SAMD-SRG: Service-Aware Microservice Deployment Using Spectral Ranking and Graph-Enhanced Reinforcement Learning

Edge computing is crucial for latency-sensitive applications, making microservice deployment in edge environments a significant area of focus. However, microservice deployment still faces challenges in handling the intricate dependencies among microservices, optimizing heterogeneous resources, and navigating the expansive state space arising from dynamic request patterns. To address these challenges, we propose SAMD-SRG, a novel microservice deployment framework that integrates spectral clustering and graph reinforcement learning. Specifically, we first employ SpecRank, a microservice ranking method utilizing Hermitian spectral clustering, to enable intelligent grouping and ordering. We then design Hierarchical Service Representation Learning (HSRL), leveraging graph convolutional networks and cross-layer multi-head attention to capture bidirectional service-edge interaction representations. Finally, we adopt Graph-Enhanced Actor-Critic (GEAC) to optimize microservice deployment decisions. Evaluations across four cross-regional scenarios show that SAMD-SRG significantly outperforms existing baselines, with 48% reduction in response time, 52.50% lower deployment costs, and 75.93% improved load balancing in benchmark Scene 1.