IDC's HPC Forum starts another Benchmark Initiative

Dallas 07 Nov 2000 Many tried, but to date all efforts failed to define benchmarks that go beyond the Linpack used in the TOP500 and have wide acceptance by users. Eurobench, Parbench, the NAS kernels, and many others. But, nevertheless, the HPC User Forum organised by IDC, representing leading supercomputer users in government, industry and academia, reported progress on a plan to create better performance tests for this most powerful class of computers. Speaking at the SC2000 supercomputing conference in Dallas, HPC User Forum officials said improved tests are needed to advance scientific research, industrial engineering and classified government work, all of which rely heavily on supercomputers. It would be too easy to predict that also this very US oriented benchmark initiative will fail, but the approach, identifying a set of "representative" benchmarks is not different from earlier attempts. Supercomputer architectures from different vendors evolving into the same direction, could, however, help this initiative.

"Despite the strategic and economic importance of supercomputers, or HPC systems, there is no widely accepted standard for evaluating their performance today", said Debra Goldfarb, IDC group vice president, Worldwide Systems and Servers. "It is not uncommon for user organisations to find that the actual performance of a new, multi-million-dollar supercomputer is a small fraction of the stellar results produced on today's limited tests. Although these Linpack and peak performance test scores are nearly meaningless beyond `chest-thumping' publicity, in the absence of anything better they continue to be emphasized in many procurements."

Team Engaged To Develop Benchmarks. IDC analyst Earl Joseph II said the HPC User Forum has engaged a group headed by Robert Lucas, head of HPC Research at the Department of Energy's National Energy Research Computing Center (NERSC), to develop a new benchmark suite to test the overall scalable performance of HPC systems, and an additional benchmark initiative to test performance on a broad range of specific application codes. In a mid-September meeting, HPC User Forum members endorsed the preliminary plan outlined by Lucas.

"Bob Lucas and the other members of this NERSC and University of Tennessee team are world-class experts at benchmarking HPC systems," Joseph said. "They have years of leading experience and were involved in the development of the NAS Parallel Benchmarks and `Top500.'"

`Short List' Identified For Overall Scalable Performance Benchmarks. Lucas reported that since the September meeting his team has identified a "short list" of candidate tests for the overall scalable performance benchmark suite. This suite is intended to measure processor performance, as well as internal memory speed, disks, and external networks.

"We will now bring this selection process quickly to a conclusion in collaboration with the HPC user community," said Lucas. "For the processor (CPU) benchmark we intend to collaborate with the `Top500' organization. Our objective is to supplement, not replace, the Linpack test on which the `Top500' rankings are based," Lucas added.

Development Schedules Outlined. For the scalable benchmark suite, Lucas said, his team plans by the second quarter of 2001 to complete the following: finalize the selection of tests; engage HPC leaders to define an initial set of orthogonal benchmark kernels; develop pencil and paper descriptions and representative codes; and establish a Web site and publish initial results.

For the full application benchmark suite, in the same timeframe the NERSC and University of Tennessee team plan to identify an initial application and a domain expert partner; develop a generic methodology for benchmarking full applications; and deploy the prototype with the partner.

Benefits to Stakeholders. Goldfarb said the new tests will benefit all stakeholders:

  • The U.S. will be better able to maintain industrial and scientific competitiveness.
  • Worldwide industries that use supercomputers (automotive, aerospace, chemical, genomics/bioinformatics, petroleum, pharmaceutical and weather) will be better able to address their most complex, economically important problems.
  • Major supercomputing centers will be better able to support large, diverse user communities.
  • All user organizations will be better able to know which third-party applications perform best on which HPC systems.
  • Applications developers will be better able to know which HPC systems are most suitable for their needs.
  • HPC vendors will not need to increase benchmarking resources; they can focus on the applications domains that are important for their markets.


Ad Emmen

[News on Advanced IT][Calendar][Analysis][IT in Medicine]