Cluster-Based Multi-Dimensional Visualization: Harnessing Computational Resources for Real-Time Visualization

Abstract

There is a growing need for high-frame-rate low-latency visualization solutions as medical practice moves toward interventional procedures. We present a cost-effective visualization system well suited for off-line visualization and interventional procedures. Users can view large time-resolved multi-dimensional datasets in real time with GPU cluster visualization. In addition, computational pre-processing can be hidden by rendering across distributed graphics cards, leading to improved frame-rates over a single graphics card solution. Finally, rendering on graphics cards offloads CPU cycles for generating the next time frame in the visualization. We have developed a network arbitration protocol for GPU cluster visualization called “token scheduling.” Our protocol reduces communication latency, which in turn lowers visualization latency and improves system stability and scalability. In addition, we evaluate GPU cluster behavior and performance through a timing analysis. This analysis leads to a better understanding of cluster size needed to achieve the desired frame rate of a given problem.