Stochastic Guarantees for Rate-Adaptive Streaming

A major challenge in mobile video streaming applications is the variability of the wireless channel. Techniques like Apple's HLS and MPEG-DASH typically offer the video in chunks where each chunk is made available with different levels of quality and bitrate. Adaptive clients use estimates of the past network conditions to select the source bitrate of the next chunk accordingly. Frequently, this selection is based on the average performance and does not consider the variability of the channel. In this work, we present a method for bitrate adaptation that uses the distribution of the past throughput to provide statistical performance guarantees obtained by the stochastic network calculus. We show how to select the bitrate so that a certain, small probability of buffer underflow is not exceeded. Theoretical results are derived for the general case of an arbitrary throughput distribution. Numerical results are included for the example of a Gaussian distribution. Additionally, we present simulation results that show the relevant performance metrics like average bitrate and buffer level of a system that adjusts its bitrate as suggested by our method. Generally, the method presented in this paper can achieve a bitrate that is close to the average available throughput with little variability of the source bitrate, while limiting the probability of buffer underflow to a predefined level.