The problem of mapping programs to parallel architectures is cast as an optimization problem. This perspective highlights several inaccuracies in the description of the problem instance that may affect solution quality. This work evaluates the sensitivity of program partitioning methods to inaccurate estimates of execution time. Other research has concentrated on evaluating potential solutions to the overall mapping problem and has ignored the important effects of such inaccuracies. The work will further overall understanding of the mapping problem and will help guide the development of new methods for program mappings. The evaluation will be accomplished for several classes of programs and target architectures and several partitioning methods. Experiments will systematically vary the accuracy of execution time estimates used by the partitioning methods. Actual execution times of the resulting mappings will be measured and analyzed to characterize the sensitivity of the partitioning methods to inaccurate execution time estimates.
