Switch Migration Based Load Balancing for Software Defined Network Multiple Controllers Using Reinforcement Learning

Abstract

SDN is a new computer networking paradigm that is built on the concept of separating the control plane and data plane to simplify network management through central network programmability and quick innovation. Because network traffic fluctuates over time and space, a static mapping between switches and controllers results in unequal load distribution among controllers. Multiple controller deployment in SDN is a viable technique to improve the SDN controller's reliability and scalability. However, it introduces a new issue of load imbalance between multiple controllers due to dynamic network traffic changes and imbalanced incoming load allocation across the controllers. This leads to load imbalance issues between controllers, causing some controllers to become overloaded and some others. Existing dynamic switch migration approaches are mostly focused on load balancing of distributed SDN controllers with switch migration schemes, and they are based on threshold definition for load judgments, which results in migration overhead. We use online Q learning with the constraints of maximum efficiency and no migration conflicts to obtain global optimal controller load balancing at the lowest cost. Our work resulted in better outcomes in a variety of scenarios with varying load distribution on switches. We demonstrate the convergence behavior of the proposed scheme to the optimal policy using simulation results. The result shows the designed scheme outperforms better in terms of load balancing rate and switch migration exchange and achieves 19% improvements compared to RLB and 12% compared to SAR-LB in load balancing rate. Also resulted in a maximized cumulative reward of the system