CAREER: Compiling for Heterogeneous Systems

异构并行系统在一台机器内或跨机器包含不同的并行模型。 这些系统为大型、多样化的应用提供高性能。不幸的是，异构化使软件开发复杂化。用户目前手动并行化每个架构的每个任务。这些程序不是可移植的，如果发生更改，必须重新编码（例如，添加或删除机器）。 用户还需要执行适当的优化以实现高性能。编译器可以解决这些问题。基于编译器的方法最重要的优点是： (1)可移植性和易用性：用途：用户为其算法编写适当的并行模型。编译器自动将它们映射到可用体系结构的具体细节。(2)高性能：编译器通过积极的、目标特定的优化进行优化，并协调各种硬件以实现出色的性能。 为了管理异构性，编译器必须：针对不同的体系结构，调整优化策略以适应目标，利用机器可用性的运行时变化，并允许语言发展。 本研究的目的是调查的编译系统，以支持这些要求。该系统将使用创新的编译器组织和表示，以提供异构性所需的灵活性。 本项目集中于此编译器基础设施的设计和实现。 该系统将被用来研究自动并行化和分区策略的异构目标。 它将解决诸如如何将程序员的并行模型映射到相应的机器上以及如何进一步优化其整体系统性能等问题。 这种基于编译器的方法将提供一个能够执行高级动态和静态优化的基础设施，异构系统的高性能需求。 ***

Heterogeneous parallel systems incorporate diverse models of parallelism within a single machine or across machines. These systems provide high performance for large, diverse applications. Heterogeneity unfortunately complicates software development. Users currently hand parallelize each task for each architecture. These programs are not portable and must be recoded if changes occur (e.g., machines are added or removed). The user is also burdened with performing the appropriate optimizations to achieve high performance. A compiler can solve these problems. The most important advantages of a compiler based approach are: (1) Portability and Ease of Use: Users program the appropriate model of parallelism for their algorithms. The compiler automatically maps them on to the specifics of the available architectures. (2) High Performance: The compiler optimizes with aggressive, target specific optimizations and coordinates diverse hardware to achieve excellent performance. To manage heterogeneity the compiler must: target diverse architectures, adjust optimization strategies to suit targets, exploit run-time changes in machine availability, and allow languages to evolve. The objective of this research is to investigate a compilation system to support these requirements. The system will use an innovative compiler organization and representations in order to provide the flexibility that heterogeneity requires. This project concentrates on the design and implementation of this compiler infrastructure. The system will be used to investigate automatic parallelization and partitioning strategies for a heterogeneous target. It will address issues such as how to map programmer models of parallelism on to the corresponding machine and how to further optimize it for total system performance. This compiler-based approach will provide an infrastructure capable of performing the advanced dynamic and static optimizations that heterogeneous systems demand for high performa nce. ***