Hierarchical Work-Stealing

dynamic load-balancing on hierarchical platforms. In particular, we consider applications involving heavy communications on a distributed platform. The work-stealing algorithm introduced by Blumofe and Leiserson is a commonly used technique to balance load in a distributed environment but it suffers from poor performance with some communication-intensive applications. We describe here several variants of this algorithm found in the literature and in different grid middlewares like Satin and Kaapi. In addition, we propose two new variations of the work-stealing algorithm : HWS and PWS. These algorithms improve performance by considering the network structure. We conduct a theoretical analysis of HWS in the case of fork-join task graphs and prove that HWS reduces communication overhead. In addition, we present experimental results comparing the most relevant algorithms. Experiments on Grid’5000 show that HWS and PWS allow us to obtain performance gains of up to twenty per cent when compared to the classical workstealing algorithm. Moreover in some cases, PWS and HWS achieve speedup while classical work-stealing policies result in speed-down.