Features:

HPC USER FORUM HOLDS THIRD U.S.-EUROPE MEETING AT CERN

Senior representatives from leading user and vendor organizations attended the HPC User Forum meeting at CERN last week and heard presentations on new developments and major issues from CERN, Ohio Supercomputer Center, SARA, Boeing, Pittsburgh Supercomputing Center, Centro Svizzero di Calcolo Scientifico (CSCS) and CINECA. HPC User Forum 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, summarized U.S. meetings held during 2004 and IDC gave an update on the HPC market. Earl Joseph, HPC User Forum Executive Director, invited everyone to join the next meeting in Japan at the Earth Simulator on January 27th.

The session at CERN was the third annual U.S.- European dialogue meeting held by the HPC User Forum, which was founded four years ago to promote the health of the global HPC industry and address issues of common concern to users. The organization has grown to 150 members worldwide. It is directed by a steering committee of users from government, industry and academia, and operated for the users by market analyst firm IDC.

The next U.S. User Forum meeting will be held April 19-21, 2004 in Sundance, Utah. The first U.S.-Japan dialogue meeting is slated for January 27, 2005 at the Earth Simulator Center in Yokohama. The next U.S.-Europe meeting will be held in May 2005 at Warsaw University's Interdisciplinary Centre for Mathematical and Computational Modeling (ICM). The fall 2005 U.S. meeting will happen September 26-28 in Richmond, Virginia. For more information about the HPC User Forum and its activities, go to http://www.hpcuserforum.com.

IDC's Addison Snell reported that in the first half of 2004, HPC revenue grew 39% over the same period in 2003, to $3.37 billion (6-month figure). Unit sales were 70,190, up 98.1%. Average selling price (ASP) was $40,100, down 29.8%. If the second half of 2004 merely matches growth for the prior-year second half, the HPC market will grow by an impressive $1 billion in revenue during 2004. Clusters are growing steadily and now represent over 30% of the HPC market.

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.

Markus Schulz reviewed developments at CERN, especially the LCG-2 project, whose aim is to deploy and operate the computing environment needed to analyze experimental data from the Large Hadron Collider. Schulz said LCG-2 has demonstrated that more than 80 sites can be integrated into a working Grid. The biggest issues are related to stable operation, and future middleware has to take this into account. The next step: proving that the systems can be scaled to handle the expected load.

Paul Buerger, Ohio Supercomputer Center (OSC), described the "Cluster Ohio" project that networks Ohio universities together and said the new OSC- Springfield facility will house a Cray X1, Cray XD1 and an Xserve G5 cluster. OSC held a successful FPGA symposium in early October and will completely replace its storage environment this year to support medical imaging needs (e.g., digitized pathology slides); high-energy physics; and subsurface geophysics for applications such as pollution transport and oil/gas exploration.

Sun Microsystems' Tony Kay discussed Grid trends. Noting that it's "hard to get the utilization you thought you would out of clusters," he said Sun sees a move toward holistic processor designs with moderate pipelines, reasonable frequencies, realistic clock speeds, good I/O into the processors, low memory latencies (NUMA) and reasonable power draw. Nodes will get bigger. Infiniband allows a single fabric but is not mature yet. Grids will move deeper into the data center. On-demand, utility computing has arrived. Sun has a simple model: $1 per CPU hour.

Axel Berg described SARA's National e-Science Support Center, the Dutch National Supercomputer and other programs and projects. Surfnet6 is an advanced research network in the Netherlands, Netherlight an optical Internet exchange for NRENs in Europe and beyond. The NL-Grid's goal is to strengthen the Grid infrastructure in the Netherlands. The mission of the Virtual Laboratory for e-Science: boost e-Science by creating an appropriate environment and doing research on methodologies. LOFAR (low frequency array) is the world's largest radio telescope. The role of EGEE (Enabling Grids for e-Science in Europe) is self-evident. AlmereGrid is a fiber-to-the-home project for this city of 160,000 near Amsterdam. Finally, DEISA (Distributed European Infrastructure for Supercomputing Applications) is a capability- computing Grid.

AMD's Francesco Torricelli said AMD has a direct connect architecture and, from the beginning, Opteron's design was done with dual- and multi-core in a single die in mind. In August 2004, AMD demonstrated its first dual-core capability. The AMD ecosystem includes compilers (Pathscale, GCC, Absoft); libraries (NAG ACML); and the AMD WW code division. AMD HPC partners and products: IBM, HP, Sun, Cray.

Paul Muzio summarized 2004 HPC User Forum meetings held in the U.S. The April meeting in Dearborn, Michigan focused heavily on structural analysis applications in HPC, with representatives from the user community, ISVs, vendors. The user community needs advanced multi-physics, multi-scale applications to support the vehicle design, manufacturing and construction industries. The commoditization of HPC has resulted in less investment from hardware manufacturers in ISVs, and industrial users have less budget for internal software development. There's a lack of available HPC platforms on which ISVs can do software development. Finally, the proliferation of commodity systems has resulted in many types of compilers, interconnects and processors, and it is hard for ISVs to support all of these. People need to recognize that hardware costs are a small portion of total development costs in high performance computers. Another dilemma: With the HPC workforce aging and an increasing reliance on COTS software, there is a loss of knowledge about what is going on in the programs.

Muzio said an important focus of the September meeting in Tucson, Arizona, was programming models and languages. MPI is ubiquitous for parallel codes but can be difficult to use, has different versions for each interconnect and has problems with fault tolerance, I/O speed, portability and robustness. But new languages need to be supported on a wide range of systems to gain popularity. OpenMP has limited scalability and Co-array Fortran has a problem with portability. There is a lack of production-quality application software across the board. Life expectancy for software is 20-30 years, whereas for hardware architectures it's 3-4 years at most. As a result, the available applications don't take advantage of the latest hardware capabilities. It's important to protect investments in existing software and user knowledge. We need to recognize that HPC is a strategic investment and application software is key to healthy HPC. The HPC industry needs objectives-based investment incentives. He summarized the aims of the DARPA HPCS program regarding programming models, noting that models currently are orthogonal to performance goals.. Higher- level languages that are easy to use generally have poor performance. Conversely, lower-level program tools that provide high performance are not easy to use. HPCS wants something in the middle. Muzio said that global programming models such as UPC, Co-array Fortran can provide higher performance with about 10% of the coding effort required with MPI.

Suresh Shukla, Boeing, said the HPC industry is focusing more on price and cost than on capability and cost-benefit. Boeing understands the benefits of HPC and in the future needs balanced systems with latency and bandwidth improvements of at least two orders of magnitude; efficient implementation of parallelism in a reliable manner; applications able to exploit hardware enhancements; and reasonable quality enhancements/cost to justify replacing experimentation with simulation. Boeing's Cray X1 is doing work, such as on new plane design, in three hours that took three days before. By saying commodity machines are good enough, he said, we are hurting ourselves, our companies, and the state of research and science. User requirements should drive buying decisions.

Luigi Brochard said IBM is introducing a scalable SMP, 2- to 16-way and Xeon clusters, scalable to 16-way, with a common software architecture for all these systems. He said Power4 was the first dual-core chip. Blue Gene is about putting more density, performance in a rack. The basic building block is an ASIC. From a programming standpoint, it's an SPMD model, i.e., a general MPI machine. IBM is currently porting libraries and real applications (life sciences, physics, earth sciences, CAE).

Marie-Christine Sawley reported that the Centro Svizzero di Calcolo Scientifico (CSCS) is working on a new organization this year. CSCS was an administrative unit of ETH-Zurich and is more autonomous now. CSCS has a new partnership model for development, education and new technologies. Examples: CSCS-EPFL-VATECH collaboration to simulate water flow inside a turbine; high- resolution weather forecasting; visualization of DNA molecule. The Horizon Project pairs CSCS and Paul Scherrer Institute to deploy a self-made 10- teraflop high-productivity system. CSCS is also involved in the Swiss BioGrid project, with multiple Swiss partner institutions, to evaluate the actual capacity of distributed Grid for computation. Target problem: world distribution of dengue fever.

Jim Kasdorf said the Pittsburgh Supercomputing Center is a technology user and producer, and has an educational role, including helping industry with HPCC. He reviewed Sangtae Kim's Tucson presentation, "The NSF Perspective." NSF supports all science and engineering, has a continuing commitment to high end computing (HEC), and is balancing capability and capacity through programs such as the Extensible Terascale Facility (ETF). PSC's current terascale computing system, LeMieux, has 6 TF peak performance and has sustained more than 4.5 TF on real applications. Kasdorf said a strong interconnect is important for scientific end users. PSC has an agreement with Cray for a Red Storm-based product. Once again, PSC is spending more money for a better- balanced system with a strong interconnect. This system is a prototype for a leadership-class computer.

According to Winston Prather, trends Hewlett Packard sees include industry- standard servers, clustered computing, Linux and open source. These trends have led to 7x price/performance increases in the past 2-3 years and have opened HPC to customers who couldn't afford it before.

The leading architecture in the Top500 is x86. Problems with clusters: management, scalability, integration of data management and visualization, interconnect/network complexity, application availability. "It's not easy to make it all work together," he said. Data center thermal management is also an issue. HP is partnering with the U.S. Government to build a Lustre file server.

Giovanni Erbacci said CINECA has a new service-oriented laboratory infrastructure, consisting of HPC and Grid elements, to provide better access to resources and services. The HPC domain is increasingly integrated into the mainstream of information computing technology (ICT) and will drive many aspects of ICT. CINECA believes it is crucial to integrate compute, data and visualization. Resources: SP4 Regatta with 512 CPUs, IBM Linux cluster, SGI Altix with 12 graphic pipelets, SAN, SEC, virtual theater. CINECA is part of HPC-Europa, the EU-funded project aimed at providing transnational access to HPC systems. Other participants: EPCC, HLRSM, IDRIS, SARA, PSNC, PARALLAB, TCD, CASPUR, NTUA. 800 researchers will benefit from access visits in the period 2004-7.

Karl Solchenbach said part of Intel's rationale for being in HPC is that rapid technology adoption in this market provides experience that is also valuable for the commercial market. Standard Intel tools are used in HPC: compilers, libraries, VTune. The company has established a new division for Parallel and Distributed Solutions, and is continuing development of threading and cluster tools as Intel software products. Intel also has an Advanced Computer Center, and Intel's UNICORE HPC Grid system, a Grid programming environment, is used at centers around the world.