Hardware performance and communication models of a distributed memory parallel computer system: A case study

This paper describes the methods to evaluate the hardware performance and to model the communication performance of a distributed-memory vector parallel computer. The multi-processor VPP300 supercomputer and the parallel language VPPF (VPP-FORTRAN) were used for this case study. For a given vector parallel system, knowledge of parallel hardware performance and inter-processor communication performance is essential for effective code vectorization, parallelization and application development. In this work, the measurement of hardware performance is based on the performance of selected vectorizable do-loops while the communication performance modeling focuses on the communication startup time and message passing time. Several models have been used to verify the experimental results for popular communication patterns such as point-to-point, one-to-all, all-to-one, all-to-all, broadcast, and exchange communications.