by Uwe Harms
Munich, GERMANY — Supercomputing throughout Europe is no longer restricted to scientific and academic users. Industrial applications mostly from the commercial area now dominate the scene, such as SAP R/3, data warehousing, data mining and telecommunications. These commercial users do not install the ultra high-end machines, as is evident from the Top500 list of the world’s most powerful supercomputers where academia and research still hold sway with a total capacity of 9 TFlop/s compared to 6.2 for industry.
This article will focus in particular on Germany for, of all European countries, Germany has the largest installed scientific supercomputing capacity and offers examples of the wide range of uses of supercomputing and the different arrangements for organizing and running supercomputing centres. Several of the country’s academic research centres have installed high-end computers to run grand challenge problems. These cover a broad spectrum of applications in different sciences: from climate research to engineering; from theoretical chemistry to superconductivity. It is rather difficult to filter the projects and measure their importance in CPU hours. But in January 1999, the German Scientific Council, a high-level advisory board for the German Government and Research Ministry, published a report discussing the future of supercomputing in Germany. The report listed the usage and the research projects at German supercomputing centres.
At the Leibniz-Rechenzentrum in Munich, Bavaria, mechanical engineering is the biggest user of the Fujitsu Siemens VPP 700, a vector machine with 52 processors: about half the usage goes on this application and includes, for example, computational fluid dynamics and mechanics for hypersonic transportation. Bavarian scientists also use the big machines elsewhere in Germany, at the Hˆchstleistungsrechenzentrum Stuttgart (HLRS); and the John von Neumann Institute for Computing (NIC) at the J¸lich Research Centre. In 1998 they accounted for about half the usage of the Cray T3E and 13 per cent of the NEC machines at HLRS, while at NIC their usage accounted for 18 per cent of the Cray T90 and 6 per cent of the Cray T3E.
The biggest user in both centres is the department of theoretical physics at the University W¸rzburg. There Professor Werner Hanke and his team simulate and conduct research in the area of high-temperature superconductivity. (High temperature means -100 degrees Celsius, compared to the former temperature of -266 degrees Celsius.) They have conducted a systematic Quantum-Monte-Carlo (QMC) study for the single-band Hubbard model, which is intended to describe the relevant low-energy degrees of freedom of these celebrated high-temperature superconductors. Most of the CPU-time of a QMC-simulation deals with matrix-matrix multiplication, matrix inversion, and different stabilisation algorithms for these large-scale matrices. This results in large vector lengths and therefore an excellent vectorisation. The algorithm is also perfectly (almost 100 percent) parallelised by simply letting each processor run independently with a different (random) start configuration of the Ising field. Therefore an astonishing total performance of 16 x 1.2 GFlop/s – in total 19 GFlop/s – was achieved on the VPP700 (using 16 parallel processors).
In its 1999 report, the German Scientific Council nominated the Leibniz-Rechnenzentrum (LRZ) in Munich as the nexta federal centre for supercomputing in Germany. This decision gave the green light for the purchase of a new supercomputer, the Hitachi SR8000-F1, which was inaugurated at the LRZ earlier this summer.
Professor Hanke’s superconductivity research will be one of the main projects on the new machine. Another important research field is theoretical chemistry. During the opening ceremony for the Hitachi supercomputer at LRZ, Professor Notker Rˆsch from the Technische Universit‰t M¸nchen) gave an impressive overview what he expects from the Teraflop power of the new machine. Without a massive increase in computer power, research in theoretical chemistry will not be possible. He sketched out new tasks in computer chemistry, for example nano-science. In biochemical research area, calculations of electronic properties of DNA will be carried out. DNA strands can be used in nanotechnology as connectors and in the biochip world. In nano-electronics researchers are trying to reduce the size of circuits to the point where they can switch with single electrons. A prototype is the SET (single electron tunnelling) transistor. Computer chemistry allows detailed insights into the electronic and geometric structures of nano-structured materials.
Another research area is catalysis, to improve and drive chemical reactions. To understand their behaviour is a complex task, proportional to the number of atoms. A final application of supercomputing mentioned in Professor Rˆsch’s speech was molecular processes in the environmental chemistry, for example, the distribution of heavy metals like cadmium in water or the take-up by micro-organisms.
Small-scale climate research in Bavaria is another fruitful research area. Global climate or weather simulations have a grid size of 100 to 500 km. Regional features such as mountains or the course of a river fall through this mesh. Meteorologists working in Bavaria have been able to downscale synoptic scale atmospheric information via multiple nesting to regional scales on the order of a few kilometres. Thus they have been able to incorporate a rather realistic representation of the complex orography of the Alps and Bavaria along with the associated dynamics and physics. The techniques mean that the local model is integrated in the global model. Within the Bavarian climate research program (BayFORKLIM), they have conducted two five-year simulations for southern Germany with a resolution of 15 km: one with today’s climate and one on a climate with doubled atmospheric carbon dioxide concentrations. Additionally 1 km simulations contain clouds and thunderstorms for example. Longer simulation times and high-performance computers are required to deliver statistically relevant results.
Nearly 40 project proposals from German universities and Max Planck Institutes have been received for the new Hitachi SR8000-F1Hˆchstleistungsrechner in Bavaria (HLRB). New simulation programs can replace or at least augment costly real world experiments and can deliver deep insights. The ‘Grand Challenge’ projects make an enormous demand for computing and memory resources: each project would be able to use the whole capacity of the HLRB. One third of the proposed projects belong to computational fluid mechanics (CFD), dealing with turbulence or high inertia-friction relationships. Simulations of complex three-dimensional geometries such as stirred vessels and turbines have significant industrial importance. Other topics consider fluid-structure interactions, for example with tube bundles in heat exchangers that are subject to vibration induced by cross-flow. Further aspects of CFD research are the simulation of unsaturated groundwater flow in soils or the simulation of production processes for silicon single crystals, which play an important role in the fabrication of electronic devices.
Physics projects work on the numerical solution of the full three-dimensional set of Einstein’s equations of general relativity. Their results will help to simulate black hole collisions and to predict details of the gravitational waves generated by such events. At the other end of the scale – in the atomic region – we find projects promising improved designs of electron microscopes or investigating models for quantum mechanical effects at boundary layers. With the help of extensive computer simulations, researchers hope to determine the mechanisms of chemical reactions that have not yet been reproduced by means of catalytic systems in the laboratory. A prominent example of such a reaction is the conversion of molecular nitrogen to ammonia, one of the fundamental reactions of nature. Further project proposals deal with simulations of seismic wave propagation of earthquakes, large-scale heat and gas transport in oceans for more realistic climate studies, stellar explosions and gravitational clustering.
At the John von Neumann Institute for Computing (NIC) in the J¸lich Research Centre, theoretical chemistry and physics are the main research areas, using the Cray vector and the NIC’s massively parallel computers. In elementary particle physics, the deeper researchers look into the microcosm, the more complicated its structure appears. The vacuum itself can be characterised by quantum fluctuations on the subatomic scale. In many-body physics, one research topic is the flow behaviour of granular matter such as cereals, plastic granulates and sand through a tube. When two grains collide, part of the energy is converted into heat, which reduces their relative velocity. The simulation shows the optimal tube size and filling velocity. Such results can be used in industry and is of particular interest to the pharmaceutical industry in the final stages of manufacturing drugs and packaging them into tablet or other controlled-dose form.
For an equally challenging but very domestic problem, consider the ceramic cooking hobs found in many of today’s kitchens. The main constituent of the CERAN cooking tops manufactured by the Schott company is an alumino-lithio-silica glass-ceramic, called beta-eucryptite. It has a remarkably small thermal expansion coefficient over a temperature range of around 1,000 degrees. For the crystalline form of beta-eucryptite (oxygen atoms are green, lithium atoms brown, silicon atoms blue, and aluminium atoms red), density function calculations – which are free of adjustable parameters – show that the thermal expansion coefficients parallel and at right angles to the lithium chains are almost constant over a large temperature range. Since the atomic movement can be followed in a precise way in computer calculations, it is possible to understand the reasons for the material’s thermal behaviour.
Because the NIC is deeply involved in theoretical chemistry, it organized a 5-day winterschool end of February 2000, “Modern Methods and Algorithms of Quantum Chemistry.” About 200 experts from the fields of quantum chemistry, computer science and applied mathematics met. Recent methodological and computational advances in the field of theoretical chemistry and their applications were presented. The twenty lectures covered the topics: density function theory, ab initio molecular dynamics, post-Hartree-Fock methods, molecular properties, heavy-element chemistry, linear scaling approaches, semi-empirical and hybrid methods, parallel programming models and tools e.g. performance analysis and automatic differentiation and industrial applications.
Printed versions of the talks as well as the abstracts of the posters of this Winterschool are published in a new book series of NIC. All the material is additionally available on the web, the talks in Postscript and PDF format, the posters in Postscript only:
A medical application, which uses their Cray T3E and distributed visualisation, is the mapping of brain functions via fast magnetic resonance measurements of the changes in blood oxygen saturation during neural activation. Magnetic resonance imaging methods have high temporal resolution (about 1 mm and 100 ms per tomogram), but current image analysis techniques for the detection of brain activation can be applied only after the measurement and are very time-consuming. A real-time analysis offers a number of advantages (e.g. optimization of the stimulation conditions and quality control) and opens up new neuroscientific applications, e.g. in the field of bio-feedback. In co-operation with Siemens Medical Systems and Algorithmicon, a Siemens Vision 1.5 Tesla whole-body scanner was modified for real-time measurements. Real-time correlation analysis of image intensity changes has been developed and implemented on Unix workstations and on the Cray T3E supercomputer at NIC. This iterative correlation technique can be limited to a time window of freely selectable width, which is moved on with every new measurement to determine the temporal dynamics of brain activation. Primary sensor activations of the visual, motor and auditory cortex can be detected within a few seconds.
In addition, new spectroscopic imaging methods were developed, which even permitted real-time detection of neural activation for the control of individual finger movements. Within the scope of a Gigabit networking project, the results of the T3E runs were transferred to the SGI Onyx 2 at GMD (the Research Centre for Information Technology) in St. Augustin, Bonn, and visualised on a responsive workbench (a virtual reality display). Alternatively the images as computed at GMD can be transferred via the Gigabit network back to J¸lich. There they can be displayed on another, smaller responsive workbench or on a graphics workstation. However, the transfer of the images puts a high load on the network.
The Hˆchstleistungsrechenzentrum Stuttgart is an unusual supercomputer facility, in that academia and industry share the same computers, through a jointly owned company. Half the shares are held by Baden-W¸rttemberg, the Universities of Karlsruhe and Stuttgart; 40 per cent are owned by debis Systemhaus, a subsidiary of DaimlerChrysler; and 10 per cent by Porsche AG. In the industrial sector, automotive research is done, such as crash simulation, computational fluid dynamics, and structural analysis. Half of the academic part is open to researchers from anywhere in Germany, while the rest is reserved for those from the local Federal Land of Baden-W¸rttemberg. The main academic users are, as in the other centres, from physics and chemistry. In the debis computer centre a Cray SV1, 20 CPUs, and the massively parallel Cray T3E with 512 processors are installed. In the University computing centre one can currently find two NEC SX-4, 32 CPUs, NEC SX-4, 8 CPUs, and two NEC SX-5e, 16 CPUs. The NEC-SX5 systems are used for vectorised programs, such as the FEM (finite element method) program packages. One example is the optimization of the combustion of power stations, which are fed with coal dust or biomass. Physics and chemists access the massively parallel system.
Another example of a shared facility is in Karlsruhe, where the Research Centre Karlsruhe (RCK) and University Karlsruhe share their supercomputers: a virtual computer centre with a 512 processor IBM SP2 at the University; and a Fujitsu Siemens VPP300/16 (and the just installed VPP5000/4) at RCK. As the University was a Technische Hochschule, it is more engineering oriented. Thus computational fluid dynamics and finite element analysis are important research fields. On the other hand theoretical chemistry with density function theory, molecular dynamics and the quantum chemical ab initio program Turbomole are also applications in these centres.
RCK concentrates on environmentally benign high technologies, environment, energy, microsystem/medicine techniques and basic research. In one molecular dynamics application, the program computes the time sequence of different physical elements, for example position and speed of single atoms or molecules and their interaction. In an atom cluster, the main question is the stability: what changes the structure or keeps it stable. One of the data inputs represents a rotor, consisting of 4,096 aluminium atoms. A laser beam forces it to rotate at 100 billion revolutions per second. Up to 75 billion revolutions per second, the rotor is stable but then it breaks up into its parts. This simulation needs about 420 seconds on the VPP300.
In July this year RCK installed a 4 processor VPP5000 with a peak performance of 9.6 GFlop/s per processor – a factor of four improvement compared to the VPP300. It will be extended next summer to 8 processors.
The look in detail at the different supercomputer centres shows that there are a lot of similarities in the research topics. Most computing time is consumed by theoretical chemistry and physics. That was one reason for the establishment of the predecessor of NIC, the Hˆchstleistungsrechenzentrum in J¸lich. Another aspect is the growing co-operation with industry in many projects.
University Karlsruhe, High Performance Supercomputing http://www.ssc.uni-karlsruhe.de
Research Centre Karlsruhe http://www.fzk.de
Interesting information about these projects, printed brochures in English can be obtained: John von Neumann Institute for Computing http://www.fz-juelich.de/nic
LRZ: Research Projects on the High Performance Computers of the LRZ, including a CD (112 projects) http://www.lrz.de
CSCS/NEC: High-Performance Instruments for Innovation …, http://www.hpc.comp.nec.co.jp/sx-e/index-html Joint Application Support SX-4 Task Force http:www.cscs.ch http://www.ess.nec.de
— Uwe Harms is a supercomputing consultant and owner of Harms-Supercomputing-Consulting in Munich, Germany.