Automatic code generation of data decomposition
Ya-Nan Shen, Rongcai Zhao, Jianmin Pang · 2006
How to decompose or map data of programs automatically onto scalable parallel processors is a key issue in developing parallelizing compilers in DSM architecture. Data locality is crucial for parallelized programs to achieve high performance. Based on a linear inequalities mathematical model a formal specification of an optimized data decomposing algorithm and its implementation in C++ are presented. The algorithm enhances data locality and minimizes communication. Experimental results indicate that the algorithm improves the performance of parallelized programs significantly.