On the Periodicity of Time-series Network and Service Metrics

The presence of an underlying periodicity in time-series network and service metrics has been used as a basis for some recent anomaly detection techniques. These techniques however assume the presence of a periodicity, and would benefit from the concept of a quantitative figure of merit for the strength of a given periodicity in the metric. We survey a number of potential techniques for this purpose, and find none suitable. As such, we construct such a figure of merit to suit our application. Use of the figure of merit allows selection of the most appropriate period for the metric, and we present an efficient automated method for this selection. Furthermore, this figure of merit is a useful indicator of whether periodic analysis for anomaly detection is in fact suitable for the given metric. Finally, we suggest a number of other areas where use of the figure of merit could enhance anomaly detection using periodic analysis.