Performance Prediction by Trace Transformation

Performance stability is an essential feature for the widespread adoption of multicomputers. In this paper, we report the preliminary steps of our research in performance prediction and extrapolation. Performance tuning, guided by extrapolation, may help achieve a substantial fraction of peak performance rates across a broader range of applications while providing guidance for code porting. We introduce a methodology for assessing stability of parallel programs, based on stability of the program execution graph, using time perturbation analysis. For programs with stable behavior, we present a model for performance prediction under architecture variations, by transformation of the execution traces with parameters that reeect the diierences in architecture between two systems. We illustrate the use of this transformation with an example of a parallel PDE solver executing on a multicomputer.