Abstract.
Abstract:
We describe the implementation of a fluid dynamical benchmark on the 256 node SUPRENUM-1 parallel computer. The benchmark, the Shallow Water Equations, is frequently used as a model for both oceanographic and atmospheric circulation. We describe the steps involved in implementing the algorithm on the SUPRENUM-1 and we provide details of performance. Optimal SUPRENUM performance requires algorithms that may be compiled into vector instructions with long vector length, and as with many other MIMD systems, relatively few communication operations. For such algorithms the system delivers a very impressive fraction of its theoretical peak rate. SUPRENUM software is excellent, including communication facilities and a fully vectorizing compiler for Fortran 77 which was used in this study. We have measured 5.33 Mflops (64-bit arithmetic) for single node performance, and 1280 Mflops aggregate performance with 256 nodes, at efficiencies up to 95%. This compares well with vector and MIMD supercomputers and shows that SUPRENUM was among the fastest MIMD computers during 1992. Performance of 1530 Mflops was measured for the same algorithm on the CRAY YMP/8, and 543 Mflops was measured on the 128-node Intel iPSC/860. The SIMD Thinking Machines CM-200 delivers 5.25 Gflops (64-bit) and 8.09 Gflops (32-bit) for the benchmark. We also discuss the influence of physical cluster interconnection topology and asynchronous communication on SUPRENUM performance.
