可变结构的并行计算中任务粒度细化可扩展方法

Extension by Refining Task Granularity for Parallel Computation with Variable Structures

摘要:

首先评估并行任务及体系结构中影响可扩展性的关键因素，并对并行任务及体系结构进行图建模.然后，提出一种DAG任务粒度细化的可扩展方法，本质上是变换图的结构、调整图节点权值和边权值.进一步推导得出一些关于新扩展方法的有用结论.最后，应用网格模拟工具SimGrid开展实验，结果表明所提出的扩展方法，能实现可变结构并行计算的等速度效率扩展，对于并行计算扩展实践有指导意义.

Abstract:

Aiming at such extension problem in parallel computation, this paper evaluates the key factors from parallel tasks and architecture which affect the scalability, and then models parallel tasks as well as architecture by the weighted graph. Especially, we propose the extension method of refining task granularity to realize an extension in parallel computation. The extension method transforms the graph’s structure and adjusts the weights of its nodes and edges in essence. Additionally, by further derivation, some significant conclusions about the new extension methods are drawn. Finally, the simulative experiments are conducted on the platform SimGrid to verify the effectiveness of the proposed extension methods. The results show that the new methods can realize isospeede extension in parallel computation with variable structures, which is helpful for its practical extension.