[5H2-OS-18a] OS-18

12:45 PM - 1:00 PM JST(3:45 AM - 4:00 AM UTC)

[5H2-OS-18a-04] Constrained Reinforcement Learning for Microservice Placement Optimization in the Edge–Cloud Continuum

〇Hirotaka Kasahara¹, Ryuichi Kitajima¹, Osamu Sato¹ (1. SoftBank Corp.)

Keywords:

Edge Computing,Graph Neural Network,Combinatorial Optimization,Reinforcement Learning

In recent years, low-latency processing has become increasingly important in applications such as smart cities and autonomous driving, leading to growing interest in the edge–cloud continuum. In such environments, microservices must be placed across geographically distributed and heterogeneous computing resources while satisfying multiple constraints and minimizing operational costs. This microservice placement problem is known to be NP-hard, making real-time decision-making difficult for conventional exact solvers in large-scale systems. This paper proposes a constrained reinforcement learning framework for microservice placement in the edge–cloud continuum. The problem is formulated as a constrained Markov decision process, and multiple hard constraints are handled using PPO-Lagrangian. Our simulation results demonstrate that the proposed method consistently outperforms genetic algorithms and greedy heuristics. The proposed approach achieves up to 20–30% cost reduction while maintaining high feasibility across different graph sizes and topologies. Moreover, our method computes a service placement within a few tens of milliseconds even for large-scale scenarios, indicating its suitability for real-time orchestration in practical edge–cloud systems.