Addressing RouteNet scalability through input and output design

Abstract

With recent advances in the field of Machine Learning (ML), a multitude of problems related to communication systems and networks can be solved with data-driven solutions. Since data in these systems is mostly represented as graphs, Graph-based Neural Networks (GNNs) are a good candidate for solving such problems. These GNNs can be used as a computer network modeling technique to build models that accurately estimate the Key Performance Indicators (KPI) such as delay or jitter in real network scenarios in order to ensure their requirements in terms of service assurance. To build GNN solutions with higher accuracy, low computational resource requirements, and easy deployment of synthetic network training results into real-world networks, it is more than necessary to develop efficient and effective GNN models. This paper presents a GNN model capable of accurately estimating the average delay per flow in networks. By designing scale-independent features and using notions from queuing theory, the proposed model successfully generalizes to large size topologies, routing configurations, and traffic matrices not seen during the training phase.

Designing graph neural networks training data with limited samples and small network sizes

“…This process repeats for 𝑇 iterations, where 𝑇 is a user-de ined parameter before the low states are inally used to compute the performance metric estimations. Furthermore, in a manner similar to the model in [24], RouteNet-Fermi also features design choices such as replacing the numerical "link capacity" value with a relative value representing the traf ic load of a link based on its capacity, and the delay is inferred from the queue occupancy rather than predicted directly, which helps retain accuracy even when testing on networks much larger than those experienced during training.…”