Performance Improvement for Parallel Message Passing Applications

Abstract

The message passing interface (MPI) application is a parallel application that solves a problem by combining various communication routines with computation. Due to the synchronization between nodes, running this program on a heterogeneous platform would result in undesirable idle times. In other words, this may lead to a computational imbalance in the system. The increase in slack time between nodes, on the other hand, wastes the computing power of the computing nodes. A dynamic load balancing algorithm is proposed in this paper, which is applied to the parallel Jacobi application. It reduces slack times by using some information gathered from a parallel application dynamically. The obtained results speedup the execution time of an application by 15.40, 12.38, and 9.54 for problem sizes 8000, 6000, and 4000 respectively, compared to the serial execution time.