Parallel Out-of-Core Occlusion Culling

We present a parallel rendering system for PC-Clusters to visualize large 3D scenes. One single visualization node, equipped with a high-end graphics adapter, is supported by a group of backend nodes with weak graphics performance. The objects of the scene are distributed among these backend nodes, they serve two purposes: First, they provide an out-of-core memory system for the visualization node. Second, they assist the visualization node's rendering by performing visibility calculations and only sending visible objects to the visualization node. In order to obtain fast rendering with our system, we have to distribute the objects among the backend nodes in a way that does not only guarantee an even distribution of the objects, but also an even distribution of the visibility calculations and the amount of data send to the visualization node. We identify necessary properties of the distribution and argue that a random distribution is a good candidate. Further, in order to reduce the number of objects sent to the visualization node per frame, we employ an approximate hierarchical occlusion culling in each backend node. For this, they are equipped, in addition to the objects assigned to them, with simplified versions of the other objects of the 3D scene. The visualization node is equipped with 512 MiB video memory and supported by 15 backend nodes. This system is able to render a approx. 350 million polygons (approx. 8.5 GiB) large aircraft model between 20 - 30 fps and thus allows a walkthrough in real-time.