PROCESSOR LOAD ANALYSIS FOR MOBILE MULTIMEDIA STREAMING: THE IMPLICATION OF POWER REDUCTION (ThuPmOR6)

Author(s) : Min Li (Zhejiang Univ., China) Xiaobo Wu (Zhejiang Univ., China) Zihua Guo (MSRA, China) Richard Yao (MSR, United States of America) Xiaolang Yan (Zhejiang Univ., China)

Abstract : Mobile multimedia streaming has been shown as one of the killer applications for 3G and beyond. Since recently released mobile processors have great processing capability, the software codec, which brings a lot of flexibility and reconfigurability, becomes feasible and popular on most mobile devices. However, the software codec incurred significant power consumption, because the energy efficiency of general processor based system is much lower than that of the dedicated hardware such as ASIC. Dynamical Voltage Scaling (DVS) is one of the most efficient techniques to promote the energy efficiency, and the key point is to lower down the supply voltage when the processor is not fully loaded. Most existing papers on this topic use simple heuristics to predict processor load, and poor prediction accuracy is observed in experiments. We advocate intensive analysis on processor load before designing DVS framework and algorithm. Hence, we conduct load analysis on more than 600 processor load trace files from 57 test sequences (different encoding parameters) and 98 representative clips from Internet. Basic statistical analysis and time series analysis are applied in order to identify major characteristics of the processor load. The analysis shows that it is feasible to predict processor load using low order linear time series model if the load is sampled using feature period, Moreover, there is indeed significant potential to reduce the energy consumption. Based on the analysis results, we develop a fully adaptive DVS framework and a set of intelligent algorithms to adjust supply voltage online with guaranteed penalty. Accurate prediction is achieved with standard error of deviation below 7%, and more than 50% energy reduction is achieved when streaming CIF test sequences.

Menu