Compute Resource Systems

1
Compute Resource Systems

• Wayne Pfeiffer
• Deputy Director
• NPACI SDSC
• NPACI All-Hands Meeting
• January 28, 1999

http//www.npaci.edu/Resources/index.html
2
NPACIs balanced complement ofhigh-end resources
for FY99

• Compute resources (LES 4 partners)
• Teraflops system at LES first for academia
• Complementary systems at partner sites
• Data resources (LES 10 partners)
• gt100 TB mass store at LES largest based on HPSS
• gt100 GB data sets at partner sites
• Network resources (LES all partners)
• gt100 Mbps access to compute data resources
• Communications backbone for metacomputing

3
Compute resources are at 5 sites
U Michigan
UC Berkeley
Caltech
SDSC
U Texas
4
Complementary rolesof 5 compute resource sites

• Leading-edge site (SDSC)
• Very high-performance resources, including
  teraflops system
• Mid-range sites (U Texas U Michigan)
• Smaller systems compatible with LES
• Support for apps with limited scalability,
  large-memory jobs, apps development, OS testing,
  education
• Alternate architecture research systems
  (Caltech, UC Berkeley, SDSC)
• Support for leading-edge apps, thrusts,
  evaluation

5
Evolution of allocable compute servers
Quarters are by fiscal year.
6
Nationally-allocable compute resourcesfor NPACI
in FY98 FY99
FY99 changes are in bold.
7
Research compute resourcesat UCSD, Caltech,
UCB in FY98 FY99
FY99 changes are in bold.
8
First Tera MTA is at SDSC
9
IBM selected as first NPACI teraflops vendor

• Briefings (with partners) from 5 vendors
• Compaq/Digital, HP, IBM, SGI, Sun
• Selection endorsed by Executive Committee
• Strong commitment to high end by IBM
• Technology being developed through ASCI
• SDSC to have largest system with next-generation
  nodes
• Growing partnership with IBM

10
SP teraflops system coming this year

• Cluster architecture
• gt1,000 Power3 processors at gt200 MHz
• 8-processor, next-generation SMP nodes
• gt500 GB of memory initially, with upgrade later
• Current generation switch initially, with upgrade
  later
• Staged installation
• 1/4 teraflops this summer
• Full teraflops this fall
• Memory switch upgrade next year

11
Reference plan for multi-teraflops computing
12
More aggressive planfor multi-teraflops computing
13
Features of more aggressive planfor
multi-teraflops computing

• More rapid deployment of multi-teraflops system
• Current 1 Tflops in FY99 -gt 4 Tflops in FY02
• Rapid 1 Tf in FY99 -gt 3.7 Tf in FY00 -gt 13 Tf in
  FY01
• Fits within existing SDSC building
• Includes balanced mass storage system
• Current 1 PB in FY02 -gt Rapid 12 PB in FY02
• Re-use of equipment
• Modest investment in alternate architecture
• Additional cost at leading-edge site of about
  60M per year