Design and Implementation of an Energy Efficient Multimedia Playback System ∗ Zhijian Lu † , John Lach † , Mircea Stan † , Kevin Skadron ‡ , Departments of † Electrical and Computer Engineering and ‡ Computer Science, University of Virginia Charlottesville, VA 22904 {zl4j, jlach, mircea}@virginia.edu, skadron@cs.virginia.edu Abstract Mobile devices capable of multimedia playback have be- come popular consumer items, making techniques for en- ergy management during multimedia decoding increasingly important. In this paper, we model the multimedia decoding process as a discrete-time system excited by random input sequences representing the incoming stream and therefore, unlike many existing techniques, do not make any assump- tions about the workload streams. Using this novel stochas- tic process model, we can apply formal methods to analyze the decoding system and design a feedback-based on-line dynamic voltage/frequency scaling (DVS) algorithm to ef- fectively reduce the energy consumption during multimedia playback. We implemented our technique in a laptop com- puter equipped with a DVS enabled processor, and results reveal good performance for real-world video clips across a wide range of video compression formats compared with existing techniques. 1. Introduction While the continuously increasing computation power provided by technology scaling enables more complex mo- bile applications, energy consumption becomes a limiting factor. Since multimedia playback is among the dominant applications in mobile devices, it is important to design en- ergy efficient multimedia playback systems. Fortunately, due to the variations in multimedia decoding complexity, the required computation power changes during the play- back, and dynamic voltage/frequency scaling (DVS) is a powerful technique to exploit this opportunity to reduce runtime power by lowering down the circuit speed (and thereby power) whenever possible. However, DVS multi- media systems should be carefully designed to avoid caus- ing large degradation in playback quality. In this paper, we systematically analyze the design of such a system, propose a new design technique and validate our design through a prototype system on a DVS enabled laptop computer. * This work is supported in part by the National Science Foundation under grant Nos. EHS-0410526, CCF-0429765, CCR-0133634 (Career) and EHS-0509245. Designing a good multimedia playback system using DVS involves trade-offs among multiple contradicting ob- jects and is subject to practical constraints. Techniques pro- posed in [5, 6, 10] require perfect predictions for decode execution timing, which is not possible in some multimedia compression formats. The methods in [7, 8] exploit buffer- ing to improve the energy efficiency of DVS. However, smaller buffer sizes might increase the frame deadline miss rate, while larger buffer sizes not only increase hardware costs but also increase the playback latency, which is unac- ceptable in many real-time applications. In general, a prac- tical solution has to seek the best trade-off between power consumption, playback quality and hardware resources and, in general, has to assume no specific information about the incoming multimedia streams. 0 10 20 30 40 50 60 70 80 90 0.2 0.4 0.6 0.8 1 1.2 Normalized cpu clock frequency Power (W) Figure 1. MPEG decoding power at different CPU speed. The DVS design presented in this paper provides a good balance between all these competing factors. We first cre- ate a model for the available design space. Then we search the design space with two desired properties in mind: speed schedule uniformity and system stability. Power consump- tion is a convex function of circuit speed. As an example, in Figure 1, the MPEG decode power consumption of a laptop computer is shown at different CPU speeds (i.e., frequency and voltage settings). Though each frame can be decoded at a different speed due to computation variations, decoding multiple frames at a uniform speed (i.e., the average de- coding speed) provides better energy efficiency according to Jensen’s inequality [9]. Therefore, in an energy efficient