This paper presents a parallel ray-tracing algorithm in order to compute very large models (more than 100 million triangles) with distributed computer architecture. On a single computer, the size of the used dataset generates an out of core computation. Cluster architectures designed with off-the-shelf components offer extended capacities which allow to keep the large dataset inside the aggregated main memories. Then, to achieve scalability of operational applications, the real challenge is to exploit efficiently the amount of available memory and computing power. Ray-tracing, for high quality image rendering, spawns non-coherent rays which generate irregular tasks difficult to distribute on such architectures. In this paper we present a cache mechanism for main memory management distributed on each parallel computer and we implement a load balancing solution based on an auto adaptive algorithm to distribute the computation efficiently.

The use of SIMD ray packets [24] has been an important step forward in ray tracing performance. It is the first significant acceleration process based on a strategy that deals with "small" ray bundles. In order to improve the outcome for wider bundles, very powerful data structures have been used. However, these new data structures are not the most powerful when dealing with single rays or SIMD rays. They are therefore unsuitable, and this incompatibility raised the problem of incoherent rays which, such as secondary rays or global illumination rays (i.e. most the rays traced in real life ray tracing) are impossible to process efficiently. An original coupled use of BSP and BVH trees is proposed to overcome this disadvantage. Each tree, used individually, can efficiently boost both small ray bundles and wide ray bundles. Used together in a coupled strategy, they can achieve a global speedup of +50%.

No abstract

No abstract