Abstract

We consider a problem which is partially motivated by the BitTorrent protocol. Suppose that a large file is initially available only at some server and we desire to disseminate it amongst N end users in the least possible time. We suppose that the file is divided in M parts of equal size and that an end user may download a part from either the server or one of the peers who has previously downloaded it. Given constraints on the rates at which the server and peers can upload to one another, we seek a strategy that minimizes the time until all peers have downloaded the entire file. Assuming that there is a centralized controller who does the scheduling of the uploads, we show how to solve this problem by solving a finite number of linear programs. Making connections with the so-called ‘broadcasting problem’, we provide a complete solution of this problem when the capacities are all equal. We also carry out simulations to assess the performance of a natural randomized algorithm that operates under distributed control. We compare the performance of this algorithm in two different information scenarios with what can be achieved by centralized control.

Citations