A Comparison of the Intel iPSC/860 and the Suprenum-1 Parallel Computers
COLORADO UNIV AT BOULDER DEPT OF COMPUTER SCIENCE
Pagination or Media Count:
The author compares the Intel iPSC860 and SUPRENUM-1 parallel computers using a well-known scientific application algorithm. The algorithm, the Shallow Water Equations, is frequently used as a model for both oceanographic and atmospheric circulation. He describes the steps involved in implementing the algorithm on the iPSC860 and the SUPRENUM-1, and he provides details of their performance. Surprisingly, the results show that the SUPRENUM-1 provides better performance on both a single node and a multiple node basis, despite the much higher theoretical peak rate of the i860 nodes. Using the latest Intel PGI compiler, he measured 4.63 Mflops per node on the iPSC860, with efficiencies of 92.7 on 16 nodes, and 91.6 on 128 nodes. Earlier measurements with the standard Greenhills compiler yielded only 2.7 Mflops per node on the iPSC860. With an early version of the SUPRENUM-1 compiler, he measured 5.11 Mflops 64-bit arithmetic for single-node performance, and efficiencies of 88.3 with 16 nodes 72 Mflops aggregate performance. While a 256-node SUPRENUM-1 was not yet available for measurement, the author provides a simple static load-balancing algorithm for hierarchical systems that effectively extends the observed single-cluster efficiency for grid algorithms to large systems.
- Computer Hardware
- Computer Systems Management and Standards