Instruction timings and message passing performance of the connection machine 2

In this study we have investigated various methods of collecting low level C/Paris instruction execution times, including some of those instructions for message passing. We have written an adaptive timing code to collect actual execution times for single instructions and code fragments. This adaptive scheme reduces the runtime needed to collect the benchmark data without loss of accuracy. Several different CM-2 front ends were used, however due to space limitations only a subset of the data for the Sun-4 front end is presented here. We have started investigating message passing performance, with early results of a simple five dimensional grid presented. Communication patterns typical in both numeric and nonnumeric large scale computations will ultimately be measured. The results obtained in this study can be useful for optimizing CM-2 programs and understanding choices during implementation of C/Paris programs. One small example of how the instruction timing data could be used to estimate code fragment execution time is presented. 5 refs.