LCA: Deep Reinforcement Learning-Based Congestion Avoidance Routing Model in SDN

Abstract

With the rapid advancement of network technologies, ensuring optimal performance across diverse applications has become increasingly challenging, particularly regarding latency and throughput. Treating online video streaming and file transfers equally can lead to congestion and inefficient resource utilization. The real-time optimization of forwarding paths for traffic flows with varying QoS requirements remains inadequately addressed. Traditional algorithms often struggle with congestion control, as they operate independently of routing protocols, making them inflexible and difficult to adjust. To overcome these challenges, we propose LCA, a Deep Reinforcement Learning-based Congestion Avoidance Routing Model for SDN. LCA dynamically adapts to diverse application needs using a sophisticated reward function to optimize routing and ensure differentiated QoS guarantees. To enhance adaptability and reliability, we introduce a Fast-Secure Validation mechanism that prevents unsafe routes, ensuring rapid convergence and reducing management overhead. LCA also integrates a congestion severity index to assess congestion levels and prevent network collapse, along with a quality of experience factor to maintain QoS. Additionally, LCA incorporates a traffic classification phase to differentiate mice-elephant flows, assigning appropriate queues to meet class-of-service requirements. Implemented within a RYU-Docker-based SDN architecture, LCA features a Live-QoS Monitor and DNC Classifier, addressing real-time traffic demands across ten service classes. Performance evaluations demonstrate that LCA outperforms state-of-the-art algorithms, achieving 5%–10% lower delay, 5%–20% lower packet loss, and 25%–30% reduced jitter under congestion. These results highlight LCA’s effectiveness in ensuring QoS, making it a promising solution for modern SDN environments.