Research Initiation Award: Compilation of Data-Parallel Programs for Scalable Shared-Address Space Multiprocessors

9309054 HUDAK共享地址分布式内存大规模并行处理(MPP)系统提供了单个地址空间的简单性，同时保持了万亿次浮点运算的潜力。然而，为了在这些系统上获得合理的效率，程序员或编译器必须指定并适当地在处理器和物理分布的存储器之间分配工作和数据。这个项目将解决为给定的并行应用程序确定最佳分布的问题。正在进行的MPP编译研究已经产生了一系列技术，用于规范单个数据并行循环的工作和数据分布。该项目将把这些技术与数据并行程序的流程图表示法结合起来，以生成具体说明工作和数据分布的创建和更改的代码。分布将被设计为跨程序构造的最大效率，用于改变程序的控制流，诸如条件语句(例如，IF语句)和过程调用。这项研究将导致开发一个名为DPF的软件系统，这是一个新的系统，它将分析数据并行程序的划分，并为面向对象的数据并行编程提供一个有效的环境。数据并行程序可以从用户定义的对象构建，通过DFT编译器自动调整每个对象的过程，以在现有的共享地址间隔多处理器系统上实现高性能。该研究将极大地提高MPPS的可用性、并行程序的可移植性和用户所达到的性能水平。***

9309054 Hudak Shared-address distributed-memory massively parallel processing (MPP) system offer the simplicity of a single address space while maintaining the potential for teraflops performance. However, to obtain reasonable efficiency on these systems, the programmer or compiler must specify and appropriate distribution of work and data among the processors and the physically distributed memories. This project will address the problem of determining the optimal distributions for a given parallel application. Ongoing research in compilation for MPP's has yielded a collection of techniques for the specification of work and data distributions for individual data-parallel loops. This project will combine those techniques with flow-graph representations of data-parallel programs to generate code which specifics the creation and alteration of work and data distributions. The distributions will be designed maximum efficiency across program constructs for altering the flow of control of a program, such as conditional statements (e.g., if statements) and procedure calls. The research will lead to the development of a software system named DPF+, a new system that will analyze an partition data- parallel programs as well as provide an efficient environment for object-oriented data-parallel programming. Data parallel program can be built from user-defined objects, with the DFT+ compiler automatically tuning each object's procedures for high performance on an existing shared-address spaced multiprocessor system. This research will greatly increase the usability of MPPs, the portability of parallel programs and the performance level achieved by users. ***