Dragon-Lab, Network States Detection and Identification Framework: Performance Investigation
Master thesis
Permanent lenke
http://hdl.handle.net/11250/262455Utgivelsesdato
2011Metadata
Vis full innførselSamlinger
Sammendrag
As IP networks have become the support of an increasingly varied range of applications, from games, video streaming, to e-commerce and online banking, it is critical to understand the functioning and performance bounds of the network in order to provide a certain QoS.To be able to achieve the understanding of the whole network huge amounts of data have to be acquired and processed. Various methods of analyzing the network have been created, such as: analysis of point-to-point packet delay, delay tomography from end-to-end unicast measurements, network unreachabilities troubleshooting using end-to-end probes and routing data, analysis of link failures in an IP backbone etc.In this research, a new approach on Internet analysis was adopted. Dragon-Lab is capable of detection, identification, and temporal and special localization of Internet backbone anomalous states based on two end-to-end metrics, packet delay and packet loss, and traceroutes information of the measured Internet paths. The anomalous states or instabilities are: congestion, queue building up and link failure.This is a new approach in the field of network troubleshooting and management. It is a bridge between methods focusing mainly on delay analysis and root cause analysis methods focusing on harvesting huge amounts of routing data in order to identify and localize network problems. Dragon-Lab is based on a three months measurement study conducted over the Internet between Norway, China and New Zealand.Relying on the Principal Component Pursuit processed data to identify anomalous delay changes and on general network knowledge, the unstable network states are identified and implicitly temporally localized.Further analysis is performed by processing instabilities into cumulative distributed functions of duration of instabilities, time between instabilities, charts of the distributions of instabilities per path, and per type of instability, in order to study instabilities impact on Internet paths. Moreover, general metrics have been defined in order to give an overview of the impact of the combination of instabilities on paths. The metrics are: availability, fatigue or stability of the Internet path.Further Dragon-Lab is improved by geo-localization of the instabilities. The research aims to use available data to pinpoint the source location of instabilities on the Internet paths, in essence to find the problem hop on the path by IP. This is achieved by combining reconstructions algorithms from the Compressive Sensing domain with Dragon-Lab knowledge on instabilities. The approach presents provides a solution bounded by certain conditions and having a few limitations and also a spacial localization algorithm.