Communication Optimization on Automatic Program Parallelization
Dong Chun-Li, Zhao Rong-Cai, Ma Zhuo-Jie, Li Dong-Hong · 2007
Minimizing communication by increasing the locality of data references is an important optimization for achieving high performance on distributed memory machines. But in the progress of decomposition, reorganization is inevitable. And the communication produced by reorganization is inevitable too. In this paper, the authors present a linear decomposition algorithm that automatically finding computation and data decomposition, including finding data and computations decomposition that has data reorganization communication. And the authors improve the method and reduce the communication cost by merging parallel regions with the same data decomposition.