Extending RouteNet to model more complex networking scenarios
Recently, a Graph Neural Network (GNN) model called RouteNet was proposed as a method to estimate end-to-end computer network performance metrics (e.g. delay, jitter, packet loss rate) given information about topology, routing con guration, and tra c demands of the network. RouteNet achieves an accu...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2020 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/335301 |
| Acceso en línea: | https://hdl.handle.net/2117/335301 |
| Access Level: | acceso abierto |
| Palabra clave: | Neural networks (Computer science) Machine learning Graph theory Modelatge de Xarxes Xarxes de Computadors Network Modelling Graph Neural Networks Machine Learning Computer Networks Xarxes neuronals (Informàtica) Aprenentatge automàtic Grafs, Teoria de Àrees temàtiques de la UPC::Informàtica |
| Sumario: | Recently, a Graph Neural Network (GNN) model called RouteNet was proposed as a method to estimate end-to-end computer network performance metrics (e.g. delay, jitter, packet loss rate) given information about topology, routing con guration, and tra c demands of the network. RouteNet achieves an accuracy comparable to packet-level simulators but at a fraction of their computational cost. In spite of its success in making accurate estimations and generalizing to unseen topologies, it does not take into account all the complex particularities of real computer networks. One of the simplifying assumptions that RouteNet makes is to assume that all nodes (forwarding devices) in the network have the same characteristics. In this thesis, we propose a new architecture that allows to introduce the concept of node into RouteNet and support the modelling of di erent forwarding device features. To assess the validity of the new architecture we build a data set composed of network scenarios where nodes have di erent queue sizes, and then we use it to train and evaluate the proposed model. Our solution constitutes one step forward towards building more realistic network models that can be used to optimize and manage real-world computer networks. |
|---|