Service capacity of peer to peer networks

We study the 'service capacity' of peer to peer (P2P) file sharing applications. We begin by considering a transient regime which is key to capturing the ability of such systems to handle bursty traffic, e.g., flash crowds. In this context our models, based on age dependent branching processes, exhibit exponential growth in service capacity, and permit the study of sensitivity of this growth to system policies and parameters. Then we consider a model for such systems in steady state and show how the average delay seen by peers would scale in the offered load and rate at which peers exit the system. We find that the average delays scale well in the offered load. In particular the delays are upper bounded by some constant given any offered load and even decrease in the offered load if peers exit the system slowly. We validate many of our findings by analyzing traces obtained from a second generation P2P application called BitTorrent.