Graph-Based Resource Optimization for Dynamic Cloud Microservice Systems

Abstract

Cloud-native applications increasingly rely on interconnected microservices that experience rapidly changing workloads and complex resource dependencies. Traditional resource allocation mechanisms often struggle to maintain application performance while minimizing infrastructure costs. This study proposes a graph-based resource optimization framework that combines graph representation learning with reinforcement learning to improve resource management in distributed microservice systems. The proposed approach models service communication patterns, resource utilization, and workload variations through dynamically updated dependency graphs. The illustrative evaluation uses 21 million distributed tracing records representing 3,600 microservice deployment configurations under heterogeneous workload conditions. Compared with conventional threshold-based autoscaling, the proposed framework reduces service-level objective violations from 5.4% to 2.8% and decreases average CPU resource consumption by 17.2%. Furthermore, the 99th-percentile request latency decreases from 218 to 184 milliseconds under comparable traffic loads. Additional experiments indicate that the framework adapts effectively to workload bursts and changing service dependencies. The findings suggest that graph-based learning can support more efficient and reliable resource management in large-scale cloud computing environments.