Performance Effect of Localized Thread Schedules in Heterogeneous Multi-Core Processors

We previously proposed a heterogeneous 16-core architecture and a nearest neighbor priority-based thread scheduling algorithm which values core affinity and permits inter-core thread migration in [1]. In this paper, we run simulation experiments under localized core affinities limited to a very small number of specific cores within core classes. These localized schedules represent applications that are user- or system-confined to specific cores or their vicinity. They may also represent stressful scenarios with high contention to specific - and possibly scarce - cores with unique functionality or computational might. The results indicate that under these localized initial core affinity scenarios, non-migratory thread scheduling algorithms perform much worse than non-affinity algorithms while the algorithms which consider core affinity and allow inter-core thread migration are best.