Evaluation of on-the-fly auto-tuning of hybrid parallelization on processors with integrated graphics
Akiyoshi Wakatani · 2019
On processors with integrated graphics, both parallelization of plural processing cores of CPU and parallelization of GPU can be exploited simultaneously. However, the optimal load balancing of CPU and GPU is hard to be determined, because performance estimation of CPU and GPU depends on the characteristics of applications. In order to cope with this problem, we propose an on-the-fly auto-tuning method, which determines the optimal load balancing in runtime. Our method, in advance, finds a ratio (AutoRatio) of preliminary calculation used for achieving the best performance of hybrid parallelization, and the effectiveness of our approach is empirically confirmed by using four applications.