by Mark Goldman
This is an abbreviated summary of Mark's talk.
|
by Mark Goldman |
On October 10, days before the Fall 1996 Cray User Group (CUG) meeting, the Department of Energy awarded Silicon Graphics/Cray Research the Accelerated Strategic Computing Initiative Blue Mountain contract to build a teraflops
computer. When one speaker unexpectedly cancelled during the week of CUG,
SGI graciously arranged to fly in Mark Goldman, ASCI project manager, to give
a first-hand account of the contract award and the project. This is an abbreviated summary of Mark's talk.
|
The Department of Energy's (DOE) Accelerated Strategic Computing Initiative (ASCI) is targeted to be a $1 billion, 10-year, high-performance computing program. Its objectives are:
- To replace nuclear weapons testing with three-dimensional numerical simulations
- To achieve a sustained rate of 1 teraflop on real simulation codes by the end of 1998
- To create advanced computational modeling and simulation capabilities to maintain the safety, reliability, and performance of the nuclear stockpile
Differences between ASCI Red and Blue
DOE has designed the ASCI program to investigate alternative supercomputing architectures. The ASCI Red project is looking at a massively parallel processing (MPP) approach, while the ASCI Blue project is investigating a cluster of shared-memory multiprocessors (SMPs).ASCI Red is headquarted at Sandia National Laboratory in Albuquerque, New Mexico. The ASCI Red contract was awarded to the Intel Paragon (1.8 peak teraflops, more than 8,000 Pentium processors). (An ironic sidenote is that after intitial delivery of the product, the Intel Paragon will be discontinued.)
The ASCI Blue project is comprised of two parts: ASCI Blue Pacific and ASCI Blue Mountain.
ASCI Blue Pacific is headquartered at Lawrence Livermore National Laboratory in Livermore, California. The ASCI Blue Pacific contract was awarded to the IBM RS6000SP (3.0 peak teraflops).
ASCI Blue Mountain is headquartered at Los Alamos National Laboratory, Los Alamos, New Mexico. The ASCI Blue Pacific contract was awarded to the SGI/Cray Origin2000 (3.0+ peak teraflops). The SGI/Cray award was the largest of the three awards.
ASCI Blue contract specifics
Los Alamos and Livermore jointly released the ASCI Blue RFP on February 12, 1996. Several venders submitted proposals on April 25, 1996. DOE elected to make two awards: one for ASCI Blue Pacific and one for ASCI Blue Mountain.In July, 1996, negotiations commenced. DOE awarded the ASCI Blue Pacific contract to IBM at Livermore in August 1996, and the ASCI Blue Mountain contract to SGI/CRI at Los Alamos in October 1996.
The SGI/CRI contract for ASCI Blue Mountain provides for three deliveries over the contract life
- The initial delivery system, 1996:
- 256-processor Cray Origin2000 (2 X 128) with 4 MB L2 cache
- 128 GB memory
- 2.5 TB RAID disk
- HiPPI-800 interconnect
- The technology refresh system, 1997
- The sustained stewardship teraflop system, 1998:
- 3072 SN1 processors
- 500 GB memory
- 75TB RAID disk
Collaboration with the Advanced Computing Laboratory
The ASCI Blue Mountain project will operate in conjunction with Los Alamos' Advanced Computing Laboratory (ACL) ASCI Grand Challenge Program, which released an RFP for complementary systems to the ASCI Blue Mountain supercomputer. The ACL system includes visualization components and is co-located with and may be coupled to the ASCI Blue System.
The ASCI/ACL partnership is a beneficial collaboration in many ways:
- Initial system, 1996
- Refresh system, 1997
- Final system, 1999
- SGI/CRI are technology leaders in simulation.
- Los Alamos is a true pioneer in high-performance computing.
- The lab will push the limits of SGI/CRI products and provide vital feedback.
- ASCI/ACL will work together to accelerate software tools and libraries system performance.
- The ASCI and the ACL systems can operate as one 4+ teraflop system with a simple programming model of shared memory.
- Bandwidth of distributed shared-memory (DSM) architecture allows the widest range of applications of any of the ASCI systems.
- The systems are configurable between classified and unclassified environments.
- The ACL system includes state-of-the-art SGI visualization applications.
- The systems will be used for hydrodynamic, structural, chemical, seismic, and environmental applications.
- New physics and algorithms will be developed.
How the SGI and Cray ASCI teams "merged"
The challenge to SGI/CRI of bidding for the ASCI Blue contract award was to merge the two different corporations and their cultures in a short period of time. We discovered that we could quickly integrate the two cultures to create a winning solution. Collaborating on ASCI Blue set the tone that SGI and Cray could effectively work together.SGI and Cray staff first met to discuss the ASCI project on March 27, 1996, three days after the SGI/Cray merger was announced. The tone of this rather odd meeting, to which both companies brought their own separate proposals, was was, "I'll show you mine if you show me yours." We discussed our respective high-level approaches on March 29. Proposal bids were due on April 11; DOE granted an extension till April 26.
We said: "OK, We have three weeks, what are we going to do?"
What made this work?
- We exchanged our proposals and began in-depth team meetings.
- Our basic architectural approach was similar and we discovered ... synergy.
- The team functioned as a team.
- Everyone was invested in making this a success.
- We all understood the importance of this program.
- Our teamwork set the tone for more than 150 participants on other teams in such areas as hardware design, software development, field support, and finance.
Comments to: lester@ucar.edu