Wavelet analysis for compression and feature extraction of network performance measurements
2008-08-12T11:14:38Z (GMT) by
Monitored network data allows network managers and operators to gain valuable insight into the health and status of a network. Whilst such data is useful for real-time analysis, there is often a need to post-process historical network performance data. Storage of the monitored data then becomes a serious issue as network monitoring activities generate significant quantities of data. This thesis is part of the EPSRC sponsored MASTS (Measurements in All Scales in Time and Space) project. MASTS is a joint project between Loughborough, Cambridge and UCL and focuses on measuring, analyzing, compressing and storing network characteristics of JANET (UK's research/academic network). The work in this thesis is motivated by the need of measuring the performance of high-speed networks and particularly of UKLight. UKLight connects JANET to U.S.A and the rest of Europe. Such networks produce large amounts of data over a long period of time, making the storage of this information practically inefficient. A possible solution to this problem is to use lossy compression on an on-line system that intelligently compresses computer network measurements while preserving the quality in important characteristics of the signal and various statistical properties. This thesis contributes to the knowledge by examining two threshold estimation techniques, two threshold application techniques, the impact of window size on the lossy compression performance. In addition eight different wavelets were examined in terms of compression performance, energy preservation, scaling be- haviour, quality attributes (mean, standard deviation, visual quality and PSNR) and Long Range Dependence. Finally, this thesis contributes by presenting a technique for precise quality control of the reconstructed signal and an additional use of wavelets for detecting sudden changes. The results of the thesis show that the proposed Gupta-Kaur (GK) based algorithm compresses on average delay signals 17 times and data rate signals 11.2 times while accurately preserving their statistical properties.