standard building blocks for grid and hpc clusters · 2006-09-06 · standard building blocks for...
TRANSCRIPT
Standard Building Blocks for Grid and HPC Clusters
LAURENCE LIEW [email protected]
Director, Open Source Grid Development Centre
Platform Computing, August 2006
© Platform Computing Inc. 2003-20062
Agenda
Platform Computing
HPC History & Today
HPC in the Future
A Standard?
Q&A
1
2
3
4
5
© Platform Computing Inc. 2003-20063
Platform Overview
� World-leading Grid software solution provider� Industry-leading partnerships with Dell, HP, IBM, S GI, SAS, …� Global Presence -- clients, offices, service, suppor t, training � 1,600 Fortune 2000 customers around the world
� Growing profitably every year since inception� 14 years experience with mature, proven production solutions � Dedicated vertical industry teams and expertise � Strong commitment to standards (OGSA, GGF etc.)
• Core Product Team: 120 Employees in Development26 Ph.D’s; Many Masters Degrees
• Support Team: 35 people10 Phd’s
• 3 Development Centres: Canada, China and SingaporeSingapore -> Open Source Grid Development Centre (US$5M investment)
© Platform Computing Inc. 2003-20064
Industry adoption and technical innovation over 14 years
1992• Aerospace
• Telco OEM’s
1992• Aerospace
• Telco OEM’s
Over 1,000 production grids deployed
0 to 1,600 clients0 to 1,600 clients
1993• Gov’tResearch
• Universities
1993• Gov’tResearch
• Universities
1994• Semiconductor
1994• Semiconductor
1995• Pharmaceutical
• Chemical
1995• Pharmaceutical
• Chemical
1996• Oil and Gas• Computer
Graphics
1996• Oil and Gas• Computer
Graphics
1997• Automotive
• Software (ISV’s)
1997• Automotive
• Software (ISV’s)
1998• Defense
• Telco services
1998• Defense
• Telco services
1999• Life
Sciences
1999• Life
Sciences
2000• Banking
and Financial
2000• Banking
and Financial
2001• Systems integrators
2001• Systems integrators
2002• Data mining
services
2002• Data mining
services
1992 - Batch workload management
1992 - Batch workload management
Platform Industry client
Platform grid technology
1996 – Multiple clusters
1996 – Multiple clusters
1997 – Reporting
1999 – License pre-emption
1999 – License pre-emption
2000 - Analytics2000 - Analytics
2001 – On-line workload management
2001 – On-line workload management
2005• Telco
2005• Telco
2005 – VM workload
management
2005 – VM workload
management
2002 – Jobflow automation
2002 – Jobflow automation
© Platform Computing Inc. 2003-20065
Applications Silos Today
The Need is for Variable Resources to Meet Variable Business Demand The Need is for Variable
Decouple the applications from resources
The Result is IT Agility
© Platform Computing Inc. 2003-20066
Platform Grid Architecture for SOI/SOA
ApplicationApplication Application Application Application
Enterprise Applications
Results: Model architecture that enables Enterprise Grid
SOA
SOI
Resource Orchestration Platform Enterprise Grid Orchestrator™ (EGO)
Platform Enterprise Grid Orchestrator Developers Edition
Platform Symphony™
Platform LSF®Family of Products
Platform Virtualization Orchestrator ™
3rd Party Integrations
Application Orchestration
© Platform Computing Inc. 2003-20067
Multi-Cluster
Platform LSF Family Products Overview
License Scheduler
Macrovision FLEXlm
Application License Optimization
Reporting
Web Application
Job Submission Script
Workload Management Cluster
Superior Workload Management
System Manageability
Cluster
Analytics Integration with Leading Applications
INTEGRATED APPLICATIONS FROM VENDORS
Abaqus
IBM
Ansys
LSTC
Fluent
MSC
© Platform Computing Inc. 2003-20068
HPC in our lives: Some common examples
Improve Business PerformanceIn a world of increasing pressure to deliver faster , better and cheaper, being first and fastest is critical to profitabilit y and competitiveness.
Fundamentally, HPC provides two core benefitsThroughput: Get more done in the same time� Example: Computer chip verification test suites ~ 1 Million jobs per day
Optimization: Do more with Less ~ Do the Important Things First
Free Thinkers From Constraint ~ remove boundariesImagine the impact of curing a life-threatening disease by saving valuable time on the road to discovery. Imagine how safe our roads would be if all cars could be quality and safety tested repeatedly at a fraction of the cost of physical prototype testing.modeling the fluid flow and combustion in a car engine to maximize fuel efficiency optimizing the scheduling of work on production lines to maximize utilization accessing huge databases for on-line transaction processing and data miningof the banks' ATM networks Hollywood, video games and animationLong running batch applications: end of day accounts or billing
© Platform Computing Inc. 2003-20069
Competitive and Financial Benefits of HPC
“It has been a continuous stream of revenue to our bottom line, giving us the ability to look into other development areas.”
“It drives innovation, R&D effectiveness, and productivity.”
“Our intent is to migrate to a management policy of disposable computing”
“The time to market would prohibit our
business from existing.”
“The ROI is returned within one year of the purchase.”
Reduced Costs
Business Survival
Return on Investment
Increased Competitiveness
“… saved us hundredsof thousands of dollars in hardware costs, improved our computingperformance, and gave us a robust system…”
Sources: Platform 2003-4; Council for Competitiveness HPC User’s Summary
© Platform Computing Inc. 2003-200610
History of HPC
In the past, “Supercomputers” ruled the day
The “supercomputers” in the 1980's used proprietary, custom-designed hardware (CPU, memory, disk and interconnect) to achieve high performance levels
Commonly referred to as “Capability Computers”
Very expensive, single vendor and proprietary
Systems were frequently “lashed together” in a loosely coupled cluster and shared
Workload schedulers kept the supercomputer fully utilized (>90%)
HPC was constrained to the “privileged few” – wealthy and savvy typically used for grand challenge problem solving
Solutions were “built” not bought
Supercomputing technology primarily evolved through significant and expensive R&D
Engineered performance: “the better mouse trap”
© Platform Computing Inc. 2003-200611
HPC Today
We have evolved from Supercomputers to “Supercomputing”
Solutions are commonly bought, not built
Clusters of computers are common and moving beyond academic, research and science
Today, HPC is the domain of the capable and enabled
Price/performance is now attractive and compelling (3:1 ratio over SMP)
The market is driven by Commoditization and Standardization
© Platform Computing Inc. 2003-200612
HPC Linux Clusters
Common today: Beowulf Style
clusters.
Multi-Cluster: a cluster of clusters interconnected
© Platform Computing Inc. 2003-200613
Why Should I Care about Commoditization and Standardization?
The effects of Commoditization change the market; the forces arecontinually at work and everyone will benefit
Customers can buy bundled, pre-integrated clusters from several vendors at attractive prices today
These solutions are built on a combination of industry standard and commodity hardware components
Today these solutions are increasingly built with low cost, best of breed, open source software stacks
Differentiation now lies with the whole product solution and is centered around support and services, not engineered performance
Less about the “better mouse trap” more about “fewer mice”
© Platform Computing Inc. 2003-200614
The Linux Paradigm Shift
HPC YESTERDAY
ProprietySoftware
ProprietyHardware
ExpensiveClosedSingle Source
HPC NEXT GENERATION
ProprietySW
Best of BreedOpen Source
Linux
Superior ValueStandards - BasedBest of Breed functionalitySingle responsibility
HPC TODAY
WLM
CM
SM
OS
Linux
Misleadingly inexpensiveOpenMultiple Vendors
Dec
oupl
edS
oftw
are
Com
pone
nts
In the past, large UNIX servers were clustered together with systems management provided by IHVs - The stack was proprietary
© Platform Computing Inc. 2003-200615
Vertical Solutions: application Prototyping / Sizing- Energy/Petroleum - Life Science
- Automotives – Manufacturing and Design
- Custom application benchmarks- Standard benchmarks- Performance studies
Resource Monitoring / ManagementResource dynamic allocationCheckpoint restarting and
Job redistributing
Compilers and math libraryPerformance tools
- MPI analyzer / profiler- Debugger
- Performance analyzer and optimizer
MPI 2.0 / Fault Tolerant MPIMPICH, MPICH-GM, MPI/LAM, PVM
Interconnect Technologies- FE, GbE, 10GE… (RDMA)- Myrinet, Quadrics, Scali
- InfinibandManagement Hardware
Interconnects Hardware
Interconnect Protocols
Operating Systems
Middleware / API
ClusterHardwareSoftware
Monitoring &Management
Application
Node Monitoring & Management
Benchmark
Development Tools
Workload Management
Platform Hardware
ClusterInstallation
ClusterFile System
- Reliable PVFS- GFS , GPFS …
- Storage Cluster Solutions
IA-32, IA64 (Processor / Platform) comparison
Standard rack mounted, blade and brick servers / workstations
Cluster monitoring Load analysis and
BalancingRemote accessWeb-based GUI
Cluster monitoringDistributed System
Performance Monitoring Workload analysis and
BalancingRemote accessWeb-based GUI
Remote installation / configurationPXE support
System ImagerLinuxBIOS
Cluster Functionality: the Building Blocks
© Platform Computing Inc. 2003-200616
HPC in the Future
The Near Future:
“Personal Supercomputing”
The domain of the average, non-savvy
What’s really happening? New Market Disruption
…enables a new class of users to acquire and use th e technology for a growing set of problems
Customer Expectations: The Requirements have shifted
Provide an end-to-end solution
Reduce business risk and consistency
Single point of contact for sales, services, support
Provide customers a “single throat to choke”
Simplify purchase, deployment and management
Reduce the complexity and make clusters easy to deploy and use – over and over again
© Platform Computing Inc. 2003-200617
A Bold Statement
HPC technologies will be adopted by mainstream IT customers as part of some larger “whole product”.
…as part of a larger Middleware footprint
…an ERP Application Suite
…or as part of standard Linux and Windows Clusters
The industry needs to place HPC products in a manner that matches buying patterns
What is needed: a standardized, consistent, replicable approach
© Platform Computing Inc. 2003-200618
What the Detractors Say
“It’s too hard”
“Why?”
“Not possible”
“Huh?”
Drop by www.platform.com/Rocks and see what we have been able to do.
Q & A