Scheduling and grain size control

Locality and storage management considerations for fork-join parallelism lead to the loss of the important List Scheduling property. The worst case performance of the constrained scheduling policies, called PaH and ToS, is very bad: (nearly) all parallelism is lost. “Practical” performance, however, differs radically from worst case performance. The PaH and ToS policies perform on average within 2 % of List Scheduling for a benchmark of Divide & Conquer applications on a wide range of architectures. A synthesised benchmark was constructed to create those schedules in which ToS might suffer great performance losses. Average results from these artificial applications are very encouraging: losses for “realistic” applications are at most 2 %. Highly improbable applications increase the losses to 3 %. We were prepared to trade performance in terms of load imbalance for improvements in data communication in the case of PaH, and additional improvements in storage management in the case of ToS. We have shown that the penalty in load imbalance is practically negligible.