Amber: Programming Support for Networks of Multiprocessors

本研究关注的是在多处理器网络上构建高度并行应用程序的编程系统的设计和实现。基于多处理器的共享内存多处理器正变得广泛可用，并有望提供经济高效的高性能计算。明年，多处理器工作站肯定会得到广泛应用：人们预计，对于许多应用程序，它们甚至比大型机的同类产品更具成本效益，就像目前的单处理器工作站通常比单处理器大型机更具成本效益一样。简单地说，这项研究的目标是提供一种对多处理器网络进行编程的工具，就好像它是一台大型的、集成的、共享内存的计算机一样。这项工作需要研究操作系统结构的几个方面。首先，要寻找为表示并行算法提供适当抽象级别的编程原语。在集成并行性和分布性的系统中，重要的权衡是在不牺牲本地性能的情况下提供跨网络的统一访问。其次，必须探索操作系统组织中的替代方案，以提高单个多处理器节点上的性能。这里的目标是为并行执行提供最低开销的机制，以便程序员可以表达和利用中粒度的并行性。第三，必须开发工具，帮助程序员评估其应用程序的性能，并根据网络多处理器环境对其进行适当优化。这个项目是我们过去8年来一直参与的分布式和并行计算系统研究项目的自然产物。本研究项目得到了美国国家科学基金会的多项资助；它还得到了工业界的大力支持，特别是数字设备公司，其Firefly原型多处理器工作站将成为我们的实验工具。（DECSystems研究中心已经为华盛顿、斯坦福、麻省理工、普林斯顿、多伦多和剑桥提供了萤火虫，这使得这些大学在进行中等规模多处理的实验研究方面装备精良。）我们在支持并行计算方面的初步工作中已经获得的结果被许多大学站点和Sequent、Digital、Microsoft和Tera等公司积极使用。

This research concerns the design and implementation of a programming system for building highly parallel applications on networks of multiprocessors. Multiprocessor based shared memory multiprocessors are becoming widely available and promise to provide cost effective high performance computing. Multiprocessor workstations will surely be widely available within the next year: one expects them to be even more cost effective than their mainframe counterparts for many applications, just as current uniprocessor workstations often are more cost effective than uniprocessor mainframes. Simply stated, the goal of this research is to provide a facility for programming a network of multiprocessors as if it were a large, integrated, shared memory computer. This effort requires study of several aspects of operating system structure. First, there is a search for programming primitives that provide the right level of abstraction for expressing parallel algorithms. The important tradeoff, in a system integrating parallelism and distribution, is providing uniformity of access across the network without sacrificing local performance. Second, one must explore alternatives in operating system organization for increasing performance on a single multiprocessor node. The objective here is to provide the lowest overhead mechanisms for parallel execution so that programmers can express and utilize medium grained parallelism. Third, one must develop tools that aid the programmer in assessing the performance of his or her application, and in properly optimizing it for an environment of networked multiprocessors. This project is a natural outgrowth of a program of research in distributed and parallel computing systems in which we have been involved for the past 8 years. This program of research has been supported by NSF under a variety of awards; it also has received significant support from industrial sources, particularly Digital Equipment Corporation, whose Firefly prototype multiprocessor workstations will be our experimental vehicle. (The DECSystems Research Center has supplied Fireflys to Washington, Stanford, MIT, Princeton, Toronto, and Cambridge, making these universities uniquely well equipped to conduct experimental research on medium scale multiprocessing.) The results that we have already obtained in our preliminary work on support for parallel computing are in active use by many university sites and by companies such as Sequent, Digital, Microsoft, and Tera.