NERSC Releases Software Test for Its Next Supercomputer
September 12, 2008
BERKELEY, Calif.—The Department of Energy’s National Energy Research Scientific Computing Center (NERSC) is looking for a new supercomputer, but is not willing to spend millions of dollars on just any machine. The computer scientists and engineers want to know that their new supercomputer can reliably handle a diverse scientific workload, so they’ve developed the Sustained System Performance (SSP) Benchmarks, a comprehensive test for any system they consider.
The benchmarks were released in conjunction with the NERSC-6 request for proposals on September 4. This version of SSP marks the first time that both vendors and the performance research community can easily access all applications and test cases.
The NERSC-6 system will be the next major system acquisition to support the DOE Office of Science computational challenges. Operated by Lawrence Berkeley National Laboratory, NERSC is the flagship scientific computing facility for DOE’s Office of Science and a world leader in accelerating scientific discovery through computation, including the deployment and support of large-scale resources.
“SSP is a key part of NERSC’s comprehensive evaluation for large-scale systems,” says William Kramer, General Manager at NERSC. “We look for balanced systems that expand our computational and analytics capability by assessing the systems’ abilities to provide sustained performance, effective work dispatching, reliability, consistency and usability, for the entire range of the Office of Science computational challenges.”
Instead of peak performance estimates, that is, the number of teraflop/s that could potentially be performed, NERSC scientists and engineers are concerned with the actual number of teraflop/s that the system will achieve in tackling a scientific problem. NERSC staff refer to this concept as sustained performance, and measure it using the SSP.
The SSP suite consists of seven applications and associated inputs, which span a wide range of science disciplines, algorithms, concurrencies and scaling methods. Kramer notes that this benchmark provides a fair way to compare systems that are introduced with different time frames and technologies. The test also provides a powerful method to assess sustained price/performance for the systems under consideration.
“NERSC-6 will be the system that provides the best value overall for supporting the DOE computational workload, taking into account Performance, Effectiveness, Reliability, Consistency and Usability, summed up in the acronym PERCU,” says Kramer.
An updated version of the Effective System Performance (ESP) test, developed to encourage and assess improved job launching and resource management, was also released with the request for proposals, as are the other tests and benchmarks NERSC uses to assess large scale systems.
The new SSP suite can be downloaded from http://www.nersc.gov/projects/ssp.php.
The NERSC Center currently serves 3,000 scientists at national laboratories and universities across the country researching problems in combustion, climate modeling, fusion energy, materials science, physics, chemistry and computational biology. Established in 1974, the NERSC Center has long been a leader in providing systems, services and expertise to advance computational science throughout the DOE research community. NERSC is managed by Lawrence Berkeley National Laboratory for DOE. For more information about the NERSC Center, go to http://www.nersc.gov/.