Runs of benchmarks or a suite of benchmarks are inadequate to characterize parallel machines or to predict the execution time of applications with a computational kernel which is not included in the suite. Furthermore, the observed results are quite sensitive to the nature of the benchmark. In this paper, we present a new approach to the benchmarking of SIMD machines. The idea is to create and use a machine characterizer which measures the performance of given system in terms of high-level language instructions (e.g. Fortran 90 or TAO). The analyzer yields a set of parameters which provides the execution time for some primitive operations in the high-level language. This approach gives us considerable insight into both the machines and the programs, since the effects of individual parameters are immediately evident. We present experimental results on Quadrics machines.
展开▼
