Efficient parallel volume rendering of large-scale adaptive mesh refinement data

Adaptive Mesh Refinement is a popular approach for allocating scarce computing resources to the most important portions of the simulation domain. This approach implies spatial compression and the large simulation sizes which necessitate it. We present a novel, cluster- and GPU-parallel rendering scheme for AMR data, which is built on previous work in the GPU ray casting of AMR data. Our approach utilizes the existing AMR structure to subdivide the problem into convexly-bounded chunks and perform static load-balancing. We take advantage of data locality within chunks to interpolate directly between blocks without the need to store ghost cells on the interior boundaries. We also present a novel block decomposition method, and analyze its performance against two alternative methods. Finally, we examine the interactivity of our renderer for multiple datasets, and consider its scalability across a large number of GPUs.