January 29, 2001
Phase One of the Pittsburgh Supercomputing Center Terascale System is Operational
Testing and development are ahead of schedule with better than expected stability and performance.
PITTSBURGH The initial 256-processor configuration of the Terascale Computing System has passed acceptance tests with flying colors, say researchers and system engineers at Pittsburgh Supercomputing Center. This first phase TCS, which will evolve to become the most powerful system in the world available for public research, has shown better than expected speed and reliability and is already running applications from "friendly user" researchers well ahead of the scheduled Feb. 1 date to become operational.
Installed at PSC in October, pursuant to a $45 million grant from the National Science Foundation, the initial system consists of 64 interconnected Compaq ES40 Alphaservers, each housing four EV67 microprocessors. This initial TCS itself ranks 70th among the world's top 500 supercomputer sites, with a peak capability of 342 billion calculations per second (Gflops). The final system, when fully installed later this year, will have a peak capability exceeding six trillion calculations per second (teraflops).
To assess performance, PSC subjected the initial TCS to a range of "benchmark" tests. "Expected performance levels, based on other Alpha systems, were consistently exceeded," said Sergiu Sanielevici, PSC assistant director for scientific applications and user support. For the LINPACK benchmark software, which tests linear-algebra capability, the 256-processor system is rated at 263.6 Gflops.
Since installation, the system has maintained a realistic workload, and during acceptance testing, from Nov. 23 to Dec. 22, had virtually zero down-time. "The system was up 99.98 percent of the time during the acceptance period," said PSC assistant director of systems and operations, J. Ray Scott. "We're encouraged by what we've seen so far."
"Test user" researchers ran a variety of applications to further assess the system. "The multi-processor performance beats any machine we have run on to date by a considerable margin," said Indiana University physicist Steve Gottlieb, a member of the MIMD Lattice Computation group of physicists who develop computational methods for Quantum Chromodynamics.
Other researchers reported similar results: "This is the fastest large-scale system I have benchmarked NAMD on to date," said Jim Phillips of the Theoretical Biophysics Group at the University of Illinois Beckman Institute for Advanced Science and Technology. NAMD is the group's parallel molecular dynamics software.
PSC officially accepted the initial TCS from Compaq at the end of December and since then has implemented hardware and software modifications to make it compatible with the existing PSC computing environment. The "friendly user" phase of testing will further stress the system to shake out bugs in preparation for April 1, when scientists with allocations of time through the NSF Partnerships in Advanced Computational Infrastructure program will begin production research.
PSC will soon begin to install and test the full-scale TCS, which when complete will consist of 682 or more Compaq Alphaservers, more powerful than the ES40s of the initial system, each housing four EV68 microprocessors. When testing is complete, production research will shift to the full TCS. The largest-scale currently installed Compaq Alphaserver systems are 512 processors. Sometime this summer, PSC will push beyond that level of scaling, and TCS testing and development will enter a phase not charted before.
More information on TCS: http://www.psc.edu/machines/tcs/status/
The Pittsburgh Supercomputing Center is a joint effort of Carnegie Mellon University and the University of Pittsburgh together with the Westinghouse Electric Company. It was established in 1986 and is supported by several federal agencies, the Commonwealth of Pennsylvania and private industry.