the microsoft perspective on where high performance computing is heading kyril faenov director of...
Post on 21-Dec-2015
215 views
TRANSCRIPT
The Microsoft Perspective On Where High Performance Computing Is Heading
The Microsoft Perspective On Where High Performance Computing Is Heading
Kyril FaenovDirector of HPCWindows Server DivisionMicrosoft Corporation
Kyril FaenovDirector of HPCWindows Server DivisionMicrosoft Corporation
Talk OutlineTalk Outline
Market/technology trends
Personal supercomputing
Grid computing
Leveraging IT industry investments
Decoupling domain science from Computer Science
Market/technology trends
Personal supercomputing
Grid computing
Leveraging IT industry investments
Decoupling domain science from Computer Science
Top 500 Supercomputer TrendsTop 500 Supercomputer Trends
Industry usage rising
Clusters over 50%
x86 is winning
GigE is gaining
HPC Market TrendsHPC Market Trends
Capability, Enterprise
$1M+
Top Challenges to Implementing Clusters(IDC 2004, N=229)
System management capability 18%
Apps availability 17%
Parallel algorithm complexity 14%
Space, power, cooling 11%
Interconnect BW/latency 10%
I/O performance 9%
Interconnect complexity 9%
Other 12%
Report of the High-End Computing Revitalization Task Force, 2004(Office of Science and Technology Policy, Executive Office of the President)
“Make high-end computing easier and more productive to use. Emphasis should be placed on time to solution, the major metric of value to high-end computing users… A common software environment for scientific computation encompassing desktop to high-end systems will enhance productivity gains by promoting ease of use and manageability of systems.”
Divisional$250K-$1M
Departmental$50-250K
Workgroup<$50K
4.2%4.2%
2004 Systems2004 Systems
5.7%5.7%
1,1671,167
3,9153,915
22,71222,712
127,802127,802
Source: IDC, 2005Source: IDC, 2005
2004-9 CAGR2004-9 CAGR
7.7%7.7%
13.4%13.4%
<$250K – 97% of systems, 52% of revenue<$250K – 97% of systems, 52% of revenue
In 2004 clusters grew 96% to 37% by revenueIn 2004 clusters grew 96% to 37% by revenueAverage cluster size 10-16 nodesAverage cluster size 10-16 nodes
Major ImplicationsMajor Implications
Market pressures demand accelerated innovation cycle, overall cost reduction, and thorough outcome modelingLeverage volume markets of industry standard hardware and softwareRapid procurement, installation and integration of systemsWorkstation-Cluster integrated applications accelerating market growth
EngineeringBioinformaticsOil and GasFinanceEntertainmentGovernment/Research
Market pressures demand accelerated innovation cycle, overall cost reduction, and thorough outcome modelingLeverage volume markets of industry standard hardware and softwareRapid procurement, installation and integration of systemsWorkstation-Cluster integrated applications accelerating market growth
EngineeringBioinformaticsOil and GasFinanceEntertainmentGovernment/Research
The convergence of affordable high performance hardware and The convergence of affordable high performance hardware and commercial apps is making supercomputing personal commercial apps is making supercomputing personal
Supercomputing Goes PersonalSupercomputing Goes Personal
19911991 19981998 20052005SystemSystem Cray Y-MP C916Cray Y-MP C916 Sun HPC10000Sun HPC10000 Shuttle @ NewEgg.comShuttle @ NewEgg.com
ArchitectureArchitecture 16 x Vector16 x Vector4GB, Bus4GB, Bus
24 x 333MHz Ultra-24 x 333MHz Ultra-SPARCII, 24GB, SBusSPARCII, 24GB, SBus
4 x 2.2GHz x644 x 2.2GHz x644GB, GigE4GB, GigE
OSOS UNICOSUNICOS Solaris 2.5.1Solaris 2.5.1 Windows Server 2003 SP1Windows Server 2003 SP1
GFlopsGFlops ~10~10 ~10~10 ~10~10
Top500 #Top500 # 11 500500 N/AN/A
PricePrice $40,000,000$40,000,000 $1,000,000 (40x drop)$1,000,000 (40x drop) < $4,000 (250x drop)< $4,000 (250x drop)
CustomersCustomers Government LabsGovernment Labs Large EnterprisesLarge Enterprises Every Engineer & Scientist Every Engineer & Scientist
ApplicationsApplications Classified, Climate, Classified, Climate, Physics ResearchPhysics Research
Manufacturing, Energy, Manufacturing, Energy, Finance, TelecomFinance, Telecom
Bioinformatics, Materials Bioinformatics, Materials Sciences, Digital MediaSciences, Digital Media
The FutureSupercomputing on a ChipThe FutureSupercomputing on a Chip
IBM Cell processor256 Gflops today4 node personal cluster => 1 Tflops32 node personal cluster => Top100
Microsoft Xbox3 custom PowerPCs + ATI graphics processor1 Tflops today$3008 node personal cluster => “Top100” for $2500 (ignoring all that you don’t get for $300)
Intel many-core chips“100’s of cores on a chip in 2015” (Justin Rattner, Intel)“4 cores”/Tflop => 25 Tflops/chip
IBM Cell processor256 Gflops today4 node personal cluster => 1 Tflops32 node personal cluster => Top100
Microsoft Xbox3 custom PowerPCs + ATI graphics processor1 Tflops today$3008 node personal cluster => “Top100” for $2500 (ignoring all that you don’t get for $300)
Intel many-core chips“100’s of cores on a chip in 2015” (Justin Rattner, Intel)“4 cores”/Tflop => 25 Tflops/chip
Key To EvolutionTackling system complexityKey To EvolutionTackling system complexity
ScenarioScenario FocusFocus
Departmental ClusterDepartmental ClusterConventional scenarioConventional scenario
IT owns large clusters due to cost and IT owns large clusters due to cost and complexity and allocates resources on complexity and allocates resources on per job basisper job basis
Users submit batch jobs via scriptsUsers submit batch jobs via scripts
In-house and ISV apps, many based on In-house and ISV apps, many based on MPIMPI
Scheduling multiple users’ Scheduling multiple users’ applications onto scarce applications onto scarce compute cyclescompute cycles
Cluster systems administration
Personal/Workgroup Personal/Workgroup ClusterClusterEmerging scenarioEmerging scenario
Clusters are pre-packaged OEM Clusters are pre-packaged OEM appliances, purchased and managed by appliances, purchased and managed by end-usersend-users
Desktop HPC applications transparently Desktop HPC applications transparently and interactively make use of cluster and interactively make use of cluster resourcesresources
Desktop development tools integrationDesktop development tools integration
Interactive applicationsInteractive applications
Workstation clusters, accelerator appliances
Distributed, policy-based management and security
HPC Application IntegrationHPC Application IntegrationFuture scenarioFuture scenario
Multiple simulations and data sources Multiple simulations and data sources integrated into a seamless application integrated into a seamless application workflowworkflow
Network topology and latency Network topology and latency awareness for optimal distribution of awareness for optimal distribution of computationcomputation
Structured data storage with rich meta-Structured data storage with rich meta-datadata
Applications and data potentially span Applications and data potentially span organizational boundariesorganizational boundaries
Data-centric, “whole-Data-centric, “whole-system” workflowssystem” workflows
Rapid prototyping of HPC Rapid prototyping of HPC applicationsapplications
Grids: Distributed application, systems, and data management
Interoperability
Interactive Computation Interactive Computation and Visualizationand Visualization
Manual, batchManual, batchexecutionexecution
IT MgrIT Mgr
SQL
“Grid Computing” A catch-all marketing term“Grid Computing” A catch-all marketing term
“Grid” Computing means many different things to many different people/companies
Desktop cycle-stealing
Managed HPC clusters
Internet access to giant, distributed repositories
Virtualization of data center IT resources
Out-sourcing to “utility data centers”
…
Originally this was all called “Distributed Systems”
“Grid” Computing means many different things to many different people/companies
Desktop cycle-stealing
Managed HPC clusters
Internet access to giant, distributed repositories
Virtualization of data center IT resources
Out-sourcing to “utility data centers”
…
Originally this was all called “Distributed Systems”
HPC Grids And Web ServicesHPC Grids And Web Services
HPC Grid ~ Compute Grid + Data Grid
Compute gridForest of clusters and workstations within an organization
Coordinated scheduling of resources
Data gridDistributed storage facilities within an organization
Coordinated management of data
Web ServicesThe means to achieve interoperable Internet-scale computing, including federation of organizations
Loosely-coupled, service-oriented architecture
HPC Grid ~ Compute Grid + Data Grid
Compute gridForest of clusters and workstations within an organization
Coordinated scheduling of resources
Data gridDistributed storage facilities within an organization
Coordinated management of data
Web ServicesThe means to achieve interoperable Internet-scale computing, including federation of organizations
Loosely-coupled, service-oriented architecture
Computational Grid Economics*Computational Grid Economics*
What $1 will buy you (roughly): Computers cost $1000 (roughly)
1 cpu day (~ 10 Tera-ops) == $1 (roughly, assuming 3 yr use cycle)
10TB network transfer costs == $1 (roughly, assuming 1Gbps interconnect)
Internet bandwidth costs roughly 100 $/mbps/month (not including routers and management)
1GB network transfer costs == $1 (roughly)
Some observationsHPC cluster communication is 10,000x cheaper than WAN communication
Break-even point for instructions computed per byte transferred:Cluster: O(1) instrs/byte => many parallel applications are economical to run on a cluster or across a GigE LAN
WAN: O(10,000) instrs/byte => few parallel applications are economical to run across the Internet
*Computational grid economics material courtesy of Jim Gray
What $1 will buy you (roughly): Computers cost $1000 (roughly)
1 cpu day (~ 10 Tera-ops) == $1 (roughly, assuming 3 yr use cycle)
10TB network transfer costs == $1 (roughly, assuming 1Gbps interconnect)
Internet bandwidth costs roughly 100 $/mbps/month (not including routers and management)
1GB network transfer costs == $1 (roughly)
Some observationsHPC cluster communication is 10,000x cheaper than WAN communication
Break-even point for instructions computed per byte transferred:Cluster: O(1) instrs/byte => many parallel applications are economical to run on a cluster or across a GigE LAN
WAN: O(10,000) instrs/byte => few parallel applications are economical to run across the Internet
*Computational grid economics material courtesy of Jim Gray
Exploding Data SizesExploding Data Sizes
Experimental data: TBs PBs
Modeling dataToday
10’s to 100’s of GB per simulation is the common case
Applications mostly run in isolation
Tomorrow 10’s to 100’s of TBs, all of it to be archived
Whole-system modeling and multi-application workflows
Experimental data: TBs PBs
Modeling dataToday
10’s to 100’s of GB per simulation is the common case
Applications mostly run in isolation
Tomorrow 10’s to 100’s of TBs, all of it to be archived
Whole-system modeling and multi-application workflows
How Do You Move A Terabyte?*How Do You Move A Terabyte?*
14 minutes14 minutes6176172002001,920,0001,920,00096009600OC 192OC 192
2.2 hours2.2 hours10001000GbpsGbps
1 day1 day100100100 Mpbs100 Mpbs
14 hours14 hours97697631631649,00049,000155155OC3OC3
2 days2 days2,0102,01065165128,00028,0004343T3T3
2 months2 months2,4692,4698008001,2001,2001.51.5T1T1
5 months5 months36036011711770700.60.6Home DSLHome DSL
6 years6 years3,0863,0861,0001,00040400.040.04Home phoneHome phone
Time/TBTime/TB$/TB$/TBSentSent$/Mbps$/MbpsRentRent
$/month$/monthSpeedSpeedMbpsMbpsContextContext
24 hours24 hours5050100100FedExFedEx
*Material courtesy of Jim Gray*Material courtesy of Jim Gray
LAN SettingLAN Setting
13 minutes13 minutes100001000010 Gpbs10 Gpbs
Anticipated HPC Grid TopologyAnticipated HPC Grid Topology
Islands of high connectivityIslands of high connectivity
Simulations done on personal and Simulations done on personal and workgroup clustersworkgroup clusters
Data stored in data warehousesData stored in data warehouses
Data analysis best done inside the data Data analysis best done inside the data warehousewarehouse
Wide-area data sharing/replication Wide-area data sharing/replication via FedEx?via FedEx?
Data warehouse Workgroupcluster
Personalcluster
Data Analysis And MiningData Analysis And Mining
Traditional approachKeep data in flat filesWrite C or Perl programs to compute specific analysis queriesProblems with this approach
Imposes significant development timesScientists must reinvent DB indexing and query technologiesHave to copy the data from the file system to the compute cluster for every query
Results from the astronomy communityRelational databases can yield speed-ups of one to two orders of magnitudeSQL + application/domain-specific stored procedures greatly simplify creation of analysis queries
Traditional approachKeep data in flat filesWrite C or Perl programs to compute specific analysis queriesProblems with this approach
Imposes significant development timesScientists must reinvent DB indexing and query technologiesHave to copy the data from the file system to the compute cluster for every query
Results from the astronomy communityRelational databases can yield speed-ups of one to two orders of magnitudeSQL + application/domain-specific stored procedures greatly simplify creation of analysis queries
Is That The End Of The Story?Is That The End Of The Story?
Relational Data warehouse Workgroup
cluster
Personalcluster
Too Much ComplexityToo Much Complexity
Relational Data warehouse Workgroup
cluster
Personalcluster
Distributed systems Distributed systems issues:issues:
SecuritySecurity System managementSystem management Directory servicesDirectory services Storage managementStorage management
Digital experimentation:Digital experimentation: Experiment Experiment
managementmanagement Provenance (data & Provenance (data &
workflows)workflows) Version management Version management
(data & workflows)(data & workflows)
Parallel application developmentParallel application development Chip-level, node-level, cluster-level, Chip-level, node-level, cluster-level,
LAN grid-level, WAN grid-level LAN grid-level, WAN grid-level parallelismparallelism
OpenMP, MPI, HPF, Global Arrays, …OpenMP, MPI, HPF, Global Arrays, … Component architecturesComponent architectures Performance configuration & tuningPerformance configuration & tuning Debugging/profiling/tracing/analysisDebugging/profiling/tracing/analysis
Domain science
2004 NAS supercomputing report: O(35) new computational scientists graduated per year2004 NAS supercomputing report: O(35) new computational scientists graduated per year
(Partial) Solution Leverage IT Industry’s Existing R&D(Partial) Solution Leverage IT Industry’s Existing R&D
Parallel applications developmentHigh-productivity IDEs
Integrated debugging/profiling/tracing/analysis
Code designer wizards
Concurrent programming frameworksPlatform optimizations
Dynamic, profile-guided optimization
New programming abstractions
Distributed systems issuesWeb Services & HPC grids
Security
Interoperability
Scalability
Dynamic Systems ManagementSelf (re)configuration & tuning
Reliability & availability
RDMS + data miningEase-of-use
Advanced indexing & query processing
Advanced data mining algorithms
Parallel applications developmentHigh-productivity IDEs
Integrated debugging/profiling/tracing/analysis
Code designer wizards
Concurrent programming frameworksPlatform optimizations
Dynamic, profile-guided optimization
New programming abstractions
Distributed systems issuesWeb Services & HPC grids
Security
Interoperability
Scalability
Dynamic Systems ManagementSelf (re)configuration & tuning
Reliability & availability
RDMS + data miningEase-of-use
Advanced indexing & query processing
Advanced data mining algorithms
Digital experimentationCollaboration-enhanced Office productivity tools
Structure experiment data and derived results in a manner appropriate for human reading/reasoning (as opposed to optimizing for query processing and/or storage efficiency)
Enable collaboration among colleagues
(Scientific) workflow environmentsAutomated orchestration
Visual scripting
Provenance
Digital experimentationCollaboration-enhanced Office productivity tools
Structure experiment data and derived results in a manner appropriate for human reading/reasoning (as opposed to optimizing for query processing and/or storage efficiency)
Enable collaboration among colleagues
(Scientific) workflow environmentsAutomated orchestration
Visual scripting
Provenance
Separating The Domain Scientist From The Computer ScientistSeparating The Domain Scientist From The Computer Scientist
Computer Computer scientistscientist
Computational Computational scientistscientist
Domain Domain
scientistscientist
Parallel domain application developmentParallel domain application development
Parallel/distributed file systems, relational data warehouses, Parallel/distributed file systems, relational data warehouses, dynamic systems management, Web Services & HPC gridsdynamic systems management, Web Services & HPC grids
(Interactive) scientific workflow, integrated with collaboration-(Interactive) scientific workflow, integrated with collaboration-enhanced office automation toolsenhanced office automation tools
Concrete concurrencyConcrete concurrency
Abstract concurrencyAbstract concurrency
Concrete workflowConcrete workflow
Abstract workflowAbstract workflow
Write scientific paperWrite scientific paper
(Word)(Word)Record experiment dataRecord experiment data
(Excel)(Excel)Individual experiment runIndividual experiment run
(Workflow orchestrator)(Workflow orchestrator)
Analyze dataAnalyze data
(SQL-Server)(SQL-Server)Share paper with co-authorsShare paper with co-authors
(Sharepoint)(Sharepoint)Collaborate with co-authorsCollaborate with co-authors
(NetMeeting)(NetMeeting)
Example:Example:
Scientific Information WorkerPast and futureScientific Information WorkerPast and future
PastBuy lab equipmentKeep lab notebookRun experiments by handAssemble & analyze data (using stat pkg)Collaborate by phone/email; Write up results with Latex
MetaphorPhysical experimentation“Do it yourself”Lots of disparate systems/pieces
PastBuy lab equipmentKeep lab notebookRun experiments by handAssemble & analyze data (using stat pkg)Collaborate by phone/email; Write up results with Latex
MetaphorPhysical experimentation“Do it yourself”Lots of disparate systems/pieces
FutureBuy hardware and softwareAutomatic provenanceWorkflow with 3rd party domain packagesExcel and Access/Sql-ServerOffice tool suite with collaboration support
MetaphorDigital experimentationTurn-key desktop supercomputerSingle integrated system
FutureBuy hardware and softwareAutomatic provenanceWorkflow with 3rd party domain packagesExcel and Access/Sql-ServerOffice tool suite with collaboration support
MetaphorDigital experimentationTurn-key desktop supercomputerSingle integrated system
Microsoft StrategyMicrosoft Strategy
Reducing barriers to adoption for HPC clustersEasy to develop
Familiar Windows dev environment + key HPC extensions (MPI, OpenMP, Parallel Debugger)Best of breed Fortran, numerical libraries, performance analysis tools through partnersLong-term, strategic investments in developer productivity
Easy to useFamiliarity/intuitiveness of WindowsCluster computing integrated into the workstation applications, user workflow
Easy to manage and ownIntegration with AD and the rest of IT infrastructureLower TCO through integrated turnkey clustersPrice/performance advantage of industry standard hardware components
Application support in three key HPC verticalsEngagement with the top HPC ISVsEnabling Open Source applications via University relationships
Leveraging a breadth of standard knowledge-management toolsWeb Services, SQL, Sharepoint, Infopath, Excel
Focused Approach to MarketEnabling broad HPC adoption and making HPC into a high volume market
Reducing barriers to adoption for HPC clustersEasy to develop
Familiar Windows dev environment + key HPC extensions (MPI, OpenMP, Parallel Debugger)Best of breed Fortran, numerical libraries, performance analysis tools through partnersLong-term, strategic investments in developer productivity
Easy to useFamiliarity/intuitiveness of WindowsCluster computing integrated into the workstation applications, user workflow
Easy to manage and ownIntegration with AD and the rest of IT infrastructureLower TCO through integrated turnkey clustersPrice/performance advantage of industry standard hardware components
Application support in three key HPC verticalsEngagement with the top HPC ISVsEnabling Open Source applications via University relationships
Leveraging a breadth of standard knowledge-management toolsWeb Services, SQL, Sharepoint, Infopath, Excel
Focused Approach to MarketEnabling broad HPC adoption and making HPC into a high volume market