Automated detection of unexpected communication network performance changes
The explosive growth of the Internet and associated services has seen an increase in dependence on communication networks. Given this level of dependence on communication networks it is important that they are well managed. ‘Data exceptions’ are a useful summary of performance information that can assist network operators in their management tasks. These represent distinct periods of network operation when performance has been different in some way from that expected. Previous research conducted using a simulation has shown that a two-stage algorithm using the Kolmogorov–Smirnov test statistic and a neural classifier is a successful approach for detecting and classifying data exceptions. However, this approach had not been verified on a measured data set from a real network. The authors present results from applying the above-mentioned technique to data from both a purpose-built test network and a UK based commercially operational network. It is shown that these results actually improve upon those from a simulation and discuss reasons for this discrepancy. Finally, the approach is applied to data from NLANR's Active Measurement Project, representing a section of Internet operation.
