by Rolf Hempel
After a successful first year of cooperation, GMD and the NEC C&C Research lab held the "Second GMD - NEC Workshop on Scientific Parallel Computing" at GMD's Schloss Birlinghoven site on October 11 to 12 1995.
The NEC Lab had been established as part of the GMD TechnoPark in 1994. NEC already had a lab in Princeton, USA, and this was the second NEC research center outside Japan. Since its opening, the GMD Scientific Computing and Algorithms Institute and NEC have been closely cooperating in the area of parallel supercomputing. Research topics include implementations of large-scale application programs and research in new algorithms and programming interfaces. NEC made a high performance parallel system Cenju-3 available to GMD free of charge, which GMD operates as part of its parallel computing center. The machine consists of 64 computing elements, each one equipped with a Mips 4400 processor and 64 Mbytes of memory.
The intention of this year's workshop was to present a broad spectrum of research activities in the scientific parallel computing area, with emphasis on joint research between GMD, NEC and other cooperation partners. The first day focussed on numerical applications, the second day was devoted to hardware, system software and tools.
A total of 36 scientists from Germany, Great Britain, Japan, the Netherlands, Switzerland and the U.S. participated in the workshop. In his keynote speech on the opening day, Prof. Arndt Bode of Munich University gave an overview of the programming tools currently available for parallel computing. Although there are many tools on the market, many of them have proved more or less useless in practical applications. He presented some interesting new developments at his institution in Munich.
The topic of the second keynote speaker, Prof. Oliver McBryan from the University of Colorado at Boulder, was the environment of modern high performance computers, consisting of high speed networks and high performance storage systems. Often in real world applications the attainable speed is not limited by the processor performance or the network bandwidth but by the throughput of the disk storage system.
The GMD Cenju-3 has so far been used mainly by the large-scale applications which were developed or parallelized at GMD. The basis for this has been the implementation of the PARMACS programming interface, which in the long run will be replaced by the Message Passing Interface (MPI). A GMD scientist talked about experiences with the application portations in general and fluid flow solvers in particular, and compared the performance of the NEC system with other parallel machines such as the IBM SP-2. In the area of programming interfaces there was a status report on the MPI standardization, while the GMD spin-off company PALLAS GmbH presented PARMACS and related programming tools.
Two other European institutions operate Cenju-3 computers: the Swiss scientific computing center CSCS and the National Aerospace Laboratory NLR in the Netherlands. Representatives from both institutes gave an overview of their application projects and programming tool developments.
A panel discussion ended the technical workshop programme. A key issue was the question of the most suitable programming model for parallel computers. The efficiency of the message passing model has to be balanced against the ease of use of a virtual shared memory environment. As the discussion did not lead to a general preference of one of the alternatives, it is obvious there is still a need for more research in this area. It is planned to hold a third workshop in this series at the end of 1996.