A new analysis of Work Stealing with latency

Abstract

We study in this paper the impact of communication latency on the classical Work Stealing load balancing algorithm. Our paper extends the reference model in which we introduce a latency parameter. By using a theoretical analysis and simulation, we study the overall impact of this latency on the Makespan (maximum completion time). We derive a new expression of the expected running time of a bag of tasks scheduled by Work Stealing. This expression enables us to predict under which conditions a given run will yield acceptable performance. For instance, we can easily calibrate the maximal number of processors to use for a given work/platform combination. All our results are validated through simulation on a wide range of parameters.