Performance Comparisons of Visualization Architectures

Abstract

Large data visualization problems generally demand higher resolutions than are available on typical desktop displays. An isosurface from a multiterabyte data set may consist of several million polygons. A common 1600 .. 1200 display (1.92 million pixels) could at most display every polygon as a single pixel. This is insufficient resolution for examining the detail of the isosurface. Commodity hardware components provide one approach for handling large data visualization problems. For example, the PC gaming industry has exponentially increased the processing power of graphics cards to the point where they now are used in graphics supercomputers, such as the SGI Onyx4 and Prism. Commodity-based visualization clusters are becoming popular in the visualization community because of the high performance-to-cost ratio. Prior to the visualization cluster, graphics supercomputers, such as SGI’s Onyx systems, were needed for high-resolution, multimonitor capability. The Scientific Visualization Center (SVC) at the U.S. Army Engineer Research and Development Center Major Shared Resource Center has purchased a visualization cluster to explore the limits of this trend with respect to large, computational data sets. Results from central processing unit and graphics benchmarking between the Graphstream visualization cluster, an SGI Onyx340, and an SGI Prism will be presented.