Time Series Analysis to Predict End-to-End Quality of Wireless Community Networks

• Published in: Electronics (Basel) 2019-05, Vol.8 (5), p.578
• Main Authors: Millan, Pere, Aliagas, Carles, Molina, Carlos, Dimogerontakis, Emmanouil, Meseguer, Roc
• Format: Article
• Language: English
• Subjects: Algorithms; Communication; Community; Community networks; Cost control; Deep learning; Developing countries; End-to-end quality prediction; Enginyeria de la telecomunicació; Internet; Internet access; LDCs; Machine learning; Routers; Routing (telecommunications); Rural areas; Telemàtica i xarxes d'ordinadors; Time series; Time-series analysis; Traffic congestion; Wireless LANs; Wireless networks; Xarxes locals sense fil Wi-Fi; Àrees temàtiques de la UPC
• ISSN: 2079-9292; 2079-9292
• DOI: 10.3390/electronics8050578

Community Networks have been around us for decades being initially deployed in the USA and Europe. They were designed by individuals to provide open and free “do it yourself” Internet access to other individuals in the same community and geographic area. In recent years, they have evolved as a viable solution to provide Internet access in developing countries and rural areas. Their social impact is measurable, as the community is provided with the right and opportunity of communication. Community networks combine wired and wireless links, and the nature of the wireless medium is unreliable. This poses several challenges to the routing protocol. For instance, Link-State routing protocols deal with End-to-End Quality tracking to select paths that maximize the delivery rate and minimize traffic congestion. In this work, we focused on End-to-End Quality prediction by means of time-series analysis to foresee which paths are more likely to change their quality. We show that it is possible to accurately predict End-to-End Quality with a small Mean Absolute Error in the routing layer of large-scale, distributed, and decentralized networks. In particular, we analyzed the path ETX behavior and properties to better identify the best prediction algorithm. We also analyzed the End-to-End Quality prediction accuracy some steps ahead in the future, as well as its dependency on the hour of the day. Besides, we quantified the computational cost of the prediction. Finally, we evaluated the impact of the usage for routing of our approach versus a simplified OLSR (ETX + Dijkstra) on an overloaded network.