Scientific Visualization Gallery
 |
| Large-Scale Distributed GPU-Based Visualization Framework Greg Abram, Byungil Jeong, Greg P. Johnson, Paul Navratil, Kelly Gaither, Texas Advanced Computing Center. Work in collaboration with Diego Donzis, Texas A&M and P.K. Yeung, Georgia Tech. We have enhanced NVIDIA's CUDA Isosurfacer to perform twice as fast by using one third of the GPU memory, and enabling efficient overlapping of GPU operations with I/O operations and sort-last image compositing to achieve high throughput, in-core rendering. We found that the enhanced isosurfacer achieves an approximate speedup for 4.5x over CPU-based visualization methods on a 2048^3 scalar volume. The images show enstrophy data, isosurfaced using the modified NVIDIA CUDA isosurfacer. It was created using 64 nodes of Longhorn. It took 2.5 minutes to isosurface 935M triangles, 68G cells, which correspond to 256 GB of data. Additionally, it took 9 seconds to render the images on 128 GPUs. The two frames show the assignment of data to process by color. |