A Competitive Analysis of Load Balancing Strategiesfor Parallel Ray Tracing
The Journal of Supercomputing
Understanding the efficiency of ray traversal on GPUs
Proceedings of the Conference on High Performance Graphics 2009
| Hi-index | 0.00 |
Ray tracing is at the core of most techniques for creating realistic imagery. Parallel implementations of ray tracing handle the irregular workload through task systems. The strengths of static and dynamic scheduling strategies are complementary to each other. Static strategies do not incur in synchronization overhead while dynamic strategies generally provide computational times closer to the optimal scheduling. Hybrid strategies combining good static initialization and dynamic task assignation have been shown to be a better alternative than pure static and dynamic strategies [Heirich and Arvo 1998]. We experiment with a novel strategy for load balancing on multi-GPU systems. We obtain a quick estimate of the cost of traversing batches of rays over bounding volume hierarchies. The estimated costs are used to achieve a tighter assignation of tasks to processing units. Results suggest that cost-based initialization can enhance common balancing strategies and reduce rendering times.