With the advent of advanced biomedical and scientific instruments and/or simulations, scientists and researchers have to cope with more and more three-dimensional data sets, such as computer tomography, 3D ultra-sound, and remote-sensing laser radar images. Volume rendering provides a mechanism to visualize these data's complicated internal structures by projecting them to 2D images according to user-specified views. Because of its 3D nature, the size of volumetric data sets is usually rather large. For example, a $1,000^{3}$ cube with each voxel represented by 24-bit numbers requires 3 Gbytes of storage space. As a result, disk I/O becomes a critical issue for reducing the end-to-end volume rendering latency. We have developed two novel techniques to attack the I/O delay problem associated with volume visualization. The first technique, Compression-Domain Volume Visualization, renders 3D data sets directly from their compressed representations, thus reducing both the I/O overhead as well as eliminating decompression CPU time. The second technique, View-Specific Subcube Compositing, tailors the traversal order of 3D data sets to user-given views so that the I/O accesses to parallel disk arrays are guaranteed to be conflict-free, thus exploiting the I/O system's bandwidth to its fullest.
In this project, we propose to integrate these two techniques with parallel computing to design and implement a high-performance volume visualization server that provides interactive response time and is completely based on off-the-shelf hardware components.