— Efficient use of network resources has long been an important problem for large-scale network operators. To this end, several recent research efforts have proposed automated methods for optimizing routes based on traffic measurements. However, these efforts have not considered the stability of the dual feedback control mechanisms of adaptive routing and congestion control, when operating together. In this paper, we demonstrate that an important class of adaptive routing algorithms can yield stable optimal routes in the presence of congestion control, provided that either the congestion control mechanism is fair or the network workload behaves under reasonable constraints. We further show that one or the other of these assumptions is necessary for this class of adaptive routing algorithms – otherwise, unstable, sub-optimal routes may result in some pathological cases.
Eric Anderson, Thomas E. Anderson