Kernel scheduling approach for reducing GPU energy consumption

Abstract

To handle the increasing data scale in broad fields, GPU (Graphic Processing Unit) is integrated with more and more cores to provide powerful computing capability. To obtain high performance, the feature of concurrent kernels is taken by GPU vendors to fully tap the performance of GPU. Although existing GPU supports concurrent kernels to improve the performance, its improvement in performance demands a sustained increasing power. In execution of concurrent kernels, the hardware utilization of GPU resources, the category of kernels and the order of concurrent kernels issued to GPU affect the energy consumption. By keeping eye on above factors, this paper presents a kernel scheduling approach for reducing energy consumption (KSRE) to relieve the energy problem. Firstly, KSRE extracts the features of several kernels to be executed and then classifies them according to their features. Secondly, it obtains potential energy saving effect of concurrent kernels by using the off-line trained regression model. At the end, KSRE schedules kernels based on above information to save energy. To validate the effectiveness of proposed energy model, the experiments are performed and their outcomes show that KSRE is not only effective but also can save energy with performance enhancement as compared to previous works.