Feedback Guided Dynamic Loop Scheduling: Algorithms and Experiments

Dynamic loop scheduling algorithms can suuer from overheads due to synchronisation, loss of locality and small iteration counts. We observe that timing information from previous executions of the loop can be utilised to reduce these overheads. We introduce two new algorithms for dynamic loop scheduling which implement this type of feedback guidance, and report experimental results on a distributed shared memory architecture. Under appropriate circumstances, these algorithms are observed to give signiicant performance gains over existing loop scheduling techniques.