Programmable Code Optimization and Empirical Tuning For High-end Computing

现代高端计算机的复杂性使得科学应用程序非常难以有效地管理资源，如极端规模的并行、单芯片多处理器以及共享/分布式高速缓存和存储器的深层层次结构。特别是，随着机器和应用程序都发展到复杂和大规模并行，编译器无法自动弥合复杂软件和不同硬件平台之间的差距。并行计算的优化模型远远落后于串行应用程序的优化模型，传统编译器越来越不能适应新兴的高端体系结构。本研究开发了一个新的优化模型，允许开发人员有效地与高级优化编译器交互，以提供特定领域的知识和高级优化策略(例如，启用新的方向或在不同的并行化策略中进行选择)；2)计算专家可以轻松定义任意特定领域的转换，以直接控制代码的性能优化；3)对体系结构敏感的优化，易于参数化和经验性调整，以实现可移植的高性能。优化模型由一个集成环境支持，该集成环境包含两个主要组件：DOE/LLNL开发的C/C/Fortran 2003源代码到源代码优化编译器ROSE和UTSA开发的转换语言POET和经验优化引擎。该框架允许不同级别的自动化和程序员干预，从全自动调谐到半自动开发再到完全可编程控制。该研究既针对计算内核的优化需求，也针对整个程序优化的更一般需求。该框架被集成为广泛采用的ATLAS库的外部开发机制，并与美国能源部SciDAC计划下的经验调整研究相连接，以提高大规模科学应用的效率。

The complexity of modern high-end computers has made it exceedingly difficult for scientific applications to effectively manage resources such as extreme-scale parallelism, single-chip multi-processors, and deep hierarchy of shared/distributed caches and memories. In particular, as machines and applications have both evolved to become complex and massively parallel, compilers have failed to automatically bridge the gap between complex software and diverse hardware platforms.Optimization models for parallel computing have lagged far behind those for serial applications, and conventional compilers are increasingly unable to accommodate emerging high-end architectures.This research develops a new optimization model that allows1) developers to effectively interact with advanced optimizing compilers to provide both domain-specific knowledge and high-level optimization strategies (e.g., directions to enable new or choose amongst differing parallelization strategies); 2) computational specialists to easily define arbitrary domain-specific transformations to directly control performance optimizations to their code; 3) architecture-sensitive optimizations to be easily parameterized and empirically tuned to achieve portable high performance.The optimization model is supported with an integrated environment that contains two main components: ROSE, a C/C++/Fortran2003 source-to-source optimizing compiler developed at DOE/LLNL; and POET, a transformation language together with an empirical optimization engine developed at UTSA. This framework permits different levels of automation and programmer intervention, from fully-automated tuning to semi-automated development to fully programmable control. The research targets both the optimization needs of computational kernels and the more general requirements of whole program optimizations. The framework is integrated as an external development mechanism for the widely-adopted ATLAS library and is connected with empirical tuning research under DOE SciDAC program to improve the efficiency of large-scale scientific applications.