Session Throughput Prediction for Internet Videos

Abstract

Predicting network throughput of video sessions using large-scale wide area measurements can enable throughput-aware initial bit rate selection and the best content servers selection. In this article, using a unique dataset from a commercial VoD system, we first dissect the impact of network throughput on user engagement, and show that the improper selection of initial bit rates, due to unavailability of network throughput estimation, leads to a high likelihood of join failure. We then study the throughput prediction problem in Internet video systems and propose a hybrid approach that combines several heuristics. Trace-driven experiments show that the hybrid approach achieves high prediction accuracy and large coverage. The prediction approach can be leveraged for improved video quality and user experience.