high performance computing on graphics processing units: hgpu.org

To improve the performance of large-scale rendering, it requires not only a good view of data structure, but also less disk and network access, especially for achieving the realistic visual effects. This paper presents an optimization method of global illumination rendering for large datasets. We improved the previous rendering algorithm based on Monte Carlo ray tracing and the scheduling grids, and reduced the remote reads by slightly organizing the original data with considerations of locality and coherence. We implemented the rendering system in a Hadoop cluster of commodity PCs without high-end hardware. The large scene data are processed in splits by MapReduce framework, which increases scalability and reliability. The result shows that our algorithm of scheduling rays for each data split fits with large-scale scene and takes less reads and rendering time than previous works.