Exploiting deep parallel memory hierarchies for ray casting volume rendering

Previous work in single-processor ray casting methods for volume rendering has concentrated on algorithmic optimizations to reduce computational work. Previous work in parallel volume rendering has concentrated on partitioning, with the goals of maximizing load balance and minimizing communication between distributed nodes. Building on our previous work at lower levels of the hierarchy, we present techniques to efficiently exploit all levels of the deep memory hierarchy of a distributed Power Challenge Array, on which we implement a logical global address space for volume blocks with caching. This focus on the optimal exploitation of the entire memory hierarchy, from the processor cache to the interconnection network between distributed nodes, allows us to efficiently render a 7.1 GByte dataset. Our results have implications for the parallel solution of other problems which, like ray casting, require a global gather operation and contain coherence. We discuss implications for the design of a parallel architecture suited to solving this class of problems.