Markov Clustering (MCL) based thread grouping and thread selection

Computing workload distribution is indispensable for resource sharing, cycle stealing and other modes of interaction in distributed systems and grids. Computations should be arranged to adapt the capacity variation of system resources. Although computational migration is the essential mechanism for move computing tasks around, the decision making of which task should be relocated is even more critical, especially when multithreaded parallel programs are involved. Multiple threads might be treated as a partial workload and moved together. Based on thread similarity and the Markov cluster algorithm (MCL), this paper proposes a novel thread grouping algorithm, Markov cluster grouping (MCG), to classify threads into hierarchical thread groups, some of which can be picked by the proposed thread selection scheme, Markov cluster selection (MCS) for load distribution. Experimental results demonstrate the effectiveness of the grouping algorithm for parallel workload distribution.