Load balancing for a parallel radiosity algorithm

Abstract

The radiosity method models the interaction of light between diffuse surfaces, thereby accurately predicting global illumination effects. Due to the high computational effort to calculate the transfer of light between surfaces and the memory requirements for the scene description, a distributed, paraUelized version of the algorithrn is needed for scenes consisting of thousands of surfaces. We present several load distribution schemes for such a parallel algorithm which includes progressive refinement and adaptive subdivision for fast solutions of high quality. The load is distributed before the calculations in a static way. During the computation the load is redistributed dynamically to make up for individual differences in processor loads. The dynamic load balancing scheme never generates more data packets than the original algorithm and avoids overloading processors through actions taken by the scheme. CR