Multi-GPU Load Balancing for In-situ Visualization

Real-time visualization is an important tool for immediately inspecting results for scientific simulations. Graphics Processing Units (GPUs) as commodity computing devices offer massive parallelism that can greatly improve performance for data-parallel applications. However, a single GPU provides limited support which is only suitable for smaller scale simulations. Multi-GPU computing, on the other hand, allows concurrent computation of simulation and rendering carried out on separate GPUs. However, use of multiple GPUs can introduce workload imbalance that decreases utilization and performance. This work proposes load balancing for in-situ visualization for multiple GPUs on a single system. We demonstrate the effectiveness of the load balancing method with an N-body simulation and a ray tracing visualization by varying input size, supersampling, and simulation parameters. Our results show that the load balancing method can accurately predict the optimal workload balance between simulation and ray tracing to significantly improve performance.