Features:

SGI CEO COMMENTS ON HPC'S IMPACT ON NATION

For years the United States has been the home of the fastest computer in the world - now Japan is ahead. The House Science Committee conducted a hearing to examine federal high-performance computing research and development activities. As a result, the committee has released the High-Performance Computing Revitalization Act of 2004. Supporting the initiative were industry experts including SGI's CEO, Bob Bishop. Below you will find his testimonial for additional background on the impact of high-performance computing on the nation. Summing it up, Bob clearly states "To out-compete in the 21st century, the U.S. industry must out-compute."

Written testimony submitted to Chairman Sherwood L. Boehlert regarding the proposed High Performance Computing Revitalization Act of 2004.

HOW DO U.S. COMPANIES DEPLOY HIGH-PERFORMANCE COMPUTING AND HOW DOES IT AFFECT U.S. INDUSTRIAL COMPETITIVENESS?

The role of HPC in US industry today is to solve technical problems quickly, gain insight into design alternatives, and bring safe and secure products and services to the market early, thus creating competitive advantage and improving the quality of our daily lives. HPC stimulates global competition, then helps companies compete in fiercely competitive markets.

HPC can also be seen as a nerve center within the corporate setting and a conduit to cross-functional thinking. It brings together specialists from different fields who, by interaction with each other, rapidly improve their understanding, insight and problem solving in matters of great complexity. HPC eliminates stovepipe thinking.

Managers and specialists leverage each other's knowledge in such an environment, asking multiple "what if" questions, evaluating countless scenarios while accelerating cooperative decision-making along the way. As a consequence, enterprise level strategy and tactics are broadened and strengthened.

HPC in US industry today is not compute-only activity conducted in glass-house isolation. HPC centers are connected via high-speed lines to other geographically dispersed decision centers both inside and outside of the enterprise. HPC may also direct-connect with laboratory instruments, sensor networks, satellite feeds or real-time video signals. In fact, it is increasingly common to find rich media from multiple sources "fused" into a single image, overlaying locally generated graphics, and effectively granting "X-ray vision" to all participants in the HPC session. In this way, HPC becomes a tool for superior decision making.

Increasingly, HPC drives a creative food chain, from innovation to operations, and increasingly delivers interactive real-time solutions. Speed and innovation are critical in the corporate race for global success.

Leading US industries have aggressively adopted HPC to improve their productivity and competitiveness. Defense, aerospace, automobile, chemical, pharmaceutical, medical, energy and media lead the way. Other US industries are adopting HPC at a more modest rate.

Worldwide deployment of HPC is found in similar industry sectors, especially in Japan, Germany, France, and the UK. China and India are beginning to rapidly adopt HPC as well. The US remains by far the most predominant supplier of HPC products however, for both hardware and software. Japan is the only other significant HPC equipment maker.

Leading-edge developments and break-through ideas in modern industries require high levels of modeling, simulation, visualization, and life-cycle data management. Biotechnology, nanotechnology and material science, for example. Vast amounts of intellectual property and future wealth are created in the process. Competitors strive to out-gun each other with in-house HPC capability, and win the right to patent, copyright and trade mark their knowledge.

HPC must be understood however, not as a single technology, but as an ecosystem of multiple technologies, each with its own set of issues and challenges: fast processors, complex memory hierarchies, interconnect fabrics, massive storage facilities, high-fidelity visualization, networking, and multi-layered software, to name but a few. A single weak factor will likely reduce the overall effectiveness of any HPC installation, dramatically.

HPC buyers must judiciously balance and combine HPC sub-system technologies appropriate to the real-world problems that they are attempting to solve. Even then, buyers need to continuously stay abreast of updates and developments, keeping their facilities relevant and at the leading edge, if they wish to survive.

Attracting and retaining talent to run an effective HPC facility is difficult for most US corporations, especially with the recent dearth of computer science graduates emerging from advanced engineering schools.

Perhaps this is because corporate IT spending in the recent past has been dominated by business-process applications a la enterprise resource planning (ERP), customer relationship management (CRM), Internet deployment and mobile computing. Such applications have improved the background context in which all corporations must operate. However, spending in these areas has not helped the core HPC user, except in the few cases where commercial technologies can be successfully re-purposed within the HPC mission. For example, Internet technology is useful for everyone, technical and commercial, as is the PC, the PDA and the cell phone. These latter devices however, are mostly used as access mechanisms to remote HPC resources, and do not constitute HPC technology in its own right.

The annual spending of US corporations on business-process applications is one hundred-fold greater than that spent on engineering and scientific applications. With few exceptions, computer vendors are therefore attracted to the commercial side.

To help spread the adoption of HPC within US private industry more broadly, and to help ensure more US government and US industry interchange in the future evolution of this critical capability, the Washington DC based organization "Council on Competitiveness" has recently begun a High Performance Computing Initiative. I am privileged to serve on this Council's Executive Committee, and would encourage the chairman of the House Committee on Science and its members to be in contact with this effort. The Initiative is gathering data that will provide a timely and accurate profile of key HPC users, application areas and bottlenecks experienced in US industry today.

This data will also highlight the multitude of factors that determine private industry HPC deployment in the US, including application software availability, ease-of-use, total-cost-of-ownership for equipment and personnel, and return-on-investment to the buyer.

As for US computer vendors, in the absence of significant HPC volume procurements by corporations, it is difficult for them to focus solely on industry HPC markets. Hence US computer vendors generally concentrate their product developments on the larger business-process markets, positioning their HPC activities as a minor sideline. Alternatively, they will repackage their commercial machines for technical purposes. Neither approach however, will allow HPC to reach its full potential. The market requires US Government HPC procurement in steady volume to sustain strong US HPC capability. This U.S. Government additional volume is especially critical to the health and survival of the few computer vendors that remain alive and dedicated to HPC today.

WHAT ARE SOME OF THE CURRENT HPC EFFORTS OF THE FEDERAL CIVILIAN SCIENCE AGENCIES? ARE THEY SUFFICIENT TO ENSURE U.S. LEADERSHIP IN HPC?

Recent events have conspired to raise alarm that the US HPC industry has fallen behind its foreign rivals. For example, the powering on of Japan's Earth Simulator in March 2002, was a "Sputnik like event", overshadowing all HPC machines on the planet. As of today, this machine is still at the head of the Top 500 Supercomputing Sites, as last published in November 2003. The machine is optimized for geoscience applications, and is front-ended by three Onyx machines supplied by Silicon Graphics Inc (SGI) that convert its numerical output to interactive immersive high-fidelity visualization. You can't drink from a firehose!

The ES-40 (Earth Simulator-40 Teraflops) price-tag exceeded $300 million, excluding the elegant new buildings in which it is housed. It was paid for by the Japanese Government and built by NEC along the lines of its SX-6 machine, a clustered-vector architecture in its sixth generation.

This is an outstanding example of government-industry cooperation in open science, but not necessarily a good example of HPC innovation or good HPC architecture. It is certainly a shining example of what money can buy. However, the recently installed ALTIX supercomputers at Tokyo University's Earthquake Research Center run several earthquake codes at similar speeds as the ES-40 at a much lower price. The two ALTIX machines, supplied by SGI, incorporate the latest Itanium 2 processor technology from Intel, the Linux operating system, and SGI's fourth generation global shared memory NUMAflex architecture.

Within the US Government, the National Weather Service (NWS), the National Center for Atmospheric Research (NCAR), and the Geophysical Fluid Dynamics Laboratory (GFDL) belonging to National Oceanographic and Atmospheric Administration (NOAA) are already heavy HPC users. All of these centers however, would benefit greatly from additional HPC capability, given the importance of weather in our daily lives and given the difficulty of weather science. Severe weather continues to wreak havoc in many areas of the US, and the cost of more accurate weather modeling and forecasting capability pales in comparison to the damages caused by unforeseen weather events. The cost of hurricane evacuation alone on the Atlantic seaboard exceeds $1 million per coastal mile, or $100 million in the case of a hurricane that cannot be predicted to come ashore within one hundred miles. A 50% improvement in forecast accuracy would lower this cost by $50 million, provided it could be accomplished in a timely manner; enough to recover the cost of HPC equipment in a single event, and more importantly, saving lives along the way.

The key to solving problems in weather, climate and environmental science is HPC. Nature can only be accurately described and computed from equations that take account of complex non-linear interactions between multiple natural systems, i.e. rivers, lakes, oceans, mountains, forests, dust, pollution, cloud cover, snow cover, ice, polar regions, etc. Such equations of motion are so interconnected and intertwined that they can only be managed when all aspects are held in the global shared memory of a large HPC machine and computed simultaneously.

We have a similar experience at NASA's Goddard Space Flight Center and at the NASA Ames Research Center. Both are heavily committed to HPC and are driving their climate modeling programs to higher performance through extensive use of leading edge HPC. NASA Ames has in fact tuned their 512-pprocessor ALTIX machine to world record-breaking memory bandwidth performance (the first machine in HPC history to break one terabyte-per-second, as measured by the STREAMS Triad benchmark). Both NASA facilities will require much more HPC capability however, to achieve the Administration's recently announced Code T program consisting of a permanent Moon-colony and manned space flight to Mars. There is an opportunity here for NASA to build Moon and Mars simulators, along the lines of the Japanese Earth Simulator. Such simulators would be less difficult however, given that neither the Moon nor Mars has an active weather or tectonic system like the Earth.

There is also the need to design and simulate a new generation of spacecraft for the long voyages entailed. Moreover, since NASA's three space shuttles will most likely stop flying by the year 2010, the design of new generation space vehicles should begin very soon.

Human and Health Services (HHS) is yet another Federal civilian science agency that must strongly encourage the deployment of HPC. Rapid recognition of pathogens and viruses and the development of their counter-acting vaccines is critical to public health. The recent global outbreaks of SARS, Ebola, Avian flu, and West Nile disease maybe an indication of worse to come. Rapid government response will only be achieved through HPC centers and laboratories that are globally connected.

Bio-terrorism is an additional threat for HHS to manage. Crisis management will ultimately require real-time modeling and simulation of toxin dispersion at the resolution of city streets and office buildings, at least in the top one hundred population centers of the US. These issues and others overlap with the newly formed Department of Homeland Security (DHS), which itself must become HPC capable to be fully effective.

The US Department of Energy (DOE) has extensive experience in HPC, although mostly for weapons design and nuclear stockpile stewardship. HPC deployment however, is recently gaining momentum within DOE's Open Science program, and this is a very encouraging trend for the US HPC community as a whole. DOE will play a critical role in guiding the nation's future energy infrastructure and building alternative energy technologies. It also has extensive experience with environmental remediation. These are grand challenge problems that require significant HPC resources.

Generally speaking, there is a clearer recognition across the Federal civilian agencies today that personal computers do not deliver the true horse-power of HPC machines, no matter how many units are networked together. One thousand bicycles do not make a truck! However, the low entry price of commodity clusters is often attractive for certain engineering and scientific applications, especially when these applications entail little inter- communications between the elements of the cluster. Even then, commodity clusters are only effective if there are no real-time interactivity requirements. Surprisingly however, the long term total cost of ownership of a commodity cluster can be higher than expected if the full cost of maintenance, software licensing and system administration is taken into account.

Finally, the recent formation of a High End Computing Revitalization Task Force (HECRTF) has been very helpful in building knowledge and momentum around the importance of HPC to both US industry and the US Federal Government. There is now a greater inter-agency discussion on the topic, and private industry is being heavily consulted. We are eagerly awaiting the outcome of this effort. Nothing will encourage more future spending by the US computer vendors on HPC research and development however, than a strong increase in US Federal HPC procurement and deployment.

SUMMARY OF SGI's HIGH-PERFORMANCE COMPUTING RESEARCH EFFORTS

SGI regularly spends 13% of its annual revenues on research and development. This entire amount is spent on high-performance computing, high-performance storage, and high-performance visualization.

SGI dedicates its R & D efforts to system-level architectures utilizing industry standard components where appropriate. The unique combination of system-level architectures built with standard high-volume off-the-shelf commodity components, yields an overall price/performance balance that is very attractive to the HPC user. Full-custom products are generally too expensive, and full-commodity products lack the required performance or productivity.

SGI is aggressively focused on the technical, engineering and scientific marketplace. Problems in this space require large numbers of processors, large amounts of memory, and large amounts of I/O bandwidth, all tightly coupled with each other. SGI servers scale-up the number of processors, the amount of memory and the level of I/O bandwidth independently. To date, SGI has shipped HPC machines with 1,024 processors and with 4 terabytes of globally shared main memory. Current R&D efforts within SGI are aimed at scaling systems to 128 thousand processors and to one petabyte of main memory, globally shared among all processors. This is an ultra-scale machine, and one that is within reach by SGI in the 2007-2008 time frame, in partnership with the appropriate funding agency.

Furthermore, it is SGI's intention to integrate scalar, vector, streaming, and special-function processors directly onto the shared memory architecture of this machine. The most appropriate processor elements will then be brought into action on-the-fly, while the user's application code is being executed. This "multi-paradigm" concept will therefore embrace the best features of several architectures that are in the marketplace today. The machine will reconfigure itself in a dynamic manner to best suit the application as it runs.

With respect to our R&D efforts in storage and data life-cycle management, CXFS from SGI is a very successful shared-file heterogeneous-connect storage area network (SAN) in the market today. It will be extended to run over a wide area network, and thus enable nationwide single-level file addressing (SAN over WAN).

With respect to SGI's R&D efforts in visualization, our work involves the interactive visualization of massive data sets stored in global shared memory, using the diverse compute elements of the multi-paradigm architecture. We will bring high-fidelity visualization to the Linux environment in the near future.

And with respect to SGI's R&D efforts in software, we will assist the Open- Source community scale its Linux-64 operating system to accommodate as large a number of processors in a single system image configuration as possible. We will also help bring high-level scientific programming tools into the market and application program interfaces (APIs) that improve the ease-of-use of HPC equipment in general.