth, remote patient monitoringResearch Agenda
6. AgendaTowards Predictable Cloud ComputingHasso Plattner Institute and Operating Systems and Middleware GroupWRK, KStruct and NTraceResource Management for Service ComputingCloud Computing status quoResource Management for Cloud Services
7.
8. Windows Research Kernel (WRK)
9. KStructKernel Runtime InspectionOS kernels are complex software systemsReading source code is not abstract enoughBut carrying out „experiments“ on the system might helpInspecting the state of the system at runtime might helpDebuggers are of only limited use Focus on single steppingTracing either execution flow or data structureHeavy side effects on system behavior (timing)Separate talk on SkyLab by Alexander Schmidt
10. Hotpatchable Images – applied for kernel-mode tracingNew compiler/linker switch/hotpatch, /functionpadmincl Svr03 SP1 and laterWindows Server 2003 SP1,Vista, Server 2008, Windows 7Retail-Kernel ist hotpatchableWRK can be recompiled to be hotpatchable10.04.2010Johannes Passing10 ... retn 10 nop nop nop nop nopNtfsPinMappedData: mov edi, edi push ebp mov ebp, esp mov ecx, [ebp+18h] mov edx, [ebp+0Ch] ...
11. NTrace operation10.04.2010Johannes PassingFoo-5:CallProxy:......EntryThunk:Foo:...
12. ChallengesInstrumentation vs. self-modifying codeModification has to be performed in a safe manner: atomic operationBehavior of the modified routine has to be equivalent to its non-instrumented counterpartWhen is it safe to unload?Retail Kernel, DriverOnly public (external) symbols availableAuxiliary StackStores original return addresses and bookkeeping infoThread-localRe-Entrancytracing system must avoid calling routines which have themselves been instrumented avoid hazardous forms of re-entrance and endless recursionException handlingInstrumentation has to integrate with Windows Structured Exception Handling (SEH)Exception un-winds need to be tracedSEH implementation requires exception frames to be on main stack inserting new frames will change stack layout and render functions‘ code invalid...12
13. AgendaTowards Predictable Cloud ComputingHasso Plattner Institute and Operating Systems and Middleware GroupWRK, KStruct and NTraceResource Management for Service ComputingCloud Computing status quoResource Management for Cloud Services
14. Rule-based resource managementService ConsumerService RegistryService ProviderPolicy Enforcement PointPolicyCentraSiteXBrokerService ProviderService Provider Failover Routing LoggingPolicy
15. Software AG’s CentraSite –a UDDI repository & brokerNew CentraSite Policy Enforcement Point (PEP) implementationIntercepts SOAP messagingInstalled on service provider sideOffers real-time resource partitioningWithout changing operating systemWithout changing application serverWithout changing Java run-timeSeamless integration with policy management in CentraSite
16. HPI Policy Enforcement Point for CentraSiteService ConsumerApplication ServerOperating SystemUDDIv3PolicyExchangeHPI PEP ServletService RegistryCentraSiteHPI Scheduling ServerJAX-WS HandlerWeb Service JAX-WS HandlerWeb Service JAX-WS HandlerWeb Service Service PoliciesXBroker PEPPerformance DataService – Thread MappingPerformanceReportService Provider
17. Resource management:Scheduling Server
18. Scheduling ServerCurrent prototype for Windows (>= 2000)Standard Windows daemon with admin rightsExisting versions also for Linux, Solaris, MacOSJava part remains the sameProvides explicit CPU share to thread processing the service requestNon-used cycles are donated to the systemService overrules all other Java / Non-Java threads in case of system overloadCapable of multiple reservations per machineCapable of whole process control
19. 19Service-Oriented ComputingDue to virtualization and server consolidation, services will eventually end up on same machine
20. Common service monitoring modelRequest packageenters platform(source: WSQM)Service reachable, but broken(source: Laprie)Time for EJB / Handler processing(source: JSR-77)Finished requests / uptime(source: WSQM)Service not reachable(source: WSLA)ServiceResource
21. AgendaTowards Predictable Cloud ComputingHasso Plattner Institute and Operating Systems and Middleware GroupWRK, KStruct and NTraceResource Management for Service ComputingCloud Computing status quoResource Management for Cloud Services
22. Cloud Computing...is a computing paradigm where the boundaries of computing will be determined by economic rationale rather than technical limitsThe concept generally incorporates combinations of:infrastructure as a service (IaaS)platform as a service (PaaS)software as a service (SaaS)
23. Cloud Computing – the three layers ServersStorageRacksHVACPowerCloud DataStoreManaged ContainerComm-unicationsVirtual ComputeVirtual MachineVirtual StorageKey-value StoreBlock StoreBusinessApplicationsAnalyticsApplicationsProductivityApplicationsInfrastructure“Infrastructure as a Service” , “Utility Computing”Platforms“Platform as a Service”Applications“Software as a Service”, “on-demand” appsChallenges to the CloudHas to abstract underlying hardwareBe elastic in scaling to demandBuild on a pay per use basisCloud Stack
24. Cloud Computing – status quoPrivate vs. Public cloud – or both?Private: IBM, SAPPublic: Amazon, Google, MicrosoftBoth: VMware, Salesforce‘ force.comUnit of granularityAmazon Machine Image VMware Virtual Appliance .... Virtual DatacenterWeb Role, Server Role (MS) Physical Machine (IBMs RC2)
25. Cloud Computing – status quo (contd.)Programming modelsWebRole, WebService in .NETNative C++ programJ2EE ( EJB )vCloud APIInteraction and CommunicationService busMessage passing APIsVirtual layer 2 network connectivitySecurity / Trust
26.
27. Servers have evolved...New form factorsHigher densityStandard architectures (x64, Itanium)Multicore/multithreaded archAdvances in operating systemsVirtualizationThrustworthiness/securityClustering Need for new programming models, Software Architectures, Services
28. Immense energy consumption per m2 ...On the agenda:Water CoolingDC power suppliesresource managementmainly on a manual basis
29. AgendaTowards Predictable Cloud ComputingHasso Plattner Institute and Operating Systems and Middleware GroupWRK, KStruct and NTraceResource Management for Service ComputingCloud Computing status quoResource Management for Cloud Services
30. Adopting existing solutions for the CloudScheduling server for CPU partitioningNTrace for local resource monitoringDetecting potential hotspots on local fabricGenerating monitoring interfaces automaticallyDeploying monitoring interface with cloud serviceInterfacing with VM monitoring infrastructure (i.e.; vProbe)Accomodate new programming modelsParallel computing, synchronizationCo-management of parallel activities required
31. NTrace for monitoringIn the local fabricPotentially also in the cloudFoo-5:CallProxy:......EntryThunk:Foo:...
32. Scheduling Server for CPU partitioning
33. Common service monitoring modelRequest packageenters platform(source: WSQM)Service reachable, but broken(source: Laprie)Time for EJB / Handler processing(source: JSR-77)Finished requests / uptime(source: WSQM)Service not reachable(source: WSLA)ServiceResource
34. Programming Models #1: OpenCL, CUDAOpenCL – Open Computing LanguageCUDA – Compute Unified Device Architecture Open standard for portable, parallel programming of heterogeneous parallel computing CPUs, GPUs, and other processors
35. OpenCL Design GoalsUse all computational resources in system Program GPUs, CPUs, and other processors as peers Support both data- and task- parallel compute models Efficient C-based parallel programming model Abstract the specifics of underlying hardware Abstraction is low-level, high-performance but device-portable Approachable – but primarily targeted at expert developersEcosystem foundation – no middleware or “convenience” functions Implementable on a range of embedded, desktop, and server systems HPC, desktop, and handheld profiles in one specification Drive future hardware requirements Floating point precision requirements Applicable to both consumer and HPC applications
36. OpenCL Platform ModelOne Host + one or more Compute Devices Each Compute Device is composed of one or more Compute Units Each Compute Unit is further divided into one or more Processing Elements
37. Programming Models #2: Intel’s Ct technologyCt adds new data types (parallel vectors) & operators to C++ Library-like interface and is fully ANSI/ISO-compliant Ct abstracts away architectural details Vector ISA width / Core count / Memory model / Cache sizes Fully leverage deterministic parallel programming models I.e. Make data races impossible Expresses complex behaviors through simple operators Presents a simple and predictable performance model Provides a forward-scaling programming model Is a project being worked on at Intel since some time already
38. Ct: Nested Data Parallelism in C/C++
39. Future SOC Lab @ HPIVision is to establish an open research platform for tomorrow’s IT landscape Industry partnersFujitsuHewlett-PackardSAPDeutsche TelekomEMC VMware Testbed:MultiCore MultiThreading Hardware,huge memories, NehalemEX-basedHP ProLiant DL980 G6: 64 Cores, 1-2TBFujitsu Primergy RX600S: 24 Cores, 1TB
40. Call for ProjectsOpen for research communityInitial deadline: 31st March 2010Research TopicsMulticore architecturesIn-memory business applicationsService-Oriented Computing and Service-Oriented ArchitecturesCloud Computing, SaaS, PaaSNew security conceptsDevelopment techniques for Multicore, Cloud Computing and SaaSNew Real-time Business Process ManagementLarge-Scale Multimedia Information Retrieval and Data Mining
41. Proposal Origins
42. State of Future SOC Lab30 Proposals as of April 1, 2010Hasso-Plattner-InstitutETH Zurich, SwitzerlandUniversity Bayreuth, GermanyBlekinge Institute of Technology, SwedenUniversity of MagdeburgTechnical University Berlin, GermanyGerman Research Center for Artificial Intelligence, Germany SAP AG, GermanyFrauenhofer FOKUS, GermanyUniversity WürzburgUniversity of California at Berkeley, USA
43. ConclusionsNew programming models on the horizon – think cloudHybrid Clouds will be the futureThe MultiThreading MultiCore ChallengeMust learn to write parallel programs (OpenMP, MPI, etc.)Parallel computing on a chipMonitoring the cloudWith respect to resource usage, security, SLAMust include the virtualization layerGreen Cloud
44.