Given a SDN topology, how many controllers are needed and where should they go?

Abstract

Software Defined Networking (SDN) presents a paradigm shift in network management and configuration. The idea of having an externalized control plane opens many unanswered questions regarding scalability, fault tolerance and performance of the controller. An important question that must be answered is, given a network topology, how many controllers are needed and where should they be placed to satisfy user-specific requirements and constraints. Such requirements range from latency constraints, failure tolerance and fair load distribution. These metrics compete with each other, thus no single best placement is available. In this paper, we focus on controller placement to minimize propagation latency (of control traffic) and CapEx associated with installing a new controller. We apply Silhouette Analysis and Gap statistics to compute the optimal number of controllers to use for a given topology. To determine the optimal locations to place the controllers, we use Partition Around Medoids (PAM) clustering algorithm. We evaluate our solution using the Internet2 topology and then expand our scope to over 10 publicly available WAN topologies. As expected, the answers to controller placement are topology-dependent. However, an evaluation of our algorithms on the Internet 2 topology, recommends two controllers as the optimal number of controllers to use. Surprisingly, our results indicate that one controller suffices to meet latency requirements (though certainly not reliability requirements). Finally, the techniques presented in this work can be extended to tackle other similar placement problems, such as baseband unit placement for 5G cloud radio access network (C-RAN) deployment and fog node placement which appears in the context of edge computing.