Efficient distribution of a parallel job across different grid sites

Abstract

The study for parallel computing application between grid sites reveals three conclusions. Using a graph partitioning based on block distribution between grid sites gives lower communication time compared to random block distribution. This can be inferred by just looking at total interface size of the mesh blocks distributed between grid sites or actual time measurement that also includes communication latency and wait time. As the number of mesh partitions increases, the elapsed time difference between the random- and graph-partition-based mesh block distributions reduces. Because of increased number of neighbors, wait time between the mesh blocks across the grid sites would be dominating the total elapsed time. Increasing number of grid sites would be even more costly compared to two grid sites because interdependency would result in more wait time. As an outcome of this research study, the conclusion will be that the bottleneck is communication speed and algorithm among grid sites for a parallel job. With the existing approaches, the parallel computing over the grid sites would be more beneficial.