Cone Based Clustering for List Scheduling Algorithms

Abstract

List scheduling algorithms attempt to minimize latency under resource constraints using a priority list. We propose a new heuristic that can be used in conjunction with any priority function. At each time-step, the proposed clustering heuristic tries to find a best match between ready operations and the resource set. The heuristic arbitrates among equal priority operations based on operation-clusters formed from the dependency graph. Based on this heuristic we have presented a new Cone-Based List Scheduling (\CBLS\@) algorithm. Results presented in this paper compare \CBLS with the well-known Force Directed List Scheduling (\FDLS\@) algorithm, for several synthesis benchmarks. In cases where \FDLS produces sub-optimal schedules, \CBLS produces better schedules and in other cases \CBLS performs as good as \FDLS\@. Moreover, in conjunction with a simple priority function (namely the self-force of an operator), \CBLS results in considerable improvement in latency when compared to \FDLS that has the same priority function. Finally, we show that \CBLS with the simple priority function performs better in execution time as well as latency when compared to the original \FDLS that has a relatively complex priority function.