TCP Congestion Control in Fast Long-Distance Networks

TCP congestion control is currently based on Jacobson’s algorithms (slow-start and congestion avoidance), devised back in 1988, with improvements such as SACK and NewReno. Although TCP has proved remarkably flexible thus far and managed to adapt to vastly different types of networks, we show in this paper that some of the assumptions behind these algorithms are no longer valid in today’s fast long-distance networks. Our analysis is backed by empirical data collected over a 622 Mbit/s transoceanic link dedicated to Grid applications. We propose that the additive increase and multiplicative decrease algorithms be changed for this type of networks, and describe a new fairness principle.