Maximising Efficiency of Network Resources in SDN through Optimal Packet Routing using Reinforcement Learning

Abstract

Software-defined networking (SDN) has increased in popularity in recent years owing to its ability to centralize network device control. This is the cornerstone of data center networks, and as more IoT-enabled devices are added, it is expected to skyrocket. The employment of Machine Intelligence to handle and route the majority of traffic can enhance centralized SDN-based control, allowing for high and occasionally subpar network management. Owing to the development of Reinforcement Learning (RL) combined with Deep Learning, which allows autonomous flow management in SDN, intelligent routing has gained popularity. However, the existing SDN routing algorithm has poor link usage and is unable to update and change according to real-time network conditions, making it challenging to forecast and predict how communication networks will evolve in the future. Consequently, we present a routing optimization strategy based on Quality of Service (QoS) variables for SDN-based Data Center Networks using Deep Reinforcement Learning. We provide a reinforcement learning-based routing method for SDN that outperforms Dijkstra’s shortest path in terms of the throughput and average latency. In the proposed method, we integrated SDN with RL, which takes input from the state in which the current network is and routes traffic to achieve the maximum QoS parameters. The parameters considered were the throughput and delay. The presented solution is a network-oriented, self-learning routing system named DeepQoSR, based on Advantage Actor-Critic Networks. We compared our proposed technique with traditional algorithms, such as the shortest path, and present our analysis.