A scheme for the visualization of large data volumes using volume rendering on a distributed memory MIMD system is described. The data to be rendered is decomposed into subvolumes to reside in the local memories of the system's nodes. A partial image of the local data is generated at each node by ray tracing, and is then composited with partial images on other nodes in the correct order to generate the complete Image. Subvolumes whose voxels are classified as being mapped to zero opacity are not rendered, giving rise to an imbalance of work amongst nodes. Scattered decomposition is used for load balancing, which on one hand: creates additional overheads in compositing and communication, but on the other, provides an improvement in throughput that is dependent on the characteristics of the data. Experimental results for a typical data set rendered on a 1024-node Fujitsu AP1000 are reported.
展开▼
