As part of the OpenSGPlus project and the result of Marcus Roth's PhD thesis OpenSG was extended for supporting a new method to reduce the network data needing to be transmitted in Sort-Last Rendering. The results were published at EGPGV06. All pictures below were rendered on a 48 node PC cluster (the HEyeWall cluster. Each node has a 2.4 GHz P4 and an nVidia Quadro FX 1100 and they are connected on GBit Ethernet. The results below are for brute-force rendering without using occlusion culling or level of detail techniques or sharing any of the data, as the focus was on network bandwidth reduction for compositing.
The David statue has about 56 million polygons, and it was rendered at 17.4 fps.
The whole set of 16 Davids comes in just short of a billion polygons. These were rendered at 1-2 fps.
Each of the Powerplants has about 16 million polygons, the set of 48 comes to about 750 million polygons. Rendering speed of this was around 2 fps.
As an example on how the load balancing works this picture shows a fleet of Beetles, with the final image in the middle and the parts rendered on the different servers around it (in this case only 12 servers were used).
The glass is visible in all images, as Sort-Last in general doesn't support transparencies (yet, but we're working on it ;)). As long as there are only rather few transparent pieces, rendering them on all servers is a workaround for this problem.
The David model was received from the Stanford's Digital Michelangelo Project, the Powerplant from UNC's WalkThrough project. Big thanks to them for releasing these models! The Beetle model was provided by Volkswagen. Sincere thanks to them for this nice example dataset!
© by Marcus Roth and Dirk Reiners. Every usage of these images without prior written permission is strictly prohibited.