A game-theoretic approach towards congestion control in communication networks

Most of the end-to-end congestion control schemes are "voluntary" in nature and critically depend on end-user cooperation. We show that in the presence of selfish users, all such schemes will inevitably lead to a congestion collapse. Router and switch mechanisms such as service disciplines and buffer management policies determine the sharing of resources during congestion. We show, using a game-theoretic approach, that all currently proposed mechanisms, either encourage the behaviour that leads to congestion or are oblivious to it. We propose a class of service disciplines called the Diminishing Weight Schedulers (DWS) that punish misbehaving users and reward congestion avoiding well behaved users. We also propose a sample service discipline called the Rate Inverse Scheduling (RIS) from the class of DWS schedulers. With DWS schedulers deployed in the network, max-min fair rates constitute a unique Nash and Stackelberg Equilibrium. We show that RIS solves the problems of excessive congestion due to unresponsive flows, aggressive versions of TCP, multiple parallel connections and is also fair to TCP.