Laziness, a Way to Improve Distributed Computation of the Ray Tracing Algorithm

Abstract

This paper raises the issue of computational workload and memory load balancing to ray trace large scenes efficiently on a network of workstations. The task splitting is done on the image to be produced and not on the scene data to obtain a high performance level. To deal with the high memory requirements of such a distribution strategy, laziness is added to the base algorithm. This reduces the computing local memory requirements to the locally computed part of the image. It also reduces the sequential parts of the algorithm by making the voxelization process parallel. Many comparisons have been done using a manager/worker distribution algorithm on different scenes computed on a conventional network of workstations. Performance, load imbalance, communication overhead, and memory requirement results are given and discussed in this paper. Furthermore, this paper demonstrates that the proposed solution improves the results obtained with conventional algorithms, no matter what network used or however complex the image is.