Below is Christopher Lazou's interview with Prof. Dr.-Ing. Michael M. Resch, Director of the High Performance Computing Center in Stuttgart (HLRS), on his work there, as well as his general thoughts on the future of HPC in Europe and beyond.

Christopher Lazou: Michael, it's good that you can spare some of your valuable time to talk to me. Glancing at your job responsibilities, one can see you have a number of important roles. Let's briefly discuss each of the main ones and, in the process, try to gain an insight of your views concerning HPC and share them with the HPC community.

You are director of the High Performance Computing Center Stuttgart (HLRS). What is the role of this center?

Prof. Michael Resch: The High Performance Computing Center in Stuttgart – Hoechstleistungsrechenzentrum Stuttgart (HLRS) provides access to supercomputers and supports R&D users from Germany and Europe in the use of leading edge supercomputer technology and its applications. The mission of HLRS is to provide its users with tools and expertise to achieve top international positions in their research field. We want to deliver excellence, engagement, empowerment and entrepreneur culture. But quality has a price. Capabilities at the high end and economy of scale are possible through a joint operation of supercomputer systems with T-Systems, T-Systems solutions for Research GmbH and Porsche AG in the system management company hww GmbH.

In using supercomputers for research, our main motivation is that we want to use the fastest system available to get insights that are not possible with slower systems. We are willing (able) to spend the necessary funds to receive a much higher level of performance and be about 10 years ahead of standard systems.

Lazou: You are saying that HLRS not only provides computing services to university researchers for experimentation, but also provides computing services to industry whose interests are more geared to timely solutions and development of products achieving competitive edge in the market. Is there a mutual benefit in this arrangement?

Resch: Yes. There is a synergistic ethos created by the interaction of the two cultures. Academics become more professional aware of deadlines and industry gets research results earlier. The academic researchers get funds to pursue their research and industry gets knowledge transfer. Industry is happy to get tested solutions so long as these solutions are on time. This arrangement works very well as is evident from the fact that HLRS gets 20% of its funds from industry.

Lazou: HLRS operates a heterogeneous computing environment, but also has a long history of using NEC SX systems and your latest is a very large SX-8 system. Can you elaborate on why HLRS chose to continue with the vector parallel architecture, and how did you arrive at this heterogeneous computing environment?

Resch: HLRS has a long history in using the latest supercomputers. It now operates 6 nodes NEC SX-6 (48CPUs) and 72 nodes (576CPUs) NEC SX-8 system (12.67Teraflop/s peak) delivering 5 Teraflop/s, sustained performance. This is equivalent to over one quarter of the Earth Simulator. It also has a series of commodity-based systems as for example the 256 AMD-Opteron processor Cray Strider, an NEC TX7 system, and other systems used to support our file system, visualization, pre-processing and other elements of our operational environment.

We surveyed our users and arrived at a profile of their requirements. We then surveyed and benchmarked a number of available systems across the whole spectrum, from commodity clusters to vector parallel systems. We did not choose based on architecture. We collected a number of attributes in a basket and looked for a vendor with a strong roadmap containing these components, i.e. we wanted a strong partner. NEC has a strong record in microprocessors and that was very important to us. Our assessments showed that the NEC SX-8 was the most appropriate for our requirements and when Total Cost of Ownership (TCO) was factored in, was also the most cost effective for our work profile.

Lazou: We read a lot about commodity clusters being cheaper. Yet your findings that the SX-8 is not only a capability system, suitable for large-scale applications, but also competes on price/performance are very interesting. Can you elaborate a little more about the criteria of selection and the conclusions you arrived on price/performance for various classes of systems?

Resch: The main criteria for our investment included performance, reliability, power consumption, footprint floor space, maintenance, software availability and human resources (i.e. Total Cost of Ownership for delivering the service to users). Taking all these together, we found that for a very mixed user profile, vector systems and clusters of PCs are comparable in price/performance, where legacy 64 bit architectures are much more expensive. Given that I can achieve the same level of performance with significantly fewer processors the choice for an SX-8 hinged on the big benefits which come from less complexity in operating systems, maintenance and programming of applications.

Look at the recent developments including 2 cores on a chip. This increases CPU speed, but the memory gap gets worse. Our survey found that most users in HPC use 512 to 1000 CPUs and capability is the reserve of parallel vector systems, while throughput can be achieved using clusters.

Lazou: The NEC SX-8 has been in operation at HLRS for several months. Can you briefly say whether reality matches the measured performance of the benchmarks during the procurement assessment?

Resch: It actually exceeds our expectations. For nearly every benchmark criterion, we see substantially better performance. When we look at application performance, for one application we expected 4 Tflop/s and we actually got 5 Tflop/s. NEC installed the system and after thorough testing on-site we started operations. In our experience, the NEC SX-8 system has an incredibly high level of reliability, considering our configuration size is one of a kind.

Lazou: I know that HLRS has a long association with industrial companies mentioned above, networked with very high speed links for joint operation of supercomputer systems to achieve economies of scale and capabilities at the high-end. Is productivity a major factor in your decision to stay with parallel vector architecture systems?

Resch: Productivity is a major factor for the whole center. The vector SX-8 is a major element, but to maintain productivity across the whole spectrum of our workload profile we also use clusters of PCs. We use the right architecture for the right task.

Lazou: Let's move to another of your roles: chair for High Performance Computing at the University of Stuttgart. Can you share your vision of what are the major challenges for academics in the HPC field, for example in training new researchers for industry?

Resch: People in Western countries lost interest and enthusiasm in technology in general. I teach computer science and all my students have PCs, but hardly any of them are interested in how the hardware or the software work. Very few ever wrote a program. Only students who come from outside Europe show interest in these technologies. This is reflected in manufacturing, moving away from developed countries to China, India and the rest of Asia. Industry wants people who understand the technology, know the computer and have the specific skills to be productive. The good thing for us at Stuttgart is that we have strong collaboration with industry. Our students have to write 3 theses during their course, one long one requiring several months of work. This is done in collaboration with industry. This helps to make them aware of industrial needs.

Lazou: I understand one of your research interests is in numerical simulation of flow phenomena, in particular blood flow simulation. Can you elaborate a little, especially on the computational requirements and how the NEC SX-8 computing capability enables this type of simulation to be performed?

Resch: The simulation of blood flow is an extremely intensive computation. It contains multi-physics – CFD for flow, structural mechanics of moving artery walls plus biochemistry. The computational resources and the performance requirements are very large. We at Stuttgart are collaborating with researchers from Munich University, using the enormous power of the SX-8 to perform large simulations, which would not have been possible on smaller systems. We are convinced that with the NEC SX-8 capability, we will get a much better understanding of the dynamic behavior of large arteries. This for me is a long-term research interest, and I believe it will also become very important to the health industry.

Lazou: You also list as one of your research interests, the Grid. I have recently attended a VIP conference in London on the state of the Grid and found that there is still a long way before we can have a mature usable Grid infrastructure in place. The problems to be solved are not just technical, such as robust middleware, but also include ISV software licensing, cross national taxation mechanisms and ethical legal issues, especially in health applications. What is your view?

Resch: In my view, the Grid evangelists peddled highly unrealistic claims and created extremely high expectations. For a long time, we aimed at the wrong goals, ignoring those which could have been achieved. Industry, for example, is ahead of the scientific community in many aspects of Grid computing. Industry is focusing on solving real problems. A company like Daimler/Chrysler looks at how to combine and integrate the available computing resources within the whole company, to maximize productivity by sharing all these resources.

Lazou: You are a speaker of the board of the center for competence in HPC of the State of Württemberg. Can you describe what this involves and how policy from this body is translated into action and implemented throughout the state of Württemberg?

Resch: The State of Württemberg decided that there should be a mechanism created to bring together various institutions so together they can improve services. This substantially increases the critical mass of expertise for solving a particular problem and, at the same time, it frees other resources for doing innovative research at other places. The establishment of this center reflects the importance of HPC for the state of Baden-Württemberg. It recognized the key role HPC technology can play in supporting a high technology based industry. The center of competence is another piece in the bigger picture for maintaining economic competitiveness.

Lazou: Are there any other comments you would like to add about HPC?

Resch: I believe there is too much discussion and focus on HPC architectures and not enough emphasis on technology. There is an obsession with peak performance and quantity, and ruefully little interest in quality and timely deliverance of application solutions. HPC lost its way; it is not concerned with solutions but merely peak performance. This ethos needs to be reversed urgently. Developing the right technology to deliver productivity and the capability for successful solutions of large-scale problems should be placed at the top of a priority list.

Lazou: Thank you for your invaluable time and in sharing with our readers your opinions and insights on the challenges facing the HPC community.

(Brands and names are the property of their respective owners)
Copyright: Christopher Lazou, HiPerCom Consultants, Ltd., UK. June 2005.