Data Distribution Calculating Algorithm to Eliminate Redundant Communications
Jie Zhao, Rongcai Zhao, Yuan Yao, Xiaoxian Liu · ASME Press eBooks · 2012
Some redundant communications are imported in parallel compilation to guarantee the consistency between computation partition and data distribution. To solve this problem, this paper proposed a data distribution calculating algorithm. Based on dependence testing analysis, the algorithm recognized the parallelizable loops, implemented calculating data distribution by analyzing the loop indexes and array subscripts so as to eliminate redundant communications. The experimental results show that using the results of the algorithm to generate MPI communications code can reduce the redundant communications significantly, and the performance of optimized programs can be improved to 70 percent of the manual programs.