Load balancing highly irregular computations with the adaptive factoring

In heterogeneous environments, employing dynamic scheduling algorithms to improve performance of scientific applications via load balancing is essential. Presently, these algorithms require prior knowledge about workload via profiling resulting in higher overhead as problem sizes and number of processors increase. In addition, variations in work load at runtime may be unpredictable, making profiling work tedious and sometimes even obsolete. Therefore, dynamic loop scheduling schemes such as Factoring, Fractiling, andWeighted Factoring have been proposed and proved to be extremely instrumental when used in scientific applications such as Monte-Carlo simulations, N-Body simulations, radar applications, and others. Adaptive Factoring, a technique that evolves from these schemes and addresses a wider range of irregularities has recently been proposed. This paper reports on performance improvements obtained by integrating the Adaptive Factoring, into a scientific application that invloves computational field simulation on unstructured grids. Performance of this scientific application using the implementation with Adaptive Factoring is compared with implementations using other dynamic loop scheduling techniques. Reported experimental results confirm the benefits of using the Adaptive Factoring and its high potential for a successful integration in other scientific applications, especially the ones characterized by highly irregular behaviour whose performance degradation is primarily due to load imbalance.