Dynamic Load Balancing in a Geophysics Application Using STAPL

Abstract

Seismic wavefront simulation is a common method to understand the composition of earth below the surface, especially for hydrocarbon exploration. One of these simulation methods is the wavefront construction algorithm. In this thesis, we reduced the load imbalance in a parallel implementation of the wavefront construction algorithm. We added a generic redistribution framework for data structures in the C++ parallel library STAPL. We present a redistribution algorithm for the parallel wavefront construction application which uses the recursive coordinate bisection method to find a near-optimal data distribution of the data. This algorithm leveraged the added redistribution features in STAPL to improve the running time of our application. We compared the run time of the application with and without redistribution on different geophysics models. We show that the proposed redistribution provides up to 9.45x speedup on a Cray XE6m cluster and 11.85x speedup on an IBM BlueGene/Q cluster.