Features:

HPC USER FORUM EXPLORES STRUCTURAL ANALYSIS, NETWORKING

A record 117 people attended the semi-annual HPC User Forum meeting held this week. The major focus of the two-day Dearborn meeting was on user requirements for structural analysis, interconnects and I/O, according to Steering Committee Chairman Paul Muzio, VP-Government Programs for Network Computing Systems, Inc. and Support Infrastructure Director of the Army High Performance Computing Research Center. The group also discussed the future technical agenda for the HPC User Forum.

Muzio said the Dearborn meeting was to discuss ways to enhance the product development process and quality in the automotive, aerospace and construction industries through the use of HPC. He noted that this meeting would try out a new format, a sequence of panel discussions on structural analysis software, and that if members liked the format it would be applied to other topics in the future. He thanked Larry Davis, of the DOD HPC Modernization Program, for his outstanding performance as steering committee chairman for the past two years, and also thanked IDC for their strong contribution.

Earl Joseph, IDC, who serves as executive director for the HPC User Forum, gave an HPC market update and reported that the HPC market grew 12% in 2003, to $5.3 billion, following two years of decline. IDC forecasts a 6.6% compounded annual growth rate (CAGR) through 2008, to $7.6 billion. Commodity processors and systems now dominate, with clusters disrupting traditional business models. The biggest growth areas are clusters and bio-life sciences, although the latter is expected to grow at lower rates than in previous estimates. Joseph said IDC is conducting a study on HPC applications benchmarks using the world's largest HPC systems, a multi-client study on the cluster and grid computing markets, and three separate studies on the use of HPC in industry.

Joseph encouraged attendees to participate in the Gt'04 grid conference, May 24-26 in Philadelphia, and the ISC2004 conference, June 22-25 in Heidelberg.

Vernon Turner, IDC group vice president for Global Enterprise Server Solutions and WHPS, reiterated IDC's strong commitment to the HPC market and introduced IDC's expanded team covering Workstations and High Performance Computing Systems. Turner noted that this year is the 40th anniversary of the CDC 6600, which boasted peak performance of nine megaflops and a 130-kilobyte memory.

The series of four panel discussions focused on current and emerging requirements for HPC structural analysis capabilities. The first panel, moderated by Ford Motor Company's Vince Scarafino, included officials from DaimlerChrysler, Ford, General Motors, NASA Ames Research Center and the National Institute of Science and Technology (NIST). Key issues identified by these users were the need to reduce model sizes because of today's limited computational and data storage capabilities; the fact that commercial software packages are not written to be coupled efficiently with each other for multidisciplinary analysis; bandwidth (compute, I/O, data management); lack of adherence to industry standards, especially for new and exotic systems; too many operating systems and OS "flavors"; power, heat dissipation and footprints; robustness of systems and models.

The second panel on this topic, moderated by David Koester of MITRE, consisted of representatives from software vendors MSC.Software, ANSYS, ESI, Abaqus and MCube. The issues they highlighted from the ISV perspective were the proliferation of hardware platforms and operating systems needing to be supported, including multiple Linux versions; the related need to choose between deploying finite human resources to support more features or more platforms; pressure by users to produce unified models (e.g., crash and NVH); the need to support different versions of MPI libraries; the difficulty of acquiring large versions of HPC systems for on-site porting work; the disparity between the scientific research community's focus on scaling problem size and industry's focus on reducing runtimes; and algorithm development and choices.

Panel three in the series, moderated by Earl Joseph, presented the hardware vendors' perspectives and included officials from IBM, Cray, Hewlett-Packard, Intel, NEC, SGI and Sun Microsystems. Key issues identified by the panelists included system balance; disk reliability; inadequate I/O performance; inefficient solvers and domain decomposition techniques; the need for better interaction between applications from different software vendors; the need for rewrites of major codes; the need for benchmarks based on actual workloads; the struggle to articulate ROI from HPC; too many Linux versions; multiple interconnect choices; and data storage.

David Koester, MITRE, gave a talk on evolving programming standards for HPC. Noting that parallel programming is difficult and expensive today, he said that the partitioned global address space (PGAS) model, represented in languages such as Co-Array Fortran (CAF), Unified Parallel C (UPC) and Java- based Titanium, is more economical and allows programmers to improve performance by understanding where programs reside and doing algorithms appropriately. This evolving distributed shared-memory programming model will match any architecture and is compatible with mixed models. Koester said the new programming language, called Chapel, that the Cray team is developing under the DARPA HPCS program is intended to simplify the creation of new parallel programs, while the goal of IBM's DARPA HPCS model, PERCS, is to support various languages: C/C++, Fortran, Java, UPC, Matlab and Streamit.

Gordon Johnson and Andrew Johnson of NCSI and AHPCRC, discussed the parallel implementation of EPIC, a Lagrangian hydrocode for large deformation contact- impact studies. In the first year of a planned three-year effort to parallelize and modernize the 200,000-line EPIC code, the researchers have developed a generalized particle algorithm and produced promising early results on what will become the future production version of EPIC.

Olaf Storaasli, NASA Langley Research Center, said a revolution is brewing in HPC, with new types of machines and new solution methods such as Stephen Wolfram's cellular automata. NASA-wide HPC systems, he reported, include a 128-processor SGI Altix and a 64-processor Cray X1 system now on order. NASA Langley itself now has 1200-1500 cluster nodes, representing a 50% annual increase. Langley also is partnering with Starbridge Systems and others to evaluate and enhance FPGA-based "hypercomputers" for NASA use.

Eric Pitcher, Linux Networx, discussed a cooperative project with MSC.Software called "Robust Engineering." The goal is to develop a stochastic approach to simulation whose outcomes are largely predictable (high correlation between input and output). The approach incorporates variability into MSC.Nastran. Its benefits are reduced cost and improved product quality, by enabling engineers to better model reality and identify risks.

George Cotter, DOD, chaired a panel discussion on the technical agenda for the HPC User Forum. Members of the technical agenda committee served as panelists. Referring to the white paper handout, Cotter noted that the technical committee had focused on four major technical areas identified as important in a recent HPC User Forum survey of the HPC community: architecture, system balance, programming models and environments, and software tools and ease of programming.

Cotter covered architectural issues, including global memory access and memory bandwidth, requirements for leadership-class systems (50+ TF), influencing architectural directions, implementing new technologies, I/O systems matching processor performance, user management of large-scale systems, and engaging systems designers and developers.

Alex Akkerman, Ford, noted that most systems available in the market today don't deliver a high percentage of their theoretical peak performance and discussed issues related to system balance: memory bandwidth and communications latency, the challenge of selecting a suite of applications that illustrate system balance well, implication for leadership-class HPC systems, and the unique balance issue related to highly distributed systems such as grids.

Brad Blasing, NCSI, reported on issues related to programming models and environments. These include the fact that MPI, the de facto standard programming language, has poor scalability for leadership-class systems, but better-scaling languages such as CAF and UPC are limited and poorly supported today. Innovative new architectures, such as FPGAs, often require new programming models that initially suffer from weak applications relevancy. There are major uncertainties in the UNIX-Linux evolution, and there are disconnects between applications and new systems development.

Blasing also covered issues having to do with software tools and ease of programming, noting that adequate tools are lacking to understand how the system is functioning, and tools needed to support effective programming and code portability are seldom sufficient on newer systems. There is also a need for better instrumentation of systems to permit accurate performance monitoring and measurement for buying decisions.

Earl Joseph led a panel discussion on interconnects and I/O. In this format, users first stated their issues, I/O vendors responded, and then the users commented on the vendor responses. Key issues mentioned by the panel of users included the near-term need for faster bus speeds, support for larger file systems (100 million files), and faster switch hardware. I/O vendors Infinicon, Topspin, Meiosys, Voltaire, Scali, Mellanox, IBM and SGI presented their current products and future plans for interconnects and I/O (in the case of Meiosys, this meant checkpoint/restart).

Penguin Computing presented a company update, highlighting its recent acquisition of Scyld Software. Products include workstations, rack-mounted servers and clusters. The company says its basic advantage is an OS built from the ground up with clusters in mind.

Peter Ungaro gave the Cray update, reviewing the expanded product portfolio that will include the existing Cray X1 system, and the AMD Opteron-based Red Storm and Cray XD1 product lines that will be introduced later this year. Ungaro said the Red Storm product's mesh interconnect will move to a 3D torus configuration, and that Cray is rebuilding its CAE team.

Earl Dodd updated attendees on IBM's "Deep View" visualization activities and plans, saying that IBM intends to deliver in 2004 a best-of-breed visual product lineup based on commodity technologies. He said Deep View is a virtualized software layer and employs two functional capabilities: Remote Visual Networking (RVN) and Scalable SW Concentrator (SSC).

Ed Turkel gave the Hewlett-Packard update. In February 2004, HP announced 3 Opteron server and blade products in the ProLiant family. ProLiant DL145 is a 1U2P server for HPTC. Also for use in large clusters. HP is bringing the Opteron products into its overall Linux cluster strategy and is also doing visualization clusters.

Cornell Theory Center's David Lifka talked about its Windows High Performance Computing Center. CTC switched from UNIX to Windows seven years ago and the switch has been successful. CTC is finding interest in this model from Wall Street firms, bioinformatics organizations and the media industry.

The fourth and final panel on structural analysis, moderated by Douglas Templeton of the US Army RDE Command, included representatives from Ford, Hewlett-Packard, ABAQUS, General Motors and AHPCRC. Key issues forwarded by the user, ISV and hardware representatives were the need for advanced multi- physics, multi-scale applications; reduced time to solution; the ability to handle larger numbers of jobs; simplified programming models and environments; and transfer of technology to new engineers. Muzio said the issues slides will be circulated more broadly for comments.

The HPC User Forum is directed by a steering committee consisting of users from government, industry and academia, and is operated for the users by market analyst firm IDC. HPC User Forum meetings are co-sponsored by HPCwire.

In July 2004, the HPC User Forum will hold a meeting at the Earth Simulator Center in Yokohama, Japan. On October 13, 2004, the HPC User Forum will hold another round of U.S.-European dialogue meetings at CERN in Switzerland.

The next semi-annual User Forum meeting will be held September 20-22, 2004 in Tucson, Arizona and will focus on bio/life sciences and government uses of HPC, with step two of the technical agenda planning and reports and updates in a broad spectrum of other HPC activities. For more information about the HPC User Forum and its activities, go to http://www.idc.com/hpc.