Profiling Software Defined Networks for dynamic distributed-controller provisioning

Software Defined Networking decouples packet-forwarding from logically centralized network control. Conceiving network control as load-aware responsive distributed system remains a key challenge. In this paper, we propose a scheme for dynamically provisioning and placing network controllers based on their usage profiles. The proposed scheme analyze the amount of calls each controller is handling at the same time, and uses a Markov chain to identify if the network load is increasing, decreasing, or stable. Our controller placement algorithm chooses the most central nodes for installing the controllers. We simulate our proposal over real network topologies for a variable network load. We compare our proposal with a centralized-controller approach and with an approach where all nodes are also controllers. Our results show that the proposal achieves an efficient controller placement and provides as many controllers as needed to respond to the demand. The proposed scheme installs 50% controllers less than the distributed approach, while keeps the minimum load over each controller.