A COMPARISON OF CLUSTERING HEURISTICS FOR SCHEDULING DIRECTED ACYCLIC GRAPHS ON MULTIPROCESSORS

Clustering of task graphs has been used as an intermediate step toward scheduling parallel architectures. In this paper, we identify important characteristics of clustering algorithms and propose a general framework for analyzing and evaluating such algorithms. Using this framework, we present an analytic performance comparison of four algorithms: Dominant Sequence Clustering (DSC) (Yang and Gerasoulis, Proc. Supercomputing '91, 1991, pp. 633-642) and the algorithms of Kim and Browne (Int. Conf. on Parallel Processing, 1988, Vol. 3, pp. 1-8) Sarkar (Partitioning and Scheduling Parallel Programs for Execution on Multiprocessors, MIT Press, 1989), and Wu and Gajski (J. Super-comput. 2 (1988), 349-372). We identify the common features and differences of these algorithms and explain why DSC is superior to other algorithms. Finally, we present some experiments to verify our analysis. © 1992.