HIGH PERFORMANCE MULTIPROCESSOR SYSTEMS FOR NUMERICAL SIMULATION.

A supercomputer architecture is derived from the analysis of algorithms for the numerical simulation of physical phenomena. In particular the class of multigrid algorithms is shown to be well suited for parallel computing because of the local behavior of the operations. The grid structure of the problem (2-D or 3-D) is mapped onto the nearest neighbor structure of the memory-coupled multiprocessor with a hierarchical control structure. Experimental results, theoretical studies, and simulation experiments show the flexibility and efficiency of such EGPA (Erlangen General Purpose Array) systems. The overall performance of the proposed system will be 1 GFLOPS, 1-GB main memory, 38-GB mass storage, and 640-Mb/s data transfer rate from mass storage. The system topology and structure of an arithmetic coprocessor for the processor-memory modules is described.