Ray tracing is a popular technique for realistic image synthesis, but it requires a huge amount of computation which makes it extremely time-consuming. There has been serveral efforts to reduce the computation time through parallel processing. The image parallel approach which decompose the image among processors can achieve almost uniform load-balancing but they suffer from the communication overhead due to a large number of data transfer. This thesis quantifies the pixel coherence in the ray-tracing context and presents a parallel processing scheme which exploits the pixel coherence in order to minimize the communication overhead. The scheme is mainly characterized by its caching scheme, the predictive caching scheme, which identifies the processor set that every processor in the set is very likely to need the data requested by a processor, and sends the data to all processors in the set. The data transfer takes advantages of the multicasting which makes much less message traffic than multiple one-to-one transfers. The performance of proposed scheme is compared with those of previous schemes by analysis and experiments on various interconnection topologies.
