Citation
Benayas de los Santos, Fernando
(2018).
Application of bayesian reasoning for fault diagnosis over a controller of a software defined network.
Proyecto Fin de Carrera / Trabajo Fin de Grado, E.T.S.I. Telecomunicación (UPM), Madrid.
Abstract
Current trends, such as the increasing quality and consumption of video services, the adoption of 5G technologies, and the growing presence of IoT devices are overloading current telecommunication. This is further aggravated by the adoption of cloud services. In order to face these challenges, exible networking policies are needed. This would enable dynamic, fast re-routing, depending on parameters such as tra_c load and Quality of Service.
This dynamic network managing is made possible by the Software Defined Networks (SDN).
SDN decouple the forwarding plane and the control plane, centralises the latter and exposes it through an API. Therefore, by delegating on the controller the task of injecting network policies into the network elements, the process of creating and injecting tra_c policies into
the network is simpli_ed, allowing for a more agile management.
Notwithstanding the advantages mentioned before, SDN have some resiliency issues.
Frequent changes in networking rules entails constant possibilities for faulty tra_c policies to be introduced in the network. This could severely hinder the bene_ts of switching from legacy to SDN technology. Hence, a system that allows users auditing these faults is needed.
The proposed system is based on a Bayesian network learned from labelled and processed data obtained from SDN, looking for causal relationships between data values and current
state of the network. In order to obtain labelled data, SDN has been simulated, creating multiple scale-free connected networks, where multiple tra_c types have been implemented.
The resultant Bayesian network is then used to diagnose the status of the network from new labelled and processed data.
Regarding the results, we have obtained a precision value of 92.2%, a recall value of 91.9%, an accuracy value of 97.8%, and a F1-Score value of 91.2% in the best of the models tested.