A Software Environment for High-Performance Parallel Computing

在高性能MIMD架构在生产科学、工程和商业应用中变得普遍之前，必须解决困扰当前系统的三个编程问题。首先，手工编写并行程序非常困难。程序员必须管理数十到数千个独立进程的通信、同步和调度。正确管理环境的负担经常让程序员不堪重负。其次，一旦在特定的MIMD体系结构上实现，生成的代码通常不能在其他MIMD体系结构上使用，这就需要大量的移植工作。这阻碍了用户并行化他们的代码，因为新的架构是如此频繁地被引入。第三，不支持在由各种高性能体系结构组成的异类“元系统”上透明地执行应用程序。这项提案解决了所有这三个问题。我们已经用Mentat攻击了前两个问题，Mentat是弗吉尼亚大学开发的一个易于使用的、面向对象的并行处理系统。我们计划在各种平台上的一组实际应用程序上测试Mentat方法的有效性。目标是确定Mentat是否易于在实际应用程序中使用，并衡量性能。我们解决第三个问题的方法是渐进的。我们将把Mentat扩展到一个异构环境中，并使用扩展的Mentat作为元系统实验的试验台。基于实验结果，将构建一个原型元系统，为用户提供易于使用的、高性能的抽象。

Before high-performance MIMD architectures become commonplace in production scientific, engineering, and business applications, three programming problems that plague current systems must be solved. First, writing parallel programs by hand is very difficult. The programmer must manage communication, synchronization, and scheduling of tens to thousands of independent processes. The burden of correctly managing the environment often overwhelms programmers. Second, once implemented on a particular MIMD architecture, the resulting codes are usually not usable on other MIMD architectures, necessitating a substantial effort to port. This discourages users from parallelizing their code, since new architectures are introduced so frequently. Third, there is no software support for transparently executing applications on heterogeneous "meta- systems" comprised of a variety of high-performance architectures. This proposal addresses all three of these problems. We have attacked the first two with Mentat, an easy-to-use, object- oriented, parallel processing system developed at the University of Virginia. We plan to test the efficacy of the Mentat approach on a set of real applications on a variety of platforms. The objective is to determine whether Mentat is easy to use on real applications and to measure performance. Our approach to the third problem is evolutionary. We will extend Mentat into a heterogeneous environment, and use the extended Mentat as a test- bed for meta-systems experimentation. Based on experimental results, a prototype meta-system that provides an easy-to-use, high performance abstraction for users will be constructed.