Using load information in work-stealing on distributed systems with non-uniform communication latencies

We evaluate four state-of-the-art work-stealing algorithms for distributed systems with non-uniform communication latenices (Random Stealing, Hierarchical Stealing, Cluster-aware Random Stealing and Adaptive Cluster-aware Random Stealing) on a set of irregular Divide-and-Conquer (D&C) parallel applications. We also investigate the extent to which these algorithms could be improved if dynamic load information is available, and how accurate this information needs to be. We show that, for highly-irregular D&C applications, the use of load information can significantly improve application speedups, whereas there is little improvement for less irregular ones. Furthermore, we show that when load information is used, Cluster-aware Random Stealing gives the best speedups for both regular and irregular D&C applications.