Communication Optimization Based on Elimination of Linear Inequalities
Xue-rong Gong · Jisuanji gongcheng · 2008
The parallel code that the compiler generates automatically leads to code expansion and unnecessary running time increase if it includes too much redundant code. This paper presents an optimization technique to eliminate redundant communication code by using Fourier-Motzkin- elimination algorithm for linear inequalities of computation decompositions and data decomposition. The test results show that the technique is able to reduce the size of the code by 10 to 30 percent compared with the original code, and the speedup achieves 1.12 on average using the same number of processors.