Evaluation of a Novel Technique for Determining Near Optimal Task Granularity in Parallel Programs
Simon Söderlind · KTH Publication Database DiVA (KTH Royal Institute of Technology) · 2015
The evolution of computers is moving more and more towards multi-core processors and parallel programs. For parallel programs to perform effciently the task granularity has to be near optimal for that specific problem. The aim of this thesis is to evaluate a novel technique, called Control of the Critical Path Delay (CCPD), for determining a near optimal granularity for parallel programs, and determine when the technique succeeds, why it performs as it does and if there is a way to improve the technique. The CCPD-technique showed promising on problems where non-optimal scheduling did not play any part. Where the non-optimal scheduling affected the problem the CCPD-technique did not perform well. To deal with potentially poor scheduling a method to improve the technique by incorporating more information from the critical path curve is proposed. Experimental results on a shared memory system as well as theoretical analysis are presented.