A Case Study on Addressing Complex Load Imbalance in OpenMP

Abstract

Load balance is an important factor that fundamentally impacts the scalability of any parallel application. In this paper we present a case study to address a complex load imbalance related to the convergence behavior of the parallel SPMD implementation of a GMRES solver used in a real world application in the field of computational fluid dynamics. In order to tackle this load imbalance in OpenMP we illustrate different approaches involving the use of nested tasks as well as nested parallel regions. Furthermore, we evaluate these approaches on a small kernel program extracted from the original application code and show how the load balance is affected by each of these approaches.