Low priority data transfer across the wide area is useful in several contexts, for example for the dissemination of large files such as OS updates, content distribution or prefetching. Although the design of such a service is reasonably easy when the underlying network supports service differentiation, it becomes more challenging without such network support. We describe an application level approach to designing a low priority service — one that is ‘lower than best-effort’ in the context of the current Internet. We require neither network support nor changes to TCP. Instead, we use a receive window control to limit the transfer rate of the application, and the optimal rate is determined by detecting a change-point. We motivate this joint control-estimation problem by considering a fluid-based optimisation framework, and describe practical solutions, based on stochastic approximation and binary search techniques. Simulation results demonstrate the effectiveness of the approach...
Peter B. Key, Laurent Massoulié, Bing Wang