Parallel program performance prediction using deterministic task graph analysis

The techniques for predicting detailed performance characteristics of a single shared memory parallel program were discussed. A deterministic task graph analysis was developed that provided detailed performance prediction for shared-memory programs with arbitary task graps. The model assumed the deterministic task execution times and was evaluated in three ways. The results show that the technique is accurate and efficient for a variety of shared-memory programs, and the deterministic assumption is crucial to permit accurate and efficient analysis of these programs.