E-Stream: An Energy-Efficient Approach to Short-form Video Streaming in Mobile Devices

Streaming short-form videos has become ubiquitous among mobile Internet users in recent years, but one troublesome issue is the rapid energy consumption of the mobile devices. Our measurement shows that one common smartphone can only support an average of 6 hours of video playback and some phone models are even less than 3 hours. The energy consumption is positively correlated with the video bitrate, so how to find the optimal tradeoff between the energy and the QoE is the major challenge. This work addresses this challenge by proposing E-Stream, an energy-efficient approach specifically for short video streaming. It hosts a neural network to determine the bitrate and the prefetch order in an energy-efficient manner. To balance the decision performance over the wide network environments, the neural network is trained by a novel curriculum-learning-based deep reinforcement learning method. Extensive evaluations show that E-Stream not only reduces the energy consumption by 27.1% ~ 48.2% (lengthening the streaming duration by 3.29 hours on average), but also can keep high QoE performance and robustness.