Evaluation Of Benchmark Performance Estimation For Parallel Fortran Programs On Massively Parallel SIMD And MIMD Computers
Thomas Fahringer · 2005
A potential problem encountered when parallelizing programs for massively parallel systems is to guide the parallelization effort through performance prediction. Estimating the performance of parallel programs based on benchmarking is getting increasingly popular in recent years. However, there was little research done so far to evaluate this approach mainly due to the lack of actual implementations. This paper discusses the advantages and disadvantages of benchmark performance estimation for SIMD and MIMD machines. The design and implementation of a benchmark performance estimator is presented. Even though benchmark performance estimations have been demonstrated to be very useful, experiments based on the described prototype uncover several severe problems of this approach. This includes time effort, portability, measurement complexity, performance influence of target machine and compiler, pattern matching of kernels, and prediction accuracy. Concrete experiments for the MasPar MP-1 a...