Parallel Ray Casting of Visible Human on Distributed Memory Architectures

This paper proposes a new parallel ray-casting scheme for very large volume data on distributed-memory architectures. Our method, based on data compression, attempts to enhance the speedup of parallel rendering by quickly reconstructing data from local memory rather than expensively fetching them from remote memory spaces. Furthermore, it takes the advantages of both object-order and image-order traversal algorithms: It exploits object-space and image-space coherence, respectively, by traversing a min-max octree block-wise and using a runtime quadtree which is maintained dynamically against pixels’ opacity values. Our compression-based parallel volume rendering scheme minimizes communications between processing elements during rendering, hence is also very appropriate for more practical distributed systems, such as clusters of PCs and/or workstations, in which data communications between processors are regarded as quite costly. We report experimental results on a Cray T3E for the Visible Man dataset.