Traffic-splitting networks operating under alpha-fair sharing policies and balanced fairness

We consider a data network in which, besides classes of users that use specific routes, one class of users can split its traffic over several routes. We consider load balancing at the packet-level, implying that traffic of this class of users can be divided among several routes at the same time. Assuming that load balancing is based on an alpha-fair sharing policy, we show that the network has multiple possible behaviors. In particular, we show that some classes of users, depending on the state of the network, share capacity according to some Discriminatory Processor Sharing (DPS) model, whereas each of the remaining classes of users behaves as in a single-class single-node model. We compare the performance of this network with that of a similar network, where packet-level load balancing is based on balanced fairness. We derive explicit expressions for the mean number of users under balanced fairness, and show by conducting extensive simulation experiments that these provide accurate approximations for the ones under alpha-fair sharing.