Quality of service analysis of internet links with minimal information
- Autores: Felipe Mata Marcos
- Directores de la Tesis: Javier Aracil Rico (dir. tes.)
- Lectura: En la Universidad Autónoma de Madrid ( España ) en 2012
- Idioma: inglés
- Tribunal Calificador de la Tesis: Daniel Morató (presid.), Jorge Enrique López de Vergara Méndez (secret.), Maria Kihl Palm (voc.), Michel Mandjes (voc.), Mikel Izal Azcárate (voc.)
- Materias:
- Enlaces
- Tesis en acceso abierto en: Biblos-e Archivo
- Resumen
Monitoring the Quality of Service (QoS) of Internet links is of paramount importance for Network Operators and Service Providers (NOSP), and consequently has received great attention from the research community. To monitor QoS, practitioners leverage on network traffic measurements and, by means of practical models and statistical techniques, make predictions and detect outliers that allow the planning of telecommunication networks and the detection of abnormal behavior, respectively.
However, obtaining detailed measurements from Internet links at current network speeds is very challenging, mainly because memory accesses¿ speeds have increased at a smaller pace than Internet links¿ speeds. Moreover, the amount of resources required to properly storage detailed network measurements make unfeasible to perform long measurement campaigns. These facts have motivated the application of different techniques to gather information from the network, such as collecting subsets of network traffic by applying sampling techniques in the packet capture process, or just collecting summarized statistics of the number of bytes transferred, such as those used in Multi Router Traffic Grapher (MRTG), where the maximum and average transfer speeds are recorded at non-overlapping time intervals of a given length. These techniques make network traffic monitoring less demanding and allow performing longer measurement campaigns.
Accordingly, this thesis proposes two methodologies to perform QoS analysis of Internet links leveraging on summarized statistics of network traffic. Each methodology relies on a network traffic model, validated using actual network traffic measurements, on which sound statistical methodologies are used on attempts of detecting relevant events that either require action from the network managers or may be related with degradations of the provided QoS.
The first methodology is designed to detect shifts in users¿ behavior, and consequently the detected events may entail capacity planning decisions. It builds on modeling the network traffic during a day using a multivariate fairly Gaussian distribution, from which changes in the parameters are detected at timescales of weeks. The change point instants are detected using clustering techniques and validated through the application of the Multivariate Behrens-Fisher Problem (MBFP). The proposed methodology is applied to real network measurements obtained from the Spanish academic network RedIRIS, showing satisfactory performance and entailing large Operational Expenditures (OPEX) reduction to NOSP in the management process of large-scale networks.
The second methodology performs anomaly detection through trend removal of network traffic measurements. It is tailored for Voice over IP (VoIP) traffic data, which is one of the most popular services provided through Internet nowadays. The methodology takes as input call count measurements of the VoIP service exhibiting seasonal trends, and outputs stationary residuals, which are used to detect anomalies by means of the application of unsophisticated statistical assumptions. Moreover, we propose a measurement alternative for monitoring VoIP systems. This alternative yields smaller correlations between the obtained measurements when some assumptions are met, which we showed to be satisfied in actual measurements we analyzed.
