GOVERNMENT

FRAMINGHAM, Mass. - The HPC User Forum, consisting of high performance computing users from government, industry and academia, held its first dialogue meetings between European and U.S. users last week in Bristol, UK and Annecy, France. The meetings were led by IDC's Earl Joseph II, Executive Director of the HPC User Forum. A combined 87 attendees heard from representatives of leading European and U.S. high performance computing users, including Airbus, Boeing, Commissariat Energie Atomique (CEA), Ecole Polytechnique Fédérale de Lausanne (EPFL), CSC Supercomputing Solutions, Hochleistungsrechenzentrum Stuttgart (HLRS), Network Computing Services, Inc. (systems integrator and computing facilities manager for the Army High Performance Computing Research Center) and the Pittsburgh Supercomputing Center. There were also updates from HPC vendors Cray, Hewlett-Packard, Intel, NEC, SGI and Sun. "High performance computing is a global industry with a global community, but the market dynamics can vary substantially from country to country and region to region. To advance the HPC industry and influence the direction vendors take, it's important that European and U.S. users understand each other's perspectives," Joseph said. "That's the primary goal of these initial meetings." Joseph explained the rationale for the HPC User Forum: "Five years ago, a varied group of HPC users/buyers asked IDC to help address their marketplace concerns. Many products and solutions were being designed for the commercial market, with only a few tweaks for HPC. Users were concerned about what they could buy to meet their most challenging requirements. They also wanted us to promote interaction among users and disseminate information about their achievements around the world." The HPC User Forum is directed by a steering committee consisting of users, and is operated for the users by IDC. At the highest level, issues described by European and U.S. users during the meetings were remarkably similar: the continuing need for both capability and capacity systems ("no single type of HPC system is right for everything"); vendor responsiveness; performance modeling and measurement; and, as always, the insatiable demand for more computing power. Airbus' Nigel Barry said his firm needs improved HPC analysis capabilities to keep new aircraft within weight limits and to tackle problems such as asymmetry caused by prop wash. While some problems have been successfully moved to Itanium-based systems, others still run best on vector supercomputers and on SMPs. Jim Kasdorf, Pittsburgh Supercomputing Center, said the big questions for PSC and NSF are about obtaining the full benefits of Linux and Grids. He described PSC's successful Terascale Computing System as "in effect, a Cray T3E or T3F built out of components from Compaq" and said "the cost of maintaining Moore's Law suggests a COTS approach." Suresh Shukla of Boeing delivered a well-received talk on HPC purchasing decisions. He said the main components of decisions-cost, applications, requirements and technology-are frequently out of phase. "Decisions on airplanes are made 8-10 years in advance. We have to go to management today, a bad time for the airline industry, about what Boeing will build a decade from now." Shukla said in recent years, HPC users have focused more on cost-reduction than solving bigger problems. "I have nothing against reducing costs, but have we forgotten the value an engineer or scientist can add to a design?" John Rawlins, CSC Supercomputing Solutions, said "large structural analysis still runs best on vector supercomputers, while midrange systems- HP Superdomes, SGIs, and IBM SPs- do well on problems with midrange memory bandwidth requirements, and Beowulf clusters are good for less-challenging things (which we all have lots of). I agree with Suresh that many users have lost sight of the value of HPC. The need for supercomputer power is increasing rapidly. The hard part is making the business case for it. In earlier times, we could talk about saving 7 to 10% off the value of the SuperFighter. Users aren't doing this today, and vendors can't." CEA/DAM's Jean Gonnord talked about the success of their TERA 1 supercomputer, a 5-teraflop peak system consisting of 640 Compaq ES-45 nodes. The system has sustained 1.3 teraflops and is currently at 80% utilization. "We have never had this much progress when changing to a new machine, although we need to adapt simulation codes and environment software to the cluster architecture and make it bulletproof. CEA/DAM's goal is 10 sustained teraflops by 2005 and 100 sustained teraflops by 2009. The talk by Paul Muzio of Network Computing Services, Inc., focused on their initial results using the Cray X1 vector supercomputer. On a CFD application, the X1 was more than 50 times faster than their Cray T3E, about 25 times faster than their Origin 3000, and 10-15 times faster than their IBM SP Power 4. Muzio cautioned that the X1 is better at MPI codes with vectorizable kernels than scalar codes or traditional "Cray-1" codes. Christine Soleil said EPFL (Ecole Polytechnique Fédérale de Lausanne) has a heterogeneous mix of HPC systems: HP Superdome, MPP cluster, IBM SP3, SGI Origin 3800, NEC SX-5, IBM Regatta, and a T1 (Alpha-based system). EPFL's HP system, with 25 nodes and a Quadrics switch, was installed January 2003 and is used primarily for bio-simulation. EPFL's goal is "to build a Swiss HPCN fabric with a heterogeneous architecture able to deliver services and tools for wide variety of problems." Uwe Kuester of HLRS (Hochleistungsrechenzentrum Stuttgart) discussed the upcoming procurement of a large system, called "the HLRS Simulation Workbench," that will require sustained performance of up to 2 teraflops and a powerful file system. Memorable quote: "We think it makes no sense to cluster floating point units without increasing memory bandwidth. The trend for all microprocessors is that memory bandwidth progress is lagging." Chris Willard, IDC, provided examples of IDC research related to HPC and discussed recent market trends. "The big trend is that vendors are leveraging commercial technologies into technical markets. The challenge is making sure these technologies are really oriented toward the technical market." Willard said the bio-sciences and clusters markets are growing rapidly, and Grids are becoming mainstream. "We see Grids as a way of organizing and managing resources, rather than as a new architecture." He also noted that the overall technical server market declined in 2002 by 7% and is projected to grow by an average of 6.1% a year to reach $6.3 billion (US) by 2007. Joseph said the IDC Balanced Benchmark Rating is adding new metrics on power consumption, system footprint, price/performance, and memory intensiveness and will add other relevant metrics as they become available. He noted that Alan Snavely, San Diego Supercomputer Center, is leading a separate HPC User Forum initiative to develop better performance modeling tools. Joseph encouraged everyone to attend the upcoming ISC2003 conference (Heidelberg, June 24-27, 2003) and future HPC User Forum meetings in the U.S. (Princeton, NJ, September 16-18, 2003 and Dearborn, MI, April 12-14, 2004). Based on the popularity of the Bristol and Annecy dialogue meetings, he said the HPC User Forum plans to hold additional meetings in Europe and will publish specifics on the User Forum website: www.idc.com/hpc.