User behavior anticipation in P2P live video streaming systems through a Bayesian network

In recent years, Peer-to-Peer (P2P) architectures have emerged as a scalable, low cost and easily deployable solution for live video streaming applications. In these systems, the load of video transmission is distributed over end-hosts by enabling them to relay the content to each other. Since end-hosts are controlled by users, their behavior directly impact the performance of the system. To understand it, massive measurement campaigns covering large-scale systems and long time periods have been performed. In this paper, we gathered and synthesized results obtained through these measurements and propose a Bayesian network that captures and integrates all of them into a synthetic model. We apply this model to the anticipation of peer departures which is an important challenge toward the performance improvement of these systems and especially churn resilience. The validation of our proposal is performed through intensive simulations that consider a streaming system composed of thousand users over two hundred days. We especially study two deployment scenarios: a system-scale one and a local one. We also compare our proposal with two standard estimators and we show under which conditions an estimator outperforms the others.