Aggregate Power Consumption Modeling of Live Video Streaming Systems Yousef O. Sharrab and Nabil J. Sarhan Electrical and Computer Engineering Department & Wayne State Media Research Lab Wayne State University Detroit, MI 48202 {yousef.sharrab,nabil}@wayne.edu ABSTRACT Power consumption of video streaming systems has become a ma- jor concern, especially in battery-powered devices, such as video sensors. Power is usually dissipated in each one of the major phases of the streaming process: capturing, encoding, and transmission. This paper develops models for power consumption in each of these phases and validates them with extensive experiments, focusing primarily on H.264 video encoding. For comparative purposes, we also study MJPEG and MPEG-4 video codecs. In addition, we an- alyze the impacts of the main H.264 video compression parameters on power consumption and bitrate. These parameters include quan- tization parameter, number of reference frames, motion estimation (ME) range, and ME algorithm. Keywords Power Consumption Modeling, H.264, Streaming Video, Video Capturing, Video Encoding, Video Transmission. 1. INTRODUCTION Power consumption has become a major concern in live video streaming systems, especially those employing battery-operated de- vices, such as automated video surveillance and wireless sensor networks. In such systems, prolonging the battery lifetimes is a pri- mary objective due to its great implications in terms of system cost and availability. In such video streaming systems, energy is con- sumed at the source in each of the three main phases: capturing, encoding, and transmission. Power consumption at the receivers (such as monitoring stations) may be important but to much lower degrees. In a streaming environment, the three phases operate in a pipelined fashion. This paper analyzes power consumption in the capturing, en- coding, and transmission phases. It develops models for all these phases and then validates them individually and in terms of the aggregate power consumption. The developed models are based on 500 different experiments, each of which is repeated at least 3 times, totaling more than 1, 500 experiments. Prior studies fo- cused primarily on one aspect/phase. In particular, study [9] de- veloped a power consumption model for video capturing for a spe- cific on-chip vision circuit. For wireless video sensor networks Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page.To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. MMSys’13, February 26-March 1, 2013, Oslo, Norway. Copyright (c) 2013 ACM 978-1-4503-1894-5/13/02 ... $15.00 (WVSNs), transmission power consumption has been analyzed in [3, 12]. Encoding power consumption has been studied by [11] and [20]. Study [11] developed a Power-Rate-Distortion (P-R-D) framework for a generic video encoder (that applies to H.263), but it did not analyze the effects of spatial resolution and frame rate. Paper [20] measured the power consumption of an H.263 encoder, but the experiments were limited to QCIF resolution and a 10 fps frame rate. Moreover, it did not consider varying the spatial and temporal resolutions and other encoding parameters. None of these two papers analyzed H.264 or considered the capturing power con- sumption. H.264 has many new features and algorithms, especially in intra-prediction and inter-prediction. In addition, the impacts on power consumption of important parameters (such as quantization parameter, number of reference frames, search range, and motion estimation algorithms) were not analyzed in prior work (up to our knowledge). Other work on encoding includes developing a cross- layer approach in [4] to tradeoff between coding and communi- cation power consumptions. Furthermore, much work considered improving video encoding, such as using a statistical approach to reduce the computation times of the most computationally inten- sive components of video coding [10], early detection of all-zero integer transform coefficients [31], and a hexagon-based search (HEXBS) pattern for fast motion estimation (ME) [34]. This paper has been motivated by our ongoing work on the power- aware design of automated video surveillance systems, which re- quires accurate, simple, and appropriate power consumption mod- els. The main contributions of this paper can be summarized as follows. (1) We analyze the three main phases in live video stream- ing systems and provide accurate and simple power consumption models. (2) For video encoding, we develop a model for the H.264 standard, considering important factors, such as mode selection, the number of reference frames, and sub-pixel ME search. The devel- oped full mode selection model can be used to find the total number of operations required by H.264 encoding and the number of oper- ations of each type (add, multiply, divide, etc.). (3) We analyze the impacts of important encoding parameters on power consumption, including quantization parameter, number of reference frames, ME algorithms, and ME range. Since tuning the parameters is often based on an energy and bitrate tradeoff, we develop models for the bitrate as well. (4) We compare the power consumption of each phase and study other encoders, including MPEG-4 and MJPEG, for comparative purposes. (5) We show that the overall computa- tion complexity for all phases can approximately be modeled as a linear function of the pixel rate. The pixel rate is the product of the spatial and temporal resolutions of the raw video. (6) We conduct extensive experiments using two different types of video cameras and streaming systems. The rest of the paper is organized as follows. Section 2 discusses