An evolutionary multipath routing algorithm using SDN

Abstract

The centralisation of the control plane in the Software Defined Network (SDN) architecture allows the implementation of a globally optimised multipath routing algorithm, something that is not possible in traditional distributed networks. In our previous work, we have already shown a performance improvement made possible by SDN when using Linear Programming (LP) to find the optimal routing solution as compared to OSPF and ECMP. However, while multipath routing may be required to utilise a network to its maximum capacity, this may cause packet re-ordering, which negatively affects TCP performance. To overcome this limitation and improve TCP performance at a flow level, in this work we replace the LP routing algorithm with a multi-objective genetic algorithm to solve the Multi Commodity Flow problem. This also resolves the scalability problem in LP, and additionally allows us to choose good solutions that offer the smallest number of flow splits whilst maintaining a good compromise in terms of the total network flow and the total cost of the selected paths. While the chosen solution may not offer the best overall network usage, it improves performance at the flow level because the packet reordering problem is reduced.