First and second generation hypercube performance

This thesis contains a detailed comparison of the computation and communication perfor­mance of two commercially-available hypercube multiprocessors, the Intel iPSC and Intel iPSC/2. In a previous performance study [7], some of the benchmarks where also run on an Ncube/four hypercube; where appropriate, the results of these tests are also included. Three kinds of benchmarks were used. The processor benchmarks measure computation performance. Many of these benchmarks are well-known programs used to measure tradi­tional uniprocessor computers [15, 29]. Because computation performance is a function of compiler as well as processor, both C and FORTRAN benchmarks are included. In addition to comparing processor/ compiler performance across machines, the relative performance of the C and FORTRAN compilers for each machine is examined The second class of benchmarks, the communication benchmarks, measures conunu­nication performance. The communication benchmarks consist of simple conununication benchmarks, which measure link level (i.e. point-to-point) communication characteristics, and synthetic communication benchmarks, which measure the performance of the entire communication network under parametrically-generated communication loads. Both the processor and communication benchmark set are based on the work of Grun­wald and Reed [16]. The third set of benchmarks is a set of application programs that depend on both the computation and communication performance of the hypercubes.National Science Foundation grants NSF DCR 84-17948, NSF CCR 86-57696, NSF CCR 87-06653 and NSF ANTI TAPESTRY 1-5-30035National Aeronautics and Space Administration NASA Contract Number NAG-1-613Air Force Office of Scientific Research grant AFOSR­F49620-86-C-0136 (URI)equipment donation from Digital Equipment CorporationU of I OnlyRestricted to the U of I community indenfinitely during batch ingest of legacy ETDs