Abstract-With the increasing popularity of Software defined network (SDN), designing a scalable SDN control plane becomes a critical problem. An effective approach to improving the scalability is to design distributed architecture of SDN control plane. However, how to evaluate the scalability of SDN control planes remains unexplored. In this paper, we propose a metric of scalability for SDN control planes, and study three typical SDN control plane structures, including centralized, decentralized and hierarchical architectures. We build performance models for response time, based on which we evaluate the scalability of these three structures. Furthermore, the comparison between different architectures are analyzed by mathematical methods. Numerical evaluations are also conducted to validate the conclusions drawn in this paper.
I. INTRODUCTIONWith the development of the Internet, the traditional network is approaching a bottleneck in network control. To tackle the challenge, a new network architecture named SDN is proposed which moves the control plane out of data plane and reduces the management complexity significantly. Taking advantage of centralized control, the controller for SDN is customized, programmable, and flexible. However, the centralized control faces some challenges in large scale networks. The first SDN controller (NOX) [1] can serve only 30,000 flow initiation requests per second while keeping the response time less than 10ms. It is challenge to serve more flows while keeping the response time within a reasonable small duration. Thus, understanding and quantifying the scalability of the SDN controller is a critical problem for successful adoption of SDN for large scale networks or networks with many flows. A SDN control plane is scalable if the control plane architecture will adapt to handle more network event requests with the increasing complexity and scale of network while satisfying the quality of service.In order to improve the scalability of SDN control planes, two approaches were proposed [1]. The first is to move some control function of the controller to switches, reducing the event requests submitted to the controller significantly. Thus, the scalability of the controller is improved. However, the switches may need highly specialized application-specific integrated circuits(ASICs) to satisfy these requirements, and general-purpose CPUs may also be required to supply some control function. Therefore, this approach needs to be embraced by major manufacturers and thus increases the dif-