Hierarchical service chain orchestration for multi-cloud environments enabled by deep reinforcement learning

Abstract With the rapid adoption of multi-cloud platforms, dynamic orchestration of service function chains faces coupled challenges. This study proposes a hierarchical service chain orchestration for multi-cloud environments, which decomposes end-to-end quality of service into real-time, economic, and stability objectives. This model uses hierarchical orchestration structure and integrates Actor-Critic networks with Graph Isomorphism Network and attention mechanism. The bandwidth allocation layer optimizes bandwidth coefficients for deadline-sensitive tasks, while the server selection layer processes dynamic effective subgraphs for cost-aware server deployment. The dynamic filtering strategy and online scheduling scheme significantly improve computational efficiency and dynamic load adaptability. Validated on a real-world cloud platform, our method achieves remarkable performance improvements across different load scenarios. Compared with the top-performing baseline model, it attains a comprehensive score enhancement of 2.56%, 3.44%, and 16.8% under idle, normal, and stress states, respectively. Meanwhile, it consistently outperforms the baseline model in terms of decision-making speed. The model enables dynamic service chain orchestration in multi-cloud, and breaks through the limitations of traditional centralized and distributed orchestration.