论文部分内容阅读
并行系统中各结点间的互联网络是高性能计算机的一个关键研究领域。30多年来传统上各种互联网络是以逻辑电路为基础所构成的〔1〕。然而,系统结构及其概念应随计算机工艺的进展而变化。本文提出一种新型的并行系统结构,它采用多端口快速静态存储器作为各结点机之间的互联机制MCIM。与传统的逻辑电路互联网络相比,MCIM可以减少MPP系统中的消息传递延迟;可以克服网络并行计算系统NPC中网络适配卡总线传递速率的瓶颈;可以大量减少网络协议开销。在当前VLSI工艺条件下,MCIM的实现是低价高效的。本文讨论了MCIM并行系统结构,通信路径的仲裁与选择。本文还阐述了MCIM仿真工具,给出了实验结果
The interconnection between nodes in a parallel system is a key area of research for high-performance computers. For more than 30 years, various types of Internet have traditionally been based on logic circuits [1]. However, the structure of the system and its concept should vary with the progress of the computer process. This paper presents a new type of parallel system architecture that uses multi-port fast static memory as the interconnect mechanism MCIM for each node. Compared with the traditional logic network, MCIM can reduce the message delivery delay in MPP system. It can overcome the bottleneck of the network adapter card bus transfer rate in the network parallel computing system NPC. It can greatly reduce the network protocol overhead. Under the current VLSI process conditions, the implementation of MCIM is cost-effective. This article discusses the MCIM parallel system architecture, communication path arbitration and selection. This article also describes the MCIM simulation tools, the experimental results are given