New Dynamic Switch Migration Technique Based on Deep Q-learning
Lin Yao, Jia Li, Guowei Wu, Bin Wu · 2021
By decoupling the control and data planes, Software-Defined Networking (SDN) can implement centralized manage-ment on the network. With the increasing scale of the network, the multi-controller SDN architecture is becoming more and more popular, because it can handle what SDN with a single controller is not able to address. However, the controller load imbalance may happen due to traffic dynamics in an SDN with multiple controllers, which results in congestion at a certain controller and seriously affects the scalability of the control plane. Though switch migration is an effective solution to this problem, how to migrate the traffic is an NP-hard problem. In this work, we propose a switch migration scheme based on deep Q-learning (DQN) by combining the powerful perception of deep learning with the decision-making ability of Q-learning. We first describe the SDN state formally. Then, the network state is represented by the two-dimensional array as the input of the Q network. The network features are extracted through the convolution layer, and the full connection layer is achieved. Finally, the output layer to predict the migration action in some states of a network is extracted. After the migration action is performed, we will get an instant reward or penalty. We implement our algorithm based on the keras deep learning framework and compare it with the classic Q-learning algorithm. The results show that our scheme is superior to the traditional method in terms of resource utilization and load balancing ability.