Statistical analysis of message passing programs to guide computer design

Little data exists on how message passing programs use parallel computers. The behavior of these programs can strongly influence design decisions made for future computer systems. The computer designer's use of incorrect assumptions about program behavior can degrade performance. In many cases simple statistical parameters describing characteristics such as message sizes, destinations, sources, and times between sends would give the designers of the communication libraries and the computer hardware great insight into how the hardware is used by actual programs. Techniques of collecting statistical information about the communication characteristics for system design have been applied to the parallel version of the NAS benchmarks. The paper describes the statistical data collected for multiprocessor runs of the NAS 2.1 benchmarks and some of the characteristics observed in that data.