AI-Driven Optimization of Inter-Service Communication and Latency Reduction in Distributed Microservices

Abstract

The proliferation of cloud-native microservice architectures has fundamentally transformed enterprise software delivery, yet the accompanying inter-service communication overhead constitutes a dominant source of end-to-end latency, often jeopardising stringent service level objectives (SLOs). Existing mitigation strategies ranging from static load balancing to threshold-based horizontal pod autoscaling remain predominantly reactive and topology-agnostic, failing to exploit the rich structural and temporal signals inherent in the microservice call graph. This paper presents GAL-RL (Graph Attention Latency Reinforcement Learner), a novel framework that tightly couples multi-head graph attention networks with a continuous-action Soft Actor-Critic agent to jointly optimise peredge traffic routing and per-service replica scaling. The GAT encoder learns predictive latency embeddings by attending over spatio-temporal neighbourhoods in the service mesh, while the SAC policy translates these embeddings into fine-grained resource orchestration decisions that balance tail-latency reduction against compute expenditure. Evaluated on the DeathStarBench social-network workload and the PetShop anomaly-injection benchmark, GAL-RL reduces 95th-percentile latency by 42 % and CPU utilization by 23 % compared with Kubernetes HPA, while maintaining a 2.1 % SLO violation rate. Ablation studies confirm that both the graph attention mechanism and the joint routing-scaling formulation are important for achieving these performance gains.