When Locality is not enough: Boosting Peer Selection of Hybrid CDN-P2P Live Streaming Systems using Machine Learning

Live streaming traffic represents an increasing part of the global IP traffic. Hybrid CDN-P2P architectures have been proposed as a way to build scalable systems with a good Quality of Experience (QoE) for users, in particular, using the WebRTC technology which enables real-time communication between browsers and, thus, facilitates the deployment of such systems. An important challenge to ensure the efficiency of P2P systems is the optimization of peer selection. Most existing systems address this problem using simple heuristics, e.g. favor peers in the same ISP or geographical region. We analysed 9 months of operation logs of a hybrid CDN-P2P system and demonstrate the sub-optimality of those classical strategies. Over those 9 months, over 18 million peers downloaded over 2 billion video chunks. We propose learning-based methods that enable the tracker to perform adaptive peer selection. Furthermore, we demonstrate that our best models, which turn out to be the neural network models can (i) improve the throughput by 22.7%, 14.5%, and 6.8% (reaching 89%, 20.4%, and 24.3% for low bandwidth peers) over random peer, same ISP, and geographical selection methods, respectively (ii) reduce by 18.6%, 18.3%, and 16% the P2P messaging delay and (iii) decrease by 29.9%, 29.5%, and 21.2% the chunk loss rate (video chunks not received before the timeout that triggers CDN downloads), respectively.