A new approach for task clustering
Weiping Zhu, Tyng‐Yeu Liang, Ce-Kuen Shieh · 2002
Static clustering has been used to group tasks for parallel processing. Most clustering methods used for current multithreaded DSM systems only consider the workload balance. In contrast, we present a static method to cluster closely related tasks of an application onto a multithreaded DSM system. This method relies on the Hopfield neural network to find optimal or near-optimal clusters. An optimal solution identified by this method tends to minimize load imbalance and communication overhead. We have implemented this method on Cohesion which is a multithreaded DSM system. Three programs, SOR, Nbody, and Gaussian elimination, are used to test the effectiveness of this method. The result shows that our method indeed can find optimal or near-optimal clustering for these programs.