Reducing Multimedia Decode Power using Feedback Control Zhijian Lu, John Lach, Mircea Stan Dept. of Electrical and Computer Engineering University of Virginia Charlottesville, VA 22904 {zl4j, jlach, mircea}@virginia.edu Kevin Skadron Dept. of Computer Science University of Virginia Charlottesville, VA 22904 skadron@cs.virginia.edu Abstract Despite recent advances, battery life continues to be a limiting factor in mobile multimedia systems. Significant energy savings can be achieved by adapting systems at run- time to match the execution requirements of different tasks. This paper introduces an on-line dynamic voltage/frequency scaling (DVS) feedback technique that reduces voltage and frequency to match the playback rate. A PI controller ad- justs the decoder’s speed to keep constant the occupancy of the buffer between the decoder and the display, effec- tively matching the average decode rate to the display rate without the need for any off-line profiling. MPEG simula- tion results show that this technique reduces decoder power consumption while providing strong real-time guarantees. 1. Introduction Mobile multimedia systems are becoming popular con- sumer items, but limited battery life continues to be a ma- jor problem. Energy-efficiency, however, must be balanced against the fact that users demand a high quality of service (QoS). This paper considers energy efficiency for multimedia playback, in which a stream of multimedia frames must be decoded before display, and frames that miss their decode deadlines must either be delayed or dropped. Therefore, the frame decode rate must keep up with the defined display rate to avoid choppy playback. Of course, the decode time for each frame in a multime- dia stream is not necessarily uniform. For example, MPEG frames come in three different coding types (intra (I), bi- directional (B), and predictive (P)), each of which requires different decoding effort. Even within these coding types, frame decode time varies. Therefore, a simple constant- rate decoder—running at a speed that enables the worst- case frame to meet its deadline—actually decodes many frames well before their deadlines, creating a great deal of slack that can be reclaimed for energy savings by slowing down the processor. Using dynamic voltage/frequency scal- ing (DVS), the system can run slower with less power con- sumption for frames that require less decoding computation. However, this frame information is normally not known be- fore the decode is performed, except at the crude level of whether a frame is of type I/B/P. This paper introduces an online feedback control tech- nique for DVS in mobile multimedia systems that makes the average frame decode rate the same as the display rate. The key observations are that a buffer between decode and playback provides protection against incorrect DVS settings due to unforeseen changes in decoding complexity, and that the occupancy of this display buffer becomes a natural con- trol object for the system: a draining buffer indicates that the decoding rate is too slow, and a growing buffer indicates that more energy can be saved by scaling down the decoding rate. Although our controller is designed by assuming con- tinuous frequency and voltage scaling, our simulation re- sults show that it works well for discrete frequency/voltage settings. The advantages of this buffer-based feedback-control ap- proach over prior solutions are that no pre-playback or server-side profiling is required and that slack can be re- claimed across frame boundaries. No explicit frame decode time prediction is needed, thus avoiding the missed dead- lines caused by prediction errors. We apply formal feed- back control techniques to effectively control the buffer oc- cupancy. The maximum buffer size needed is about 10 de- coded frames, which is inexpensive in common hand-held devices [10]. Simulation results demonstrate that our feed- back control method approaches the energy savings achiev- able by the optimal scaling method, and up to 20% en- ergy reduction beyond the best on-line technique [6] we are aware of. Moreover, our method provides strong real-time guarantees, for a zero deadline miss rate, even without any prior frame information. The remainder of the paper is organized as follows. Re- lated work is discussed in Section 2. A real-time model of