Insensitive Traffic Splitting in Data Networks

Bonald et al. have studied insensitivity in data networks assuming a fixed route for each flow class. If capacity allocation and routing are balanced and the capacity of a given class is shared equally between the flows, the network state distribution and flow level performance are insensitive to any detailed traffic characteristics except the traffic loads. In this paper, we consider optimal insensitive load balancing executed at packet level so that the traffic of each flow may be split over several routes. Similarly to the case with fixed routing, the most efficient capacity allocation and traffic splitting policy can be determined recursively. We formulate the problem as an LP problem using either a set of predefined routes or arbitrary routes and present numerical results for two toy networks. Traffic splitting gives a clear performance improvement when compared to flow level balancing or fixed shortest path routing.