Reverse profiling

This paper addresses the problem of designing parallel message passing programs with a reasonable idea of how well they will actually perform before they are run. Models with very few parameters (e.g. LogP, PRAM) sacrifice accuracy to simplify design. By contrast, simulation techniques provide a good degree of accuracy by incorporating sophisticated architectural models, but present a “black box” to the user. This paper suggests a compromise between the two extremes, using an automatically generated model with a large number of parameters (a separate equation for each MPI function) which is presented to the user rather than being hidden within a black box. The profiling interface of MPI may be used “in reverse” to insert (rather than measure) expected timings from the model.