Achieving topology generalization without retraining in service function chaining through cooperative multi-agent reinforcement learning

The accelerating growth in mobile devices, along with advancements in technologies like 5G, IoT, and cloud computing, has driven a rising demand for flexible and diverse network services. Service Function Chaining (SFC) has emerged as a pivotal technique to address these demands, leveraging the capabilities of Software Defined Networking (SDN) and Network Function Virtualization (NFV) to enable dynamic service deployment. However, deploying SFCs across dynamic and heterogeneous network topologies remains a formidable challenge, as current solutions are often constrained by their reliance on specific network structures. In this paper, we propose Topo-G, a novel framework for SFC deployment, built upon Cooperative Multi-Agent Reinforcement Learning (MARL), designed to achieve topology generalization under a “train once, run anywhere” paradigm. Unlike existing methods that necessitate retraining for new network topologies, Topo-G is inherently adaptable, efficiently generalizing across diverse topologies without retraining. The framework leverages a decentralized partially observable Markov decision process (Dec-POMDP) to decouple the routing and placement tasks while preserving their interdependence. By incorporating a shortest-path routing algorithm and a candidate pool selection method for VNF placement, Topo-G efficiently coordinates the decision-making process, independent of the underlying network topology. Extensive experiments demonstrate that Topo-G significantly outperforms existing methods in both deployment efficiency and adaptability across various topologies. Additionally, when retraining is required, Topo-G achieves a marked improvement in convergence speed, further underscoring its potential to enhance the scalability and flexibility of SFC deployment in dynamic, large-scale networks.