RLMR: Reinforcement Learning Based Multipath Routing for SDN
Chao Chen, Feifan Xue, Zhengyong Lu, Zhongyun Tang, Chuanhuang Li
Abstract
In recent years, as a new subject in the computer field, artificial intelligence has developed rapidly, especially in reinforcement learning (RL) and deep reinforcement learning. Combined with the characteristics of Software Defined Network (SDN) for centralized control and scheduling, resource scheduling based on artificial intelligence becomes possible. However, the current SDN routing algorithm has the problem of low link utilization and is unable to update and adjust according to the real‐time network status. This paper aims to address these problems by proposing a reinforcement learning‐based multipath routing for SDN (RLMR) scheme. RLMR uses Markov Decision Process (MDP) and Q‐Learning for training. Based on the real‐time information of network state and flow characteristics, RLMR performs routing for different flows. When there is no link that meets the bandwidth requirements, the remaining flows are redistributed according to the Quality of Service (QoS) priority to complete the multipath routing. In addition, this paper defines the forward efficiency (FE) to measure the link bandwidth utilization (LBU) under multipath routing. Simulation results show that compared with the current mainstream shortest path algorithm and ECMP algorithm, the routing algorithm in RLMR has advantages in FE, jitter, and packet loss rate. It can effectively improve the efficiency and quality of routing.