Detecting changes in network performance from low level measurements
The Internet and associated network technologies are an increasingly integral part of modem day working practices. With this increase in use comes an increase in dependence. For some time commentators have noted that given the level of reliance on data networks, there is a paucity of monitoring tools and techniques to support them. As this area is addressed, more data regarding network perfonnance becomes available. However, a need to automatically analyse and interpret this perfonnance data now becomes imperative. This thesis takes one-way latency as an example perfonnance metric. The tenn 'Data Exception' is then employed to describe delay data that is unusual or unexpected due to some fundamental change in the underlying network perfonnance. Data Exceptions can be used to assess the effect of network modifications and failures and can also help in the diagnosis of network faults and perfonnance trends. The thesis outlines how Data Exceptions can be identified by the use of a two-stage approach. The Kolmogorov-Smirnov test can initially be applied to detect general changes in the delay distribution, and where such a change has taken place, a neural network can then be used to categorise the change. This approach is evaluated using both a network simulation and a test network to generate a range of delay Data Exceptions.