The Sensitivity of Methods for Partitioning Parallel Programs to Inaccurate Execution Time Estimates

将程序映射到并行体系结构的问题被视为一个优化问题。这个透视图强调了可能影响解决方案质量的问题实例描述中的几个不准确之处。这项工作评估了程序划分方法对不准确的执行时间估计的敏感性。其他的研究集中在评估整体制图问题的潜在解决方案上，而忽略了这种不准确性的重要影响。这项工作将进一步全面理解映射问题，并将有助于指导程序映射新方法的发展。本文将对几类程序和目标体系结构以及几种划分方法进行评估。实验将系统地改变分区方法所使用的执行时间估计的准确性。将测量和分析结果映射的实际执行时间，以表征分区方法对不准确的执行时间估计的敏感性。

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.