Heuristic tuning of parallel loop performance

Abstract

A closed-loop feedback approach to tuning the performance of parallel programs across several executions is presented. It involves extracting relevant information from an execution trace, building a history of executions, and using the accumulated knowledge to guide scheduling of future executions. In particular, the technique is applied to minimizing parallel loop execution time by adjusting the number of iterates issued to each processor requesting work--the chunk size. The heuristic developed chooses a chunk size after three complete loop executions and monitors further executions for significant changes.