Collaborative: CSR-AES: System Support for Auto-tuning MPI Applications

原子（MPI软件的自动调优）项目正在研究一种能够自动提高大规模科学应用程序性能的软件系统。需要越来越多的计算资源的科学代码对现代科学至关重要，但科学家们常常不得不花时间构建运行速度很快的程序，而牺牲了他们的基本研究。随着计算机包含越来越多的计算元素，问题变得更加严重。AToMS项目的目标是通过应用自动应用程序调优开始解决这个问题。优化编译器将程序转换为语义上等价的、性能更好的程序。在处理任何复杂的体系结构时，很难知道哪些转换将提高性能。自动调优采用尝试许多转换并根据经验评估结果版本的方法。AToMS通过结合基于静态分析的代码转换引擎（称为沥青）和OpenMPI库中的运行时支持来执行这种自动调优。编译时和运行时支持的组合允许代码重构以重叠计算和通信以及创建优化的数据打包例程。此外，还可以生成代码来利用多核处理器架构。智力上的优点：这个项目的优点在于理解了支持自动可调MPI程序所需要的东西。更广泛的影响：该项目将通过使并行计算机更容易实现良好的性能来影响高性能和科学计算社区以及并行计算机的用户。

The AToMS (Automatic Tuning of MPI Software) project is investigating a software system that can automatically improve the performance of large-scale scientific applications. Scientific codes that demand more and more computing resources are critical to modern science, but too often scientists must spend time constructing programs that run fast at the expense of doing their primary research. As computers contain an increasing number of computing elements, the problem worsens. The goal of the AToMS project is to begin to address this issue by applying automatic application tuning.An optimizing compiler transforms programs into sematically equivalent ones that perform better. When dealing with any complicated architecture it is difficult to know which transformations will improve performance. Auto-tuning takes the approach of trying many transformations and empirically evaluating the resulting versions. AToMS performs this auto-tuning with a combination of a static analysis based code transformation engine (called ASPhALT) and runtime support in the OpenMPI library. The combination of compile-time and run-time support allows for code restructuring to overlap computation and communication and the creation of optimized data-packing routines. In addition, code can be generated to take advantage of multicore processorarchitectures.Intellectual Merit: The merit of the proposed pro ject is in gaining understanding about what is required to support automatically tunable MPI programs. Broader Impacts: This project will impact the high-performance and scientific computing community and users of parallel computers by making it easier to achieve good performance.