Some further investigation on maximum throughput: Does network coding really help?

Network coding has been shown to be the solution that allows to reach the theoretical maximum throughput in a capacitated telecommunication network [1]. It has also been shown to be a very appealing and practical alternative to routing-based approaches to send traffic from sources (servers) to terminals (clients) for many different applications. However, the initial theoretical claim of throughput benefit remains relatively unclear, mainly because the multicast throughput maximization problem is difficult to solve (it is closely related to the fractional Steiner tree packing problem which is NP-hard). In this paper, we show that these optimization problems are still tractable even for instances with a significant size (up to 50 nodes and 300 edges). We also propose and solve the multicast maximum throughput problem with an additional constraint on the number of multicast trees. We apply our algorithms on large sets of randomly generated instances, mainly based on bidirected graphs, because they are the most relevant to model fixed telecommunication infrastructures. The main result of our intensive experimental study is that, in practice, network coding does not increase throughput compared to traditional multicast. Instances showing a throughput gain can only be generated somewhat artificially by imposing some structure or trying to maximize the throughput gap. However, when we limit the number of multicast trees, then, most of the times, very significant throughput gaps appeared. Since management constraints often impose on network administrators a very limited use of multicast trees, network coding appears clearly as a very nice alternative for delivering content to customers.