A Statistical Approach to Predicting the Performance of Concurrent Programs

We suggest various highly simpliied models which may be employed to make estimates of the performance of a concurrent program running on a multiprocessor machine. We try to nd a small set of parameters which have an important innuence on performance, and conduct simulation experiments on synthetic programs whose parameter values are varied systematically. We then apply standard statistical methods to analyse the results and derive a quantitative relation between a performance metric and the model parameters. We give an example to illustrate the kind of information that can be derived from this approach. We describe ongoing research which is directed towards understanding the way in which diierent types of models (representing diierent classes of program) behave under various compile and run-time allocation strategies.