Hard-sphere collision simulations with multiple GPUs, PCIe extension buses and GPU-GPU communications

• Authors: K. A. Hawick;D. P. Playne
• Venue: AusPDC '12 Proceedings of the Tenth Australasian Symposium on Parallel and Distributed Computing - Volume 127
• Year: 2012

Simulating particle collisions is an important application for physics calculations as well as for various effects in computer games and movie animations. Increasing demand for physical correctness and hence visual realism demands higher order time-integration methods and more sophisticated collision management algorithms. We report on the use of singe and multiple Graphical Processing Units (GPUs) to accelerate these calculations. We explore the performance of multiple GPUs (m-GPUs) housed on a single PCIe bus as well as the use of special purpose PCIe bus extender technology using GPU housing chassis systems such as Dell's C410x PowerEdge. We describe how a hard sphere collision system with gravitational interactions was developed as a benchmark. We compare the performance of various GPU models and show how algorithms that use GPU-GPU communications with NVidia's Compute Device Unified Architecture (CUDA 4) can considerably aid communications amongst multiple GPUs working on a single simulated particle system.