As an important component of computer system, GPU has been used more widely in the system under the support of general computing. In addition to focusing on its performance, the issues of its energy consumption and environmental problem have gradually attracted the concerns of researchers, computer architects, and developers. Current researches only consider single-task scheduling for saving energy, lacking the focus on energy saving from scheduling the overall tasks. In view of the shortcomings of current researches, we propose a METS (Minimizing Execution Time Slot) approach to reduce energy by rationally allocating the tasks across GPUs. It first collects the number of tasks and the corresponding estimated performance information. Next, it decides whether to turn the problem into a 0–1 knapsack problem or to use FIFO method based on the number of tasks. Then, we conduct our experiment on typical platform to verify our proposed approach. The experimental results show that METS can save on average 8.43% of energy when compared with the existing approaches. This shows that the proposed METS method is effective, reasonable and feasible.
