RL-MR: Multipath Routing Based on Multi-Agent Reinforcement Learning for SDN-Based Data Center Networks

With the advent of the 5G era and the rapid development of emerging industries such as cloud computing and big data, data centers, as resource management platforms and key transmission hubs, are experiencing a dramatic increase in business data volume. Therefore, efficient routing optimization algorithms are needed. Improvements in Software Defined Networks (SDN) and programmable network devices allow different data traffic to be transmitted through various network paths, enabling fine-grained network performance optimization. In this paper, we utilize the emerging Hybrid Knowledge-Defined Network (KDN) architecture to propose a Multi-Agent Reinforcement Learning (MARL)-based multipath routing algorithm, termed RL-MR. RL-MR organizes intelligent agents to generate routes in a hop-by-hop manner, offering excellent scalability. The algorithm comprehensively considers flow latency, flow transmission rate, and packet loss rate to formulate routing strategies. To ensure reliability and accelerate the learning process, we introduce an auxiliary learning mechanism into RL-MR. Experiments conducted using Ryu and Mininet demonstrate that the RL-MR algorithm significantly outperforms baseline methods in terms of network throughput, link bandwidth utilization, and latency. Additionally, it exhibits excellent adaptability and reliability in scenarios involving flow changes, unknown large topologies, and partial deployments.