1. Cost-based Workload Balancing for Ray Tracing on a Heterogeneous Platform Mario Rincón-Nigro PhD Showcase Feb 17 th, 2012

2. Background Heterogeneous Computing Platforms – Widely available at all scales Ray Tracing – Most popular technique for photorealism – Base of many rendering algorithms – Computationally intensive – Embarrassingly parallel Cost-based Workload Balancing for Ray Tracing on a Heterogeneous Platform

3. Background: BVH Ray Traversal Cost-based Workload Balancing for Ray Tracing on a Heterogeneous Platform 7 boxes + 15 triangles tests Workload of ray tracing is irregular Per ray BVH traversals are highly variable Cost is hard to know beforehand Dynamic workload balancing What about if we could predict the traversal costs? How to use them to improve the balancing efficiency and reduce rendering times? 5 boxes + 9 triangles tests

4. Overview of the Approach Offline Build a BVH for the scene Compute expected number of primitive intersections Online Predict costs of batch of rays Initialize workload balancer based on predicted costs Heterogeneous launch of ray tracer Repeat for generations of secondary rays Cost-based Workload Balancing for Ray Tracing on a Heterogeneous Platform

5. Ray Traversal Cost Estimation 20 1812 675 66 44 879 EB = 4.2 ET =9.65 EB = 2 ET = 8.75 EB = 0 ET = 7 EB = 0 ET = 9 EB = 0 ET = 8 EB = 2 ET = 4 EB = 0 ET = 4 EB = 3.33 ET = 4.89 EB = 0 ET = 4 Boundary Cost-based Workload Balancing for Ray Tracing on a Heterogeneous Platform (4) In this example C(r) = 7 KB + 15 KT We traverse the BVH to 60% of its depth and sample 10% of the rays within each task

6. Workload Balancing Task for ray tracing are fixed size group of rays Two-level workload balancing – Inter-processor: need to split work between “big” processing units (CPUs, GPUs) – Intra-processor: need to split work between “small” processing units (SPs within a GPU) A variation of one of these strategies is commonly used: – Centralized queue – Distributed static assignation – Distributed dynamic assignation with task stealing Cost-based Workload Balancing for Ray Tracing on a Heterogeneous Platform

7. Balancing Strategies Distributed Queues Distributed Queues with Task Stealing Centralized Queue Static Balancing Dynamic Balancing Cost-based Workload Balancing for Ray Tracing on a Heterogeneous Platform

8. Experiments Test platform – AMAX machine running Intel Xeon E5600 processor, 16GB of RAM memory, and 3 NVidia Tesla C1060 GPUs. We have compared regular and cost-initialized versions of the workload balancing policies over a number of test scenes – For task of varying size – For rays exhibiting different degree of spatial coherency (high to medium) In general, cost-based initialized versions outperform regular versions for large task sizes Results not sensitive to degree of spatial coherency of tested rays Cost-based Workload Balancing for Ray Tracing on a Heterogeneous Platform

9. Conclusions and Future Work Cost-based initialization of task system can improve balancing efficiency of strategies. – The most benefit is gotten by the static strategy (comparable to dynamic balancing) – Dynamic strategies also showed improved balancing efficiency Cost based approach is particularly attractive for coarse grained task systems. – Best results achieved for large size tasks Work limited by degree of variability that rays can have. Rays with low spatial coherency pose a challenge due to the estimation overhead they impose. – Approach cannot be used in its current state for some rendering algorithms – We believe that fast ray reordering can help in this regard We have not considered yet using directly the costs for GPU workload balancing – An implementation of distributed queues on GPUs might also get some benefit from the estimated costs Cost-based Workload Balancing for Ray Tracing on a Heterogeneous Platform