Towards Class-Based Dynamic Voltage Scaling for Multimedia Applications

As more and more of computing has become mobile, and thus reliant on battery or solar power, reducing energy consumption has become critical. One significant consumer of energy in any computer system is the CPU. To reduce the CPU energy consumption, many CPUs now allow dynamic scaling of the CPU voltage, known as DVS. As a linear reduction in CPU voltage leads to a quadratic reduction in its energy consumption [1], this approach is very attractive. Nevertheless, reducing the voltage also entails reducing the CPU frequency, and thus increasing the computation time, which is unacceptable for many applications. One kind of application for which increasing the computation time can be acceptable is a video player. In practice, a video player only needs to meet its frame rate; any further increase in performance gives no benefit for the user. Furthermore, while decoding some frames may require the full computing power of the host platform, others are often much more simple to decode, providing intervals in which the CPU frequency can be reduced with no loss in performance. Video has thus been an attractive target for DVS algorithms [2,3,4,5]. Nevertheless, most of these approaches require either modifying the operating system process scheduler or the video codec, and thus they have not achieved wide use in practice. In previous work, we have proposed a DVS algorithm appropriate for use in the context of video kiosks, where we exploit the fact that the video is played over and over to identify the best CPU frequency for each frame based on its history of execution [6]. In this paper, we begin to consider how to extend this approach to the more common case, where a video is played only once, on heterogeneous platforms. For this, we propose to classify frames according to a model of their computation requirements and then to use this classification on an arbitrary platform to predict the treatment time of arbitrary frames at various CPU frequencies, based observation of the treatment time of a few frames in each class.