Network-decentralized robust congestion control with node traffic splitting

Abstract

We consider a traffic control problem defined on a network graph, whose nodes represent buffers and whose arcs represent flow channels. We consider network models with a peculiar aspect: each element of the flow arriving at each node must be redirected towards a precise other node of the network, hence each buffer is naturally split in several queues, characterized according to statistics about the flow splitting at the nodes. Precisely, each node is modelled as a Markov chain, in which some states are specifically associated with the arcs leaving the node: state j represents the amount of traffic waiting to be directed through arc j. We show that such a network can be stabilized by means of a network-decentralized control, in which the flow through each arc is controlled by an agent which only knows the congestion situation at the nodes it connects. The main result is that the proposed network-decentralized strategy is robust (namely it assures stability under all possible values of the Markov chain parameters) provided that zero is a simple eigenvalue for all the Markov chains, which includes the irreducible case.