Fair congestion control for long-lived background flows

We consider a new economic framework for designing congestion control algorithms for long-lived background flows. The algorithms aim at the fair sharing of the network capacity left over by the much shorter (and usually more interactive and delay sensitive) flows, between all competing background flows even if some use TCP in particular. The algorithm acts as a weighted variant of TCP with adaptive weight adjustment based on an estimate of the average throughput that would be obtained by a TCP flow following the same path, with the resulting throughput behavior ranging from low priority to that of TCP. It turns out that this design approximately maximizes the social welfare for the background traffic for appropriately defined utilities for bandwidth averaged over long time scales and a social cost function reflecting the negative externalities on the average download time of interactive flows, over all possible protocol designs. We illustrate our approach by proposing two concrete congestion control algorithms, each of them guaranteeing the throughput of TCP in the case of resource scarcity or resource abundance respectively. Two implementation alternatives are explored; one based on adjusting the congestion window ramp up speed and another based on alternating transmitting and idle periods. We use both analysis and simulations to validate our approach of implementing new background transfer protocols.