EP-NCO slashes edge-cloud latency by 50% with neural combinatorial optimization

A team of researchers from Karlstad University and other institutions has introduced EP-NCO (Edge Placement Neural Combinatorial Optimiser), a learning-based framework that tackles the NP-hard problem of placing interdependent microservices across heterogeneous edge-cloud infrastructures. The system uses a dual-graph model: one graph captures computing infrastructure nodes and their resource relationships, while the other represents service dependencies within the application. Graph neural networks (GNNs) learn structural embeddings of both graphs, and a reinforcement learning policy then constructs feasible placements that minimize total response time by accounting for execution latency, communication link delays, and bandwidth-sharing effects.

Extensive simulations across multiple system scales demonstrate that EP-NCO consistently outperforms traditional metaheuristics like genetic algorithms and particle swarm optimization by 46–50%, and controlled RL ablation baselines by 25–35%. Critically, once trained, EP-NCO supports fast online inference, making it practical for dynamic, large-scale edge-cloud environments with hundreds of compute nodes hosting thousands of latency-sensitive IoT applications—a scale beyond the reach of current scheduling systems. The work has been published on arXiv (2606.25553) and represents a significant step toward intelligent, automated service placement in the compute continuum.