Improving Cache Utilization of Nested Parallel Programs by Almost Deterministic Work Stealing

Nested (fork-join) parallelism eases parallel programming by enabling high-level expression of and leaving the mapping between tasks hardware to runtime scheduler. A challenge in dynamic scheduling nested is how exploit data locality, which has become more demanding deep cache hierarchies modern processors with a large number cores. This paper introduces almost deterministic work stealing (ADWS) , efficiently exploits locality deterministically planning cache-hierarchy-aware schedule, while allowing little variety facilitate load balancing. Furthermore, we propose an extension our prior on ADWS achieve better shared utilization. The improved version scheduler called xmlns:xlink="http://www.w3.org/1999/xlink">multi-level ADWS . idea that only part computation whose working set size small enough fit into scheduled within recursively, thus avoiding excessive capacity misses. Our evaluation benchmark decision tree construction demonstrated multi-level outperformed conventional random Cilk Plus 61% it showed 40% performance improvement over previous design.