Data Locality in Parallel Rendering

One of the main challenges in global illumination is rendering scenes with millions of polygons and megabytes of textures. Combining the processing power and the memory of multiple processors or workstations to render these complex scenes is an attractive proposition but the complex interactions between data and processing introduces a significant amount of data communication. Data locality methods may improve cache coherence and cache access coherence by finding an optimal data partitioning, by re-ordering computations, and by replacing complex geometry with simplified image-based representations. We review the different data locality methods and focus on local caching of global radiance values. We present the results of an implementation in the ray tracing program Radiance.