Modeling and analysis of seed sheduling strategies in a BitTorrent network

BitTorrent has gained momentum in recent years as an effective means of distributing digital content in the Internet. Despite the remarkable scalability and efficiency properties that characterize BitTorrent in the long haul, several studies identify the source of the content as the main culprit for the poor performance of the system in a transient regime where user requests for a popular content swamp the source and in case of high node churn. Our work models the scheduling decisions made at the source (called the seed) for selecting which pieces of the content to inject in the system through a stochastic optimization process and provides an analytical framework to compare different strategies. We define a new piece selection algorithm (called proportional fair scheduling, PFS) that incorporates the seed's limited vision of the system dynamics in terms of user requests so as to ensure a better content distribution among the users. We prove convergence of PFS and compare its short and long term performance against the mainline BitTorrent implementation and the "smart seed" technique recently introduced in [9]. Our results show that PFS induces substantial improvements on both system performance, by decreasing the download time at the users, and system robustness against peer dynamics, by quickly reacting to sudden changes in the request patterns of the users.