A scheduling strategy for shared memory multiprocessors

An efficient scheduling strategy for shared memory multiprocessors is described. The rapid dissemination of tasks to available procesors and ready queues is crucial to the performance of any parallel system. Such overheads determine the attainable speedup and performance of the system. Poor techniques used to address this can lead to severe degradation in performance particulary with high processor counts. This work has been conducted in the context of a parallel functional language-CoF, where the parallelism is usually fine grained and the efficient assignment of tasks to processors even more important. In such systems, observing strict queue semantics (i.e., FIFO) is not essesntial. This allows for very efficient algorithms such as that described here. On the BBN GP1000, our technique was superior in performance to the centralized queue and has the potential of performing well on a fully configured GP1000.