CNFP: Optimizing Cloud-Native Network Function Placement with Diffusion Models on the Cloud Continuum

TL;DR

This paper introduces a diffusion model-based framework for optimizing cloud-native network function placement across the cloud continuum, addressing scalability and constraint handling limitations of classical methods.

Contribution

It proposes a novel diffusion-based approach using graph neural networks for scalable, constraint-aware CNF placement in cloud networks, outperforming traditional methods.

Findings

Consistently generates feasible solutions with high accuracy.

Achieves faster inference compared to existing solvers.

Demonstrates robustness on diverse and larger network topologies.

Abstract

The placement of Cloud-Native Network Functions across the Cloud-Continuum represents a core challenge in the orchestration of current 5G and future 6G networks. The process entails the implementation of interdependent computing tasks, which are structured as Service Function Chains, over distributed cloud infrastructures. This is achieved while satisfying strict resource, bandwidth, connectivity, and end-to-end latency constraints. It is widely acknowledged that classical approaches, including mixed-integer (non)linear programming, heuristics, and reinforcement learning, face practical limitations in terms of scalability, robust constraint handling, and generalization to unseen network conditions. In this study, a diffusion-based theoretical and algorithmic framework for CNF placement is proposed, based on Denoising Diffusion Probabilistic Models. The placement process is reconceptualised as a conditional graph-to-assignment generation task. Each scenario is encoded as a heterogeneous graph, capturing infrastructure and service-chain structure. A Graph Neural Network denoiser is trained to iteratively refine noisy CNF-to-cloud assignment matrices. In order to bias the generation process towards valid deployments, the model incorporates constraint-aware penalties during training. At inference, a multitude of candidate placements are sampled, and the best suboptimal, feasible solution is selected. Extensive experimentation on diverse topologies, incorporating out-of-distribution evaluations with larger instances and shifted constraint regimes, demonstrates that the proposed approach consistently generates feasible solutions with considerably accelerated inference compared to other solvers. The findings of this study demonstrate the potential of diffusion-based generative modelling as a scalable tool for constrained network placement and embedding in cloud-continuum orchestration.