Optimal load balancing in insensitive data networks

Bonald et al. have recently characterized a set of insensitive dynamic load balancing policies by modelling the system as a Whittle network. In particular, they derived optimal "decentralized" strategies based on limited state information and evaluated their performance in simple example networks. In this paper, we consider the specific case of a data network where each flow can be routed on one of a set of alternative routes. By using the linear programming formulation of MDP theory we are able to analyze optimal routing policies that utilize the full global state information. In the ordinary LP formulation of MDP theory, the global balance condition appears as a linear constraint on the decision variables. In order to retain insensitivity, we impose stricter detailed balance conditions as constraints. As a further extension, the MDP-LP approach allows joint optimization of the routing and resource sharing, in contrast to the earlier work where the resource sharing policy was required to be separately balanced and fixed in advance. The various schemes are compared numerically in a toy network. The advantage given by global state information is in this case negligible, whereas the joint routing and resource sharing gives a clear improvement. The requirement of insensitivity still implies some performance penalty in comparison with the best sensitive policy.