CSR-PSCE, SM: MPI-PPA: Improving Efficiency of Large-Scale Clusters Through Statistical Performance Prediction

该项目开发了一个系统，可以提高大量处理器（高达数万或数十万）的并行效率，而无需大规模运行程序。该系统被称为MPI-PPA：MPI性能预测和建议。 MPI-PPA将科学计算应用沿着输入变量一起作为输入，输入变量包括期望的处理器数量p。在仅在少于p个处理器上执行的情况下-使得这些执行将快速发生- MPI-PPA将产生被预测为实现差的可扩展性的程序阶段的列表，允许程序员快速处理并可能重新实现这些阶段-以及对整个程序运行的预测。PPA使用统计回归来进行这些预测，以开发可以与任何数量的处理器一起使用的预测函数。 MPI-PPA不需要大量的程序理解，这是考虑到计算科学家通常是其科学领域而不是计算机科学领域的专家时的一个重要方面。 MPI-PPA的方法涉及对统计技术的严重依赖，因此该项目的工作将是计算机科学（PI）和统计学（co-PI）之间的跨学科。 MPI-PPA将通过使用NAS和ASCI代码等基准套件进行验证，沿着国家实验室感兴趣的大规模应用，如帕拉迪斯和Raptor。这项工作的广泛影响是多方面的。 首先，MPI-PPA将有利于计算科学家和集群管理员。 其中的好处将是一个简单而快速的性能调优系统，提高整体集群效率，并减少单个应用程序的响应时间。 该项目开发的技术将以性能调整和预测软件的形式转让，并通过与劳伦斯利弗莫尔国家实验室的合作向公众提供。 第二，将通过格鲁吉亚大学的监督统计咨询中心促进统计学和计算机科学之间更多的跨学科互动。第三，将继续努力从该地区历史悠久的黑人学院和大学（如莫尔豪斯大学）招收学生。

This project develops a system that improves parallel efficiency on large numbers of processors - up to tens or hundreds of thousands - without running a program at scale. This system is called MPI-PPA: MPI Performance Prediction and Advisement. MPI-PPA takes as input a scientific computing application along with the input variables, including the desired number of processors, p. With executions on fewer than p processors only - so that these executions will occur quickly - MPI-PPA will produce a list of program phases that are predicted to achieve poor scalability, allowing the programmer to quickly address and possibly re-implement these phases - as well as a prediction for the entire program run.MPI-PPA makes these predictions using statistical regression to develop a prediction function that can be used with any number of processors. MPI-PPA will not require significant program comprehension, an important aspect when considering that computational scientists are typically experts in their scientific domain and not in computer science. The approach of MPI-PPA involves heavy reliance on statistical techniques, so the work in this project will be interdisciplinary between computer science (the PI) and statistics (the co-PI). MPI-PPA will be validated by using benchmark suites such as NAS and ASCI codes, along with large-scale applications - such as Paradis and Raptor - that are of interest to national labs.The broader impact of this work is multifold. First, MPI-PPA will be beneficial for computational scientists as well as cluster administrators. Among the benefits will be a simple and fast performance tuning system, an increase in overall cluster efficiency, and a reduction in response times for individual applications. The technology developed in this project will be transferred, in the form of performance tuning and prediction software, and made available to the public through cooperation with Lawrence Livermore National Laboratory. Second, more interdisciplinary interaction between statistics and computer science will be fostered through the supervised statistical consulting center at the University of Georgia. Third, efforts will continue recruiting students from strong historically black colleges and universities in the area, such as Morehouse University.