Dr. Akiyoshi Wakatani1
1Konan University, Kobe, Japan
Performance tuning is to control several parameters of a system to accelerate the performance of the system. For numerical simulations using a large size computation, performance tuning is one of the most important techniques as well as high-speed computers. Some recent processors have GPU as well as plural processing cores of CPU, so both the parallelization of plural processing cores of CPU and the parallelization of GPU can be exploited simultaneously.
Since the optimal load balancing of CPU and GPU is hard to be determined, in the prior art, the authors proposed an on-the-fly auto-tuning method, which determines the optimal load balancing in runtime by using the pre-computation of fixed data size (AutoRatio).
However, in order to cope with the case that the value of AutoRatio is not determined in advance, we propose a new auto-tuning method that determines a load balancing in runtime by repeating small size pre-computations. By determining the optimal size of the pre-computation incrementally, this method does not require to determine the size of precomputation in advance. The effectiveness of our approach is empirically confirmed by using four applications.
Akiyoshi Wakatani received the B. Eng. degree from the Department of Applied Mathematics and Physics, Faculty of Engineering, Kyoto University, Kyoto, Japan, in 1984. He received the M. Eng. degree from the Division of Applied Systems Science, Faculty of Engineering, Kyoto University in 1986. He also received the Dr. Eng. degree from the Division of Information Engineering, Faculty of Engineering, Kyoto University in 1996. He was with Matsushita Electric Industrial (currently Panasonic) from 1986 to 2000, as a researcher. From 2000 to 2006, he was an Associate Professor of the Department of Information Science and Systems Engineering, Faculty of Science and Engineering, Konan University, Kobe Japan. Since 2006, he has been an Full Professor of the same university. His research interest includes parallel processing and programming education.