GPU Cluster Based Parallel Volume Rendering

Direct volume rendering is a very useful way for visualizing volumetric data. Such data can be acquired from different sources, like data from Computed Tomography (CT) or Magnetic Resonance Imaging (MRI) scanners, computational fluid dynamics (CFD) and seismic data. Many scientific and medical researches can produce high-resolution volume data sets that cannot be rendered on a single PC. For example, medical CT scanners can produce large sizes of scalar data sets, which can be in the range of megabytes even gigabytes. Time-dependent CFD simulation data can comprise several gigabytes for a single time step and several hundred or thousand time steps.An important step in volume rendering is to design of transfer functions that will highlight those aspects of the volume data that are of interest to the user. For many applications, boundaries reveal most of the important information. We present a method to identify the materials that form the boundaries by the Runge-Kutta method. They are then used in a LH domain to help interactive and semiautomatic design of appropriate transfer functions.Parallel volume rendering solves the large data visualization problem by distributing both the data and rendering calculations among computer nodes. In this paper, we propose a technique for sort-last parallel volume rendering. The volume data is split between the nodes and each node renders its own portion. Then, compositing is used to form a final image from each node's rendering, i.e. the volume rendering integral is utilized numerically through back-to-front compositing all sub images.Load balancing is important to parallel programs for performance reasons. When using level of detail techniques or when zooming on parts of the datasets, load unbalance becomes an important issue. In this paper, we propose a technique to achieve good load balancing for parallel volume rendering by dynamically distributing the data among the rendering nodes according to the cost of the previous frame, which is used on the kd-tree to rebalance the load.By using the GPU cluster based parallel volume rendering, we show that it is indeed a useful way for visualizing high-resolution volumetric data, e.g. The Visible Human Image Data Set of U.S. National Library of Medicine and the performance of the parallel volume rendering system is improved about 70%.