Conditionally Optimal Parallelization for Global FP on Multi-core Systems

Abstract

Throughout the last decade, the importance of parallel computing has risen greatly to match the ever-increasing computational demand. Frameworks such as OpenMP and OpenCL allow easy parallelization of computing tasks into desirable number of threads, opening up a chance to greatly utilize the parallel computing resources. We call this "parallelization freedom". However, this does not come for free, as parallelization overhead increase with parallelization option (i.e. the number of thread each task is parallelized). Thus parallelization option must be carefully decided to better utilize a given computing resource. This paper addresses the problem of assigning parallelization option to each task for global FP scheduler. For this, we extend the approaches made by Cho, which is limited to the global EDF scheduler case. We prove that a conditionally optimal parallelization assignment of parallelization option also exists for the global FP case. Through extensive simulations, we show a significant improvement of schedulability.