TorusVis^ND: Unraveling High-Dimensional Torus Networks for Network Traffic Visualizations

Abstract

Torus networks are widely used in supercomputing. However, due to their complex topology and their large number of nodes, it is difficult for analysts to perceive the messages flow in these networks. We propose a visualization framework called TorusVisND that uses modern information visualization techniques to allow analysts to see the network and its communication patterns in a single display and control the amount of information shown via filtering in the temporal and the topology domains. For this purpose we provide three cooperating visual interfaces. The main interface is the network display. It uses two alternate graph numbering schemes -- a sequential curve and a Hilbert curve -- to unravel the 5D torus network into a single string of nodes. We then arrange these nodes onto a circle and add the communication links as line bundles in the circle interior. A node selector based on parallel coordinates and a time slicer based on ThemeRiver help users focus on certain processor groups and time slices in the network display. We demonstrate our approach via a small use case.