An Approach to Architecture - Independent Parallel Programming

这个项目解决了编写同时清晰、可移植和高效的并行程序的问题。最初的重点是MIMD多处理器，包括各种共享内存和分布式内存体系结构。该方法将并行算法的表达与获得良好性能所必需的依赖于机器的细节分开。程序员首先以经典的、显式并行的风格编写算法，假设有无限的并发并且对象是共享的。然后用调度指令和数据映射来注释这个基本算法，以产生一个有效的程序。因此，注释会影响算法在给定体系结构上的性能，但不会影响其正确性。一种新的、面向对象的编程语言已经开发出来，它支持这种方法。这个项目正在通过检查额外的并行计算问题，进行实验来评估其性能，并确定这种方法可以应用于现有语言和SIMD体系结构的程度，来改进语言和方法。

This project addresses the problem of writing parallel programs that are simultaneously clear, portable, and efficient. The initial focus is on MIMD multiprocessors, including various shared memory and distributed memory architectures. The approach separates the expression of a parallel algorithm from the machine-dependent details that are necessary to achieve good performance. A programmer first writes an algorithm in a classic, explicitly parallel style assuming there is unlimited concurrency and that objects are shared. This basic algorithm is then annotated with scheduling directives and data mappings to yield an efficient program. The annotations thus affect the performance of the algorithm on a given architecture but they do not affect its correctness. A new, object-oriented programming language that supports this method has been developed. This project is refining the language and approach by examining additional parallel computing problems, conducting experiments to assess its performance, and determining the extent to which this approach can be applied to existing languages and to SIMD architectures.