Ethernet performance anomalies have been shown to be a symptom of 'soft' network faults. Soft network faults are generally characterized by peformance degradation rather than a total failure. Previous work on detecting performance anomalies has focused on using a set of parallel single-dimension detectors. This research extends these single-dimensional detectors to a multi-dimensional detection algorithm.
The multi-dimensional algorithm, the HMP algorithm, was designed based on the study of Ethernet data using dynamic visualization techniques. Dyanmic visualization is a means for extending the dimensionality that can be represented visually through the use of motion and user interaction. Dynamic cisualization tools are used to study the normal and faulty behavior of the Ethernet. Methods for applying these tools to identify peformance anomalies visually are presented along with several case studies of network faults. The analysis further reveals properties of the data that were exploited to develop the HMP algorithm. An injection experiment was performed to compare the performance of the HMP algorithm with the existing single-dimensional algorithms. To further assess the state of the art in anomaly detection, the injection data was presented to human volunteers to compre human detection capabilities against those of the machine algorithms.
Sorry, the full paper is not available online yet, so you will have to be satisified with the snazzy 3D stereogram of network data for now (you will need colored 3D glasses to view it).