KG-Orchestrator Graph Neural Network-Driven Resource Orchestration for 6G Distributed Networks

Abstract

The highly distributed infrastructure and the hetero geneous services and dynamic workloads of future 6G networks impose severe requirements on resource orchestration. We in troduce KG-Orchestrator, a unified framework which combines a Neo4j knowledge graph and a multi-GNN ensemble (Graph SAGE, HashGNN, GAT, GCN) to perform topology-aware and constraint-driven data center selection, and a knowledge graph continuously captures telecom resources and dependencies while GNNs learn multi-objective embeddings in terms of proximity, capacity and semantic compatibility, which is dynamic greedy policy exploits for real-time allocation. The results show that all GNNs support 100% satisfaction under baseline conditions, but attention-based models degrade under low feature variance. Compared to other RL and ML baselines (DQN, Random Forest) and a static First-Fit heuristic, our KG-Orchestrator approach maintains comparable service latency while reducing decision time by over 50%, consequently reaching sub-0.02 ms per-request decisions and retaining robust performance under workload surges up to 40% above baseline capacity.