A Method for Performance Prediction of Parallel Programs

In general there are many possibilities to parallelize an algorithm and to distribute the tasks to the processors. In some cases the selection of a processor configuration depends on the given problem, too. To decide which implementation for any configuration may be the best, the programmer can use methods of performance evaluation and prediction. If the structure of the parallel program can be modeled by a seriesparallel graph we know a method to obtain the runtime of the total program. This paper deals with the more difficult problem of predicting the runtimes of non-seriesparallel structured programs. It is shown how to compute approximately the average runtime of a program which consists of tasks with deterministically and/or exponentially distributed runtime variables. The method is based on the well known transient state space analysis.