近程作用离散系统大规模并行模拟概念模型

Simulation of Discrete Systems with Local Interactions: A Conceptual Model for Massive Parallel Processing

通用大规模并行计算 (MassiveParallelProcessing ,MPP)的主要困难在于有关消息传递和内存共享的硬件必然随着处理单元的增加而复杂化。而实际上许多MPP课题都可抽象为对近程作用离散系统的模拟 ,它们具有能完全并行的统一的总体算法和纯粹的局部性。若建立其专用MPP系统 ,理论上能无限增加处理器数量而不使系统结构复杂化 ,而计算速度也将随处理器数目线性增长 ,从而获得最大的经济性。本文对这类课题的界定和算法及硬件系统的设计作了初步探讨 .

A major difficulty of massive parallel processing (MPP) is that the hardware for communications and memory sharing inevitably gets more and more complicated with the increase of processing units. But in fact and many application of MPP can be abstracted as simulations of locally interactive discrete systems that have a fully parallel and general algorithm with purely local operations. Theoretically, a specialized MPP machine for this purpose can be expanded forever without complicating the system architecture, and a linear increase of computational capacity with the number of processing units can be reached. This would be the most economic system we can expect. In this paper, the algorithm, system architecture and applicable fields for such specialization are explored preliminarily.