Data Distribution Strategies for Domain Decomposition Applications in Grid Environments

In this paper, we evaluate message-passing applications in Grid environments using domain decomposition technique. We compare two domain decomposition strategies: a balanced and unbalanced one. The balanced strategy is commonly strategy used in homogenous computing environment. This strategy presents some problems related with the larger communication latency in Grid environments. We propose an unbalanced domain decomposition strategy in order to overlap communication latency with useful computation. This idea consists in assigning less workload to processors responsible for sending updates outside the host. We compare the results obtained with the classical balanced strategy. We show that the unbalanced distribution pattern improves the execution times of domain decomposition applications in Grid environments. We considered two kinds of meshes, which define the most typical cases. We show that the expected execution time can be reduced up to 53%. We also analyze the influence of the communication patterns on execution times using the Dimemas simulator.