基于区域图数据流分析的通信优化算法

A Communication Optimization Algorithm Based on Data-Flow Analysis of Region Graph

减少通信开销对于并行化编译器生成高效的分布代码是非常重要的.首先提出了一个冗余并行执行模型(RPEM)作为通信优化算法生成的目标程序的执行模型,之后给出了区域图的概念和区域最大化算法,在最大化区域图的基础上进行数据流分析可以增大数据流分析粒度,提高分析的效率,同时也有助于通信的提前与合并.最后提出了一种基于区域图数据流分析的通信优化算法.该算法能够进行跨循环、跨过程的数据流分析,提高分析的精度,改善通信优化效果.实验结果表明,该算法对于通信量较大的程序能够有效地减少通信的次数和通信量,具有良好的可扩展性.

Reducing communication overhead is extremely important for parallelizing compiler to generate efficient codes for distributed-memory systems. In this paper, a redundant parallel execution model (RPEM) is proposed as an execution model for target programs optimized by the new algorithm. The region graph is introduced, and an effective algorithm is proposed to maximize the regions in the region graph. A region-based data-flow analysis algorithm is proposed to perform communication optimization. The overhead of data-flow analysis can be reduced by performing analysis on the maximized region graph. The coarse grain analysis also helps to communication lift up and aggregation. This communication optimization algorithm is able to perform inter-loop and inter-procedure analysis. Experimental results show that this algorithm is effective in reducing both communication volume and number of messages in programs with a large communication amount.