Can Neural Networks Provide Latent Embeddings for Telemetry-Aware Greedy Routing?

TL;DR

This work tackles telemetry-aware routing by learning latent node embeddings that support fast greedy routing without solving all-pairs shortest paths. It introduces Placer, a Graph Neural Network-based method that maps telemetry-rich network states to embeddings and performs next-hop decisions in embedding space, trained with proximal policy optimization in a PackeRL/ns-3 framework. On a five-node topology, Placer achieves higher goodput and lower drops than telemetry-oblivious baselines, and provides a pathway toward explainable neural routing via embedding visualizations. However, embeddings appear quasi-static under centralized inference, motivating future work in distributed multi-agent training and exploration of hyperbolic embedding spaces for better scaling.

Abstract

Telemetry-Aware routing promises to increase efficacy and responsiveness to traffic surges in computer networks. Recent research leverages Machine Learning to deal with the complex dependency between network state and routing, but sacrifices explainability of routing decisions due to the black-box nature of the proposed neural routing modules. We propose \emph{Placer}, a novel algorithm using Message Passing Networks to transform network states into latent node embeddings. These embeddings facilitate quick greedy next-hop routing without directly solving the all-pairs shortest paths problem, and let us visualize how certain network events shape routing decisions.

Figures (2)

• Figure 1: a): Five-node network topology mini-5 used in our experiments. b, c): PackeRL's simulation backend turns network topology, incoming traffic and routing actions into an attributed state graph. d, e): Placer obtains latent node embeddings using its Message Passing Network, then greedily converts these into next-hop selections.
• Figure 2: a) An embedding of mini-5 produced by Placer$_{d=2}$. b) Goodput vs. fluctuation of Placer$_{d=2}$ for $8$ random seeds.