FEATURES & COMMENTARY
San Diego, CALIF. — Sixty-six people participated in first annual meeting of [email protected], a user led forum in high performance computing. Participants came from government agencies, research labs, industry and academic computing centers. In addition representatives from each of the major HPC vendors also attended.
The plenary meeting session ensued with the major topic being development of metrics and benchmarks to ascertain HPC system performance. Earl Joseph introduced the topic and pointed out the shortcomings of LINPAK as an appropriate benchmark for HPC systems. He also introduced the two phases of the IDC2000 Users Forum benchmarking plan: 1) development of new simple benchmarks and metrics to characterize performance, and 2) the creation of a table of applications benchmarking results contributed by the member organizations that would be available to all members. Larry Davis chaired the session on performance, which included presentations by the HPCMP, NSA, Ford, Boeing, Pittsburgh Supercomputer Center (PSC), SDSC, and NERSC.
Terry Blanchard gave the HPCMP presentation and gave an excellent overview of their ongoing benchmarking activities and their use in current HPC hardware acquisitions. Mike Merrill of NSA presented their principal benchmark (GUPS), which measures performance on random memory accesses. This led to a discussion on the general issue of memory vs. CPU performance and their relative importance in HPC systems, which continued later in the panel discussion. Alex Akkerman of Ford and Barry Sharp of Boeing discussed the types of applications benchmarks that they use, and stated that they would be willing to contribute benchmark results to the User Forum activity to create a table of application benchmark results, provided the HPC hardware vendors agreed to release the results.
Jim Kasdorf of PSC discussed their use of benchmarks and strongly supported continued updates to the NAS parallel benchmarks. Allan Snavely of SDSC discussed the general problem of how to predict HPC system performance over a range of application problem sets from a sparse set of benchmark data points. He proposed a new effort to develop scalable synthetic benchmarks based on a careful study of how applications codes behave that would produce a “system signature” of performance in several performance dimensions. He discussed a simplistic performance model for memory-bound applications that seems to produce a reasonable estimate of actual system performance in this case. He readily admitted to the complexity of the problem and stated that accurate predictions of HPC system performance can only come from an accurate understanding of both the applications software and the hardware.
Bill Kramer and Bob Lucas of NERSC then discussed the NERSC effective systems performance test, which measures a system’s efficiency in scheduling and executing a typical system workload, and a new set of synthetic application benchmarks that will focus on memory and communications performance. This set of synthetic applications benchmarks form the heart of the IDC2000 initiative to define a more useful systems performance metric than LINPAK. NERSC’s eventual goal is to combine this set of benchmarks with their effective systems performance test to accurately characterize a system’s performance in a multi-user environment. Earl Joseph stated that the first milestone for the IDC2000 initiative is to announce the selection of the first set of pseudo-applications to be used as benchmarks by SC2000. Eric Strohmeier reviewed some of the details of the IDC 2000 initiative.
The next meeting session began with feedback on the IDC2000 performance metrics plan from the three sets of users represented at the meeting: government, university and industry, and HPC hardware vendors. Each group generally endorsed the plan. Government comments included a request to the vendors to allow release of benchmark results on their systems, and vendor comments expressed a willingness to dedicate resources to benchmarking efforts, but asked that government acquisition officials be mindful of the effort required to perform benchmarking activities.
Several of the hardware vendor representatives gave short presentations on their current products and future directions. Vendors represented included SRC, Fujitsu, Compaq, IBM, SGI, Sun, and Cray. Of notable interest was the recent success of Compaq in capturing several very large acquisitions (ASCI 30 TF system and PSC NSF system). In addition, Etnus, who makes the TotalView parallel debugger, gave a short presentation on their products and plans.
IDC provided a brief update on a number of different HPC research studies including automotive/aerospace research, clustering trends, Linux market directions, vendor market shares and HPC forecasts.
Jose Munoz of DOE presented an update on the ASCI program. The ASCI White system at Lawrence Livermore National Laboratory is currently being constructed, and will have a 12 TF IBM SP P3 system. The ASCI Q system, planned for 2002, will be a 30 TF Compaq system with alpha processors. Jose presented some utilization results from the ASCI Blue Mountain system (SGI Origin architecture) that showed typical utilization on that system ranges from 60 to 80 percent. He stated that most of ASCI’s current simulations run on 2,000 to 4,000 processors, and that they are still trying to determine which kinds of platforms best run their applications. He discussed the need to maintain overall system balance over several system parameters and presented diagrams showing capabilities in these dimensions for their current and future systems. He also discussed their scientific visualization activities. His presentation created considerable discussion concerning ASCI’s acquisition policies, which seem to stress maximum CPU performance in peak GFs. Jose pointed out that while this may have been true, ASCI has made major investments in software applications to be able to use these theoretically very powerful systems efficiently.
David Kahaner of the Asian Technology Information Program then introduced Dr. B.T. Cheok of the Singpore Institute for High Performance Computing. This organization has a mandate to use leading edge high performance computing resources to enhance Singapore’s global competitiveness. The institute sponsors both people development programs and research collaborations. They also operate relatively modest HPC systems for Singapore industry and academia. Dr. Cheok extended an open invitation for Users Forum members to visit their institute if they are in Singapore.
Paul Muzio and Steve Karwoski then presented reviews of their respective distributed centers. Both discussed their respective center capabilities and activities.
Chris Willard of IDC then presented an overview of the IDC market forecast for the technical computing market, which was approximately a $5B market in 1999. They segment the market into the following four categories: capability, enterprise, divisional, and departmental. They consider high performance computing to be those systems in the capability and enterprise categories, with the dividing line between enterprise and divisional systems being $1M. IDC also presented a fairly detailed taxonomy of HPC systems categorized by memory and processor distributions. IDC forecasts fairly strong growth in the HPC marketplace over the next several years, with a higher growth rate for technical computing in general. IDC also produced a market forecast for the use of Linux in these systems which shows substantial growth through 2004, but with Unix remaining the primary operating system for technical computers throughout the period.
The full users forum meeting was preceded and followed by short meetings of the group’s steering committee. All user site members were invited to participate in the decisions guiding the development and direction of the forum. Vendors do not attend the steering committee meetings. Earl Joseph convened the first steering committee meeting and discussed the structure of the users forum, including the newly-formed Performance Advisory Group. Discussions about the level of formality in organization and accompanying rules and bureaucracy eventually led to an agreement that a smaller selection of user representatives (10 to 15) would need to be selected to provide overall direction to the user forum, but the members chose to not move to a highly formalized structure at this time. For now members agreed to ask Earl Joseph of IDC to continue as executive director of the forum.