Since June 1993 two times a year the Top500 list is published, the most powerful computers in the world, measured by the Linpack benchmark. These Top500 allow a detailed and well based analysis of the market for high-performance systems, the developments and trends over this timeframe. This article analyses in detail, how the HPC systems developed in Europe and their usage in the different fields, academia, research and industry. This analysis is based on the June 1998 list that will be published on June 18, 1998.
The Top500
The Top500 list is published in June at Prof. Hans Meuer's Supercomputer Seminar - the first time in 1993 - and in November at the IEEE Supercomputing Conference and has been fully established in the international scene. The base of the performance is the Linpack Benchmark, the measured performance rate for solving a huge linear equation in GFlop/s (Billions of floating point operations per second) - Rmax. On the Intel Red a system with 235 000 unknowns has been solved. As the Top500 team has collected a lot of information over these years, the lists are an excellent base for studies in the High-Performance (HPC) market. Especially the reader can rate the national competitiveness compared to the international scene. Furthermore, the timely aspects of the HPC market are interesting, e.g. the changes of the technologies that are used in the HPC systems.
The linear system has to be solved by an LU factorisation. The Strassen algorithm that requires less operations but is numerically not so stable is not allowed. Some vendors used it in the past and divided by the number of LU operations by that time. That led to performances that exceeded the peak performance - the vendors promised in former times not to exceed. This effect can be directly seen with vector computers but not with massively parallel systems.
Within the list one finds the rank, the manufacturer, the computer, the Linpack performance Rmax in GFlop/s (Billion Floating Point Operations per Second), installation site, country, year of installation, area of installation, number of processors, Rpeak (Peak Performance in GFlop/s), Nmax (dimension of the linear system) and N1/2 (dimension of the linear system at which the machine reaches half of the peak performance - a good measure for the inherent parallelism, this was defined by Hockney). In the cases, when no data for the HPC is available, the data from the next smaller system is used, for example the Max-Planck-Ges. (nr. 9) with 672 processors has the same Rmax as Research Center Juelich with 512 processors.
In this analysis the situation in Europe will be compared with the world-wide trend. Specific considerations will be given to Hewlett-Packard's situation in this context.
Literature:
This analysis is based on German articles, published by Erich Strohmaier and me:
Uwe Harms, Erich Strohmaier, HPC Nutzung in Deutschland - Eine Analyse anhand der TOP500, S. 37 - 42, PIK (Praxis der Informationsverarbeitung und Kommunikation), 1/98, K.G. Sauer Verlag, Muenchen, 1998
Uwe Harms, Erich Strohmaier, HPC Nutzung in Deutschland - Eine Analyse anhand der TOP500, S. 41 - 52, FOKUS Band 15, Supercomputer 1997, K.G. Sauer Verlag, Muenchen 1997
Uwe Harms, Erich Strohmaier, HPC usage in Germany, an analysis based on the Top500 in Primeur, May issue, http://www.hoise.com/primeur