Visualizing the Five-dimensional Torus Network of the IBM Blue Gene/Q

Abstract

Understanding the interactions between a parallel application and the interconnection network over which it exchanges data is critical to optimizing performance in modern supercomputers. However, recent supercomputing architectures use networks that do not have natural low-dimensional representations, making them difficult to comprehend or visualize. In particular, high-dimensional torus networks are common and are used in four of the top ten supercomputers and eight of the top ten on the Graph500 list. We present a new visualization of five-dimensional torus networks. We use four connected views depicting the network at different levels of detail, allowing analysts to observe general large-scale traffic patterns while simultaneously viewing individual links or outliers in any specific section of the network. We demonstrate this approach by analyzing network traffic for a pF3D simulation running on the IBM Blue Gene/Q architecture, and show how it is both intuitive and effective for understanding and optimizing parallel application behavior.