Features:

HPC USER FORUM EXPLORES CURRENT STATE OF HPC

The HPC User Forum held its semi-annual meeting in Tucson last week and also described the agenda for the third U.S.- Europe dialogue meeting, to be held October 13, 2004 at CERN.

At the CERN meeting, leading users from both sides of the Atlantic will discuss notable HPC developments and common issues. IDC will present a global market update that includes impressive new growth. European speakers will represent CERN, CINECA, CSCS, SARA and others. The U.S. contingent will include representatives from Boeing, the Ohio Supercomputing Center, the Pittsburgh Supercomputer Center, Network Computing Services and IDC. Registration is free and all HPC users are invited to participate. For more information about the agenda and to register, go to http://www.hpcuserforum.com.

The semi-annual U.S. meeting explored the current state of HPC in the life sciences, current and emerging programming models and benchmarking activities, and major government HPC programs, according to Steering Committee Chairman Paul Muzio, VP-Government Programs for Network Computing Services, Inc. and Support Infrastructure Director of the Army High Performance Computing Research Center. The group also heard technical presentations from HPC vendors. There were 116 attendees from government, industry, academia and vendor organizations.

Chris Willard, IDC, presented the HPC market update. Among other things, he noted that cluster market share continues to grow steadily and now represents 30 percent of the HPC market. IDC's Addison Snell summarized the U.S. Council on Competitiveness' recent study of U.S. industrial HPC users. Nearly 97 percent of the users surveyed said HPC is indispensable for their ability to compete and exist. The largest issue is the shortage of human experts able to apply HPC resources effectively. IDC reviewed key findings of a multi-client study on clusters and grids. About half of those surveyed said they are not ready to implement grids or not interested in doing so, but half are using or planning to implement grids.

Vendors attending for the first time: Level 5 Networks, CPU Technology, Inc., Paracel, Orion Microsystems and Look Dynamics.

Dr. Sangtae Kim, NSF, described the new Division of Shared Cyberinfrastructure. The new model: massive data generated at the periphery, needing to be fed to the high-end computers at the core. This will require novel systems and architectures. He said HEC will likely get 1,000 times more powerful in the next decade.

Michael Banda, Lawrence Berkeley National Laboratory, said that using the IBM SP NERSC system, Berkeley Lab and the University of California have created the first 3D global map of the protein structure universe. The model closely mirrors results based on visual observations.

Jeff Nichols, Oak Ridge National Laboratory, talked about ORNL's new National Leadership Computing Facility, which houses a large Cray X1, with plans to move to a 20-TF Cray X1E and 20-TF Cray Red Storm in 2004 and build to 250 TF by 2007. Priorities: nanoscale science, taming the microbial world, health and environment, ITER for fusion energy, the search for the genesis (astrophysics). He also described the Computational Biology Institute. The big challenge is finding and hiring top talent. ORNL expects to hire 45 in the next year.

Eric Stahlberg, Ohio Supercomputer Center, summarized OSC's life sciences research. OSC has Intel clusters, Crays, Suns, an Apple cluster. Cray X1 is a new addition, and Cray XD1 is one of the platforms OSC will use to develop FPGA algorithms. The center will hold a symposium on FPGAs Oct. 4-5, 2004, with Xilinx, Starbridge and Cray. The State of Ohio is interconnecting its many medical institutions, including OSC's Center for Computational Medicine.

Jack Collins updated attendees on the NCI's Advanced Biomedical Computing Center (http://www.abcc.ncifcrf.gov). He stressed the need to have the right applications available on omputers, and human experts to help researchers use the computers. NCI and Starbridge Systems are collaborating on a chimp-human comparison. A Smith-Waterman for this problem requires ~5 Tops/sec sustained.

Steve Reinhardt said SGI is adding new capability to do Smith-Waterman sequence comparisons rapidly by coupling FPGAs into the fabric. Results: 10- 100x speedups. Another innovation, Star-P (developed at MIT) enables parallel execution of Matlab programs and large speedups of algorithms.

Igor Tsigelny, San Diego Supercomputer Center, discussed protein structure prediction and a future automated system for drug design. SDSC uses Sun systems and plans to move to an IBM terascale system.

A panel moderated by IDC's Michael Swenson explored the fit between HPC and the bio/life sciences. Users were followed by vendors.

OSC's Eric Stahlberg said, "Irrational drug design is where things are today…science isn't all that important." He said simulation and systems modeling are good roles for HPC after a drug has been designed.

Jack Collins, NCI, said the migration of PC software to clusters and SMPs will allow news questions to be addressed routinely. He stressed the need for new software able to handle meta-data, and for better cluster bandwidth.

IBM's Hershel Weintraub stressed the need for more real-time analysis of data, "so we can modify trials based on early results and not wait six months," and for better interaction between IT and R&D.

Michael Banda, Berkeley Lab, said HPC systems have been designed for transaction processing, not scientific applications. Architectures are needed that achieve high sustained performance on a broad range of scientific applications.

Bobb Rick, Platform Computing, talk about the need to help users identify bad drug candidates early on and to get people to accept the new wave of computing, such as grid computing.

SGI's Dan Stevens said, "If we all use the same tools, we should all expect to get the same answers" and underscored the importance of applying the right kind of HPC system to specific problems.

Lionel Binns, HP, said blade technologies will make a huge difference, and better algorithms and codes are needed. Comparative genomics is not progressing as fast as expected.

John McAlpin said IBM is offering multiple types of systems to meet varying needs and has sold "a few Blue Gene/Light machines, which are somewhat customized." Power6 design is well under way with 65 nm lithography. IBM's goal: sustained PF performance in the 2010 timeframe.

Following the panel discussion Chris Hipp presented Orion Multisystems' high- performance cluster workstation products, including a 12-node desktop cluster (36 GF) for under $10,000 and a 96-node deskside cluster (300 GF) for under $100,000. The systems use the Transmeta Efficeon processor.

Scott McClellan and Lionel Binns gave the HP life sciences update. Interestingly, HP Storage Grid technology is based on "smart cells" to which unique traits can be added to enable specialized services. This gives advantages of scale- up and scale-out together.

David Bruce, Rikk Crill talked about Look Dynamics' (www.lookdynamics.com) digital-analog system, due out by mid-2005, that will optically encode shapes from images and then search databases for similarity, without requiring prior knowledge of the desired objects.

Nathan Wichmann, Cray, compared results for various HPC systems on the new HPC Challenge benchmark, versus Linpack alone, and stressed that HPCC is a substantially better predictor of real-world performance.

Dave Nelson reported that the federal HECRTF program "is now out of business, succeeded by the implementation phase." The 2006 budget is the first real opportunity for increased funding. The OSTP-OMB memo on FY06 research priorities (http:www.ostp.gov/html/m04-23.pdf) gives HEC high Administration priority. (Proto) Leadership-class systems include NASA Project Columbia (with SGI and Intel) and the DOE Leadership-class system at ORNL (mainly Cray).

Paul Muzio, chairman of the HPC User Forum Steering Committee, said that in his opinion, a major problem with HPC is the lack of production-quality software. Application software has a life expectancy of 20-30 years; hardware architectures are changing much more rapidly. Economic incentives are needed to encourage the development of production-quality software, and it needs to become easier for scientists to develop software themselves.

A second panel, introduced by Muzio and moderated by MITRE's David Koester, examined future programming models.

Koester said HPCS is aiming for high performance, high level languages. Today's languages require much work from users. If your architecture is limited to only reliable message transfer, it's limited in the programming models you can support. Portability should be relatively easy so you can support all the models. HPCS is developing a new procurement model based on productivity, defined as utility/cost. If it takes too long to do weather prediction, the value drops to zero.

Steve Finn, DigitalNet, reviewed problems with MPI and current language models and stressed the need to protect software investments and for software portability.

Hans Zima, Jet Propulsion Laboratory, Caltech and University of Vienna, said today's languages and models are inadequate. Petascale architectures pose new problems but may provide better support for high level languages and models. It will be essential to pursue different paths of research, including general- purpose languages and domain-specific frameworks. A key role will be played by intelligent tools supporting fault tolerance, performance tuning and autonomous operation.

Bob Numrich, University of Minnesota Supercomputing Institute and Goddard Space Flight Center, discussed the benefits of Co-array Fortran. He said we know how to build a very high performance machine that's easy to program that none of you can afford. We can also build cheap machines with high performance that are not easy to use. We can't do both.

Katherine Yelick, University of California/Berkeley and Lawrence Berkeley National Lab, said the real question is the difficulty of compilation and programming. You have challenging applications vs. challenging hardware (esp. unpredictable memory systems). You can’t get around "bad" (low bandwidth) machines.

Kevin Harris said HP is committed to porting UPC to all significant systems, with RDMA on shared memory interconnects. Also committed to optimization, both compile time and run-time. He added that the business model for programming tools within the industry is broken.

John McAlpin said the UPC effort is nearly complete using IBM compiler technology. As part of HPCS, IBM is working on other Stream-oriented languages. IBM has an aggressive Java-based program with higher-level functionality for HPC. He said, "customers are buying dirt-cheap hardware and still expecting free software."

Vito Bongiorno reported that Cray is supporting Co-array Fortran and UPC, as well as standard MPI. Customers and ISVs demand portability.

Albert Reuther, MIT Lincoln Laboratory, discussed pMATLAB (parallel Matlab) projects, such as flying a ruggedized Linux cluster on a Boeing 707 that the lab owns. The lab is headed toward petascale Matlab.

Cray's Bongiorno reported early performance data for the Cray XD1 (formerly OctigaBay 12K) and Red Storm products. Cray XD1 MPI latency is 1.8 usec, 4x lower than Infiniband. MPI bandwidth is 2x Infiniband. Cray Red Storm has 2 usec latency for nearest neighbor communications and 5 usec across the whole 10,000-processor Sandia system.

PERFORMANCE MODELING AND BENCHMARK ACTIVITIES UPDATES:

Leonid Oliker, LBNL/NERSC, previewed his SC2004 paper. Modern parallel vectors achieve 40-50% of peak, vs. <10% for superscalar architectures. He reported performance comparisons of the Earth Simulator and various U.S. supercomputers on four different applications and noted that codes are running faster than ever before on vector architectures. His team will return October to the Earth Simulator to evaluate new codes and higher scalability studies.

Laura Carrington said SDSC's automated framework predicted performance Cobalt60 on three IBM machines with accuracy within 11-20% and a sensitivity study showed that the biggest win would be to increase L2 cache. She gave other impressive examples.

Earl Joseph reported that IDC has multiple benchmarking activities, including the IDC Balanced Rating, and is involved in a new application benchmark study that will take a set of real-world applications, not application kernels, run them on the largest machines in world, and make results public via press releases and on the website. Will include the Earth Simulator, large ASCI machines and others.

Olaf Storaasli, NASA Langley Research Center, discussed reconfigurable computing using FPGAs. His partnership with Starbridge Systems has developed multiple algorithms for FPGAs. FPGAs will figure prominently in missions to the moon, Mars and beyond.

NASA's Bryan Biegel talked about "Kalpana," the first 512-processor SGI Altix system. In 30 days, it went from 0 to 60 TF. His facility also has a small Cray X1.

Robert Graybill, DARPA IPTO, reported on "cognitive systems," that is, systems that know what they're doing. Cognitive computers can reason, learn, explain and respond robustly to surprise. This is an 8-year program. A major challenge: what are the algorithms?

On January 27, 2005, the HPC User Forum will hold a meeting at the Earth Simulator Center in Yokohama, Japan. The next semi-annual User Forum meeting will be held in April, 2005 in Sundance, Utah. Major focus areas will include computational fluid dynamics, continued examination of programming models and benchmarking activities, and the future technical agenda for the User Forum. As always, there will be updates on a broad spectrum of other HPC disciplines and activities.

For more information about the HPC User Forum or to register for meetings, go to http://www.hpcuserforum.com.