Adaptive load balancing for raycasting of non-uniformly bricked volumes

We describe a sort-last parallel volume rendering system based on single pass volume raycasting performed in the fragment shader unit. The architecture is aimed for displaying data sets that utilize the total distributed texture memory at interactive framerates. We use non-uniform texture bricks that are constructed and distributed by means of a kd-tree to employ object space partitioning. They are further used for implementing empty-space-skipping and a load balancing mechanism, which also makes use of the kd-tree, to increase the overall performance of the rendering system. Performance numbers are given for a mid-range GPU-cluster system consisting of eight render nodes with an Infiniband interconnection.