Scalable routing and wavelength assignment in large optical networks

Recent advances in Graph Neural Networks (GNN) have shown a dramatic improvement in computer networks problems. As a result, GNN seems promising to solve many relevant network optimization problems (e.g., routing and wavelength assignment) in self-driving software-defined networks. However, most state-of-the-art GNN-based networking techniques fail to generalize, which means that they perform well in network topologies seen during training, but not over large topologies. The reason behind this important limitation is that existing GNN networking solutions use standard graph neural networks that are not suited to learn large graph-structured information in routing purposes. In this Thesis, we transform the RWA problem into an ML-based classification problem, where the routing solution is provided by a classifier in response to a given input graph. To this end, a Message Passing Neural Network which is a type of Graph Neural Network (GNN) is trained based on various types of graphs. Once trained, such a classifier is able to provide a route for newly-incoming traffic requests in an online fashion, offering an RWA configuration within a few milliseconds, thus allowing to perform dynamic network adaptation and reconfiguration in response to frequently changing traffic patterns. MPNN is tailored to learn and model information structured as graphs and as a result, our model is able to generalize over arbitrary topologies and variable traffic intensity. To showcase its generalization capabilities, we evaluate it on a SDN-based Optical Transport Network (OTN) scenario, where traffic demands need to be allocated efficiently. Our results show that our model is able to achieve outstanding performance in large topologies never seen during training.