Upload
others
View
15
Download
0
Embed Size (px)
Citation preview
© 2004 IBM Corporation
IBM Systems and Technology Group
May 11/12, 2005
David V. GelardiVP, Industry Solutions and Proof of Concept Centers [email protected]
Supercomputer Best Practices Seminar I
HPC Platforms in the Industry Solutions and Proof of Concept Centers
© 2004 IBM Corporation
IBM Systems and Technology Group
Agenda
Industry Solutions and Proof of Concept CentersMissions
Locations
Skills and Resources Supercomputing center best practices
Performance Tuning
Leveraging Results
© 2004 IBM Corporation
IBM Systems and Technology Group
Industry Solutions and Proof of Concept Centers
Design Center foron Demand Business
Customer BenchmarkProof of Concept
Deep ComputingCapacity on Demand
Development Center forSolution Integration
High AvailabilityCenter of Competency
High Availability
pSeries
zSeriesiSeries
xSeriesIBM TotalStorage
GRIDWebSphere DB2
SAP
Oracle
IT Architects
Project Managers
IT SpecialistsSystem Administration
Database AdministrationWAS
Tivoli
AIX
Middleware
Linux
i5/OS
z/OS
Security
SAP
Cisco
Virtualization
HPCAIX
i5/OS
z/OS
Windows
DB2
OracleNetworking
BladeCenter OpenPower
Linux Cluster
Blue Gene
Tactical, Closing Sales Driving Value, InfluencingClients
P&L CPU’s by the HourFirst of a Kind forOn Demand Technology
Solution OfferingArchitecture and Integration
STG PerformanceMarketing
eServer and TotalStorageBenchmark Planning & Marketing
20041,000+ Engagements
$1B STG Revenue7,000+ CPU’s
1.5 Petabytes Storage15 Locations WW
© 2004 IBM Corporation
IBM Systems and Technology Group
HPC MissionsWW Customer Benchmark Centers
Provide customer demanded benchmark capability WW for eServer and TotalStorage
Include proof of concept, scaling and performanceAssist in the execution of ISV Application benchmarks
Deep Computing Capacity on Demand CentersDevelop and deploy a Capacity on Demand offering that provides clients with an
alternative means to meet their peak computational needs, do so in a manner that is complementary with traditional server and storage sales.
Provide increased Business Value to our customers through innovative On-Demand delivery mechanisms.Provide a competitive differentiator for IBM.Drive incremental revenue and profit.
Performance MarketingLead series in the creation of plans for industry benchmarks and assist in the
marketing of results (e.g. TOP500)Participate in planning with Series product marketing and development to
optimize cross-series plans and leverage cross-series opportunitiesCreate eServer and TotalStorage marketing collateral and provide customized
assistance to sales opportunities
© 2004 IBM Corporation
IBM Systems and Technology Group
WW Benchmark & Design CentersPoughkeepsie
pSerieszSeriesxSeriesDesign CenterDCCoDHACoCLinux Clusters
MontpellierpSeriesiSerieszSeriesxSeriesDesign CenterFSS SolutionsDCCoDStorageLinux ClustersKirkland
xSeries (MS)
Rochester iSeriesDCCoDLife Sciences
Solutions
DallaspSeriesxSeries
Gaithersburg zSeriesStorage
TokyopSeriesxSeriesDesign CenterStorage
MainzStorage
BoeblingenzSeries (Linux)
Other Regional FacilitiesSeoul pSeriesSydney pSeriesSilicon Valley Lab HVWSGreenock xSeries
Examples of ProjectsDesign EngagementsHPC BenchmarksCommercial BenchmarksISV Enablement &
Benchmarking (w/R. Warren)BICoC (w/SWG)
HoustonDCCoD
BeijingpSeriesxSeriesiSerieszSeries
© 2004 IBM Corporation
IBM Systems and Technology Group
What is a Client Benchmark
A Demand to Prove a 'Capability' to a Specific ClientOften Part of an RFP ResponseOften CompetitiveClient Sometimes Needs Help Defining RequirementsCategories of Information are Performance, Scaling and Proof of Capability, etc.Client Sets Criteria against their Data and WorkloadStrict Response Date, Usually Very Short Term
Different from...Industry Standard Benchmarks (usually done in development)TPC-C, TPC-H, SPECnnn, STREAM , Pallas, NPB, Linpack HPL etc.Application Benchmarks (usually done in ISV Enablement)SAP SD, Peoplesoft, BAAN, Siebel, etc.
© 2004 IBM Corporation
IBM Systems and Technology Group
The Challenge: Optimizing HPC Capacity UtilizationTraditional infrastructure build-out increases in step-function phasesCompanies that build for average demand must be able to respond quickly to peak workload demands or suffer lost opportunity Companies that over-build to address peak workloads are left with over-capacity and under-utilization in business downturnIBM Deep Computing Capacity on Demand serves unfulfilled peak workload requirements
In-house Resources or Outsourced Data Center
Time
Cap
acity
UnfulfilledBusiness Peaks
UnutilizedResources
© 2004 IBM Corporation
IBM Systems and Technology Group
Management of Resources Multiple clients access “virtual”clusters made up of compute and storage resources
Dedicated to one client at a time – dedicated custom environment or– timesliced optimized environment (no root access, job scheduling)
Some resources open to all for testing, prototyping, tuningAllocations subject to availability
Highly secure and resilient infrastructure Client components
Application software and licensesDataCustom hardware
Compute NodeStorage Nodes + Storage HWManagement Node
Client A
Client B
Clients C, D, E Open nodes
Master Mgmt NodeSwitches
VPN Router
Tape Server
© 2004 IBM Corporation
IBM Systems and Technology Group
HPC Benchmark Systemsp5 Machines Open Powerp520 x 1 p720 x 4p550 x 2p570 x 26p595 x 3
p4 Machines PowerPCp615 x 1 JS20 x 8p630 x 2p630+ x 6p655+ x 12p690 x 1p690+ x 1p690++ x 1
Note: Quantities can vary based on active projectsAccess to Loaner equipment leveraged
Clusters32 x p575 w/ High Performance Switch8 x p575 w/ Gigabit Ethernet
16 x p710 w/ Gigabit Ethernet/Myrinet2 x Blue Gene
StorageESS x 1 DS4500 x 26SAN Switches x 11EXP 700 x 117
Pentium4x335 x 16
Intel EM64Tx336 x 36 w/Myrinet
and IBx346 x 4
Opteron250e325 x 64e326 x 68 w/ Myrinet
and IB
Xeonx345 x 26
© 2004 IBM Corporation
IBM Systems and Technology Group
IBM DCCoD: Scale Beyond In-House HPC Limits
Client HPCInfrastructure
Virtual Private Network
Houston, TX
Poughkeepsie, NY Montpellier, France
= Variable Capacity / Variable Cost
= Fixed Capacity / Fixed Cost
4 DCCoD centersIntel® Xeon ™AMD Opteron™IBM POWER™IBM Blue Gene®
7000+ CPUs
Rochester, MNNew!Available
March 2005
Secure Internet access to supercomputing power owned and hosted by IBM enables clients to rapidly and temporarily flex up/down HPC capacity proportional to business demands - to
respond to peak workloads and capture business opportunities that would otherwise be out of reach.
IBM HPC Grid
© 2004 IBM Corporation
IBM Systems and Technology Group
IBM Blue Gene® Capacity on Demand
Rack2,048 PowerPC® CPUs
2/8/5.6 TF512GB Memory
Flexible and convenient pay-for-use access to reserved capacity for scientists, researchers, anddevelopers - to a new family of supercomputers optimized for scalability, bandwidth, and massive data handling while consuming a fraction of the power and floor space required by today’s fastest systems.
Multi-use on demand access to …Up to 2x 1,024 dual PowerPC®processor compute nodes per rack64 IO Nodes per rackFront-End Node I/O File Server Service Node
… on a sub-rack basis
World’s fastest supercomputer!Ultra scalable performance Ultra floor space densityUltra performance per W of powerInnovative architecture and system designFamiliar programmer/user Linux-based Environments
#1 on TOP50070.72 TFLOPs
sustained
AvailableMarch 2005
© 2004 IBM Corporation
IBM Systems and Technology Group
Skills Scientific and technical application skills
Architecture portingAlgorithmsCompilers: Fortran 77, Fortran 90, C , C++, Libraries & ToolsApplication tuningParallel Programming, message passing Performance analysis
System Administration skillsJob scheduling: Load Leveler, PBSOS skills: AIX, Red Hat, SLES distributionssystem tuning SW installs and systems management, ID managementInterconnect: High Performance Switch, Myrinet, Topspin IB, Voltaire IBDisk subsystems, GPFSTroubleshooting
Infrastructure SkillsHardware provisioningNetworkingSecurity and Access
Project Management skills
© 2004 IBM Corporation
IBM Systems and Technology Group
HPC Typical Workloads
Cluster SizeStandalone Large
Uni
High
MPI Codes
Ope
nMP
Cod
es
Univer
sities
CAE:Nastran,STAR-CD
Petroleum:ReservoirModeling
Weather: MM5Life Sci: Blast
Petroleum:SeismicGov't: RYO
Life Sci: Charmm,
AmberCAE:Fluent,
AbaqusLife Sci:Gaussian
Kernels
© 2004 IBM Corporation
IBM Systems and Technology Group
Performance Tuning Process
Result correctness is the basis of all tuningReference data from customerComparison of results from very different levels of compiler optimizations.
Platform analysisProfiling and analyzing code for potential improvements
Compiler optimizationPerformance libraries such as IBM ESSL and MASSExplore auto-parallelization by compiler and ESSLSMP (part of ESSL library) libraryApply hand tuning (optimization, OpenMP parallelization and/or MPI parallelization) based on the profile
Repeat, Repeat, RepeatWithin time and cost/benefit constraints
© 2004 IBM Corporation
IBM Systems and Technology Group
Discussion PointsMission – What is it?Planning is KeyInfrastructure is KeySecurity a NightmarePower / CoolingSkillsBroad Applications (and long lived!)No Test Systems (FOAK)SLAsChiba City Like OS/Software StackBenchmarking v. Production (DCCoD)Metrics / Cust Sat / SurveysUser InteractionRapid Turnover of Assets In/OutWhat is not 7x24x365?