GRL-RR: A Graph Reinforcement Learning-based resilient routing framework for software-defined LEO mega-constellations

Abstract

In the emerging Low Earth Orbit (LEO) mega-constellations, intelligent routing strategies that integrate Software-Defined Networking (SDN) and Deep Reinforcement Learning (DRL) demonstrate enhanced control over data transmission. However, existing schemes have poor scalability and robustness when facing large network scales and frequent failures. To address this, we propose a resilient routing framework called GRL-RR, specifically designed for software-defined LEO mega-constellations. Initially, we model network failure scenarios using cascade failure theory, analyzing network isolation caused by random failures and traffic overloads due to network attacks from both topological and routing strategy dimensions. Furthermore, we formulate SDN control domain partitioning and resilient routing as optimization problems under corresponding constraints. Subsequently, for the partitioning of resilient control domains, GRL-RR employs a benchmark topology template approach, enabling rapid division of control domains through rectangular constraints. Lastly, to compute resilient routing strategies, GRL-RR introduces a traffic topology model within each control domain to transform physical topology changes caused by various failures into traffic diversity. Through a link selection algorithm focused on critical failure scenarios, GRL-RR utilizes a Graph Neural Networks (GNN)-based DRL algorithm to control critical links, achieving scalable and robust routing. Simulations on mega-constellation topologies, such as Starlink and OneWeb, demonstrate that GRL-RR outperforms other existing resilient routing schemes in various link failure scenarios, improving path reliability by over 12.45%, reducing maximum link utilization by at least 12.92% while maintaining end-to-end latency within an acceptable range.