Quick-VDR: Out-of-Core View-Dependent Rendering of Gigantic Models
IEEE Transactions on Visualization and Computer Graphics
Memory sharing for interactive ray tracing on clusters
Parallel Computing - Parallel graphics and visualization
| Hi-index | 0.00 |
We present a practical system to visualize large datasets interactively on commodity PCs. Interactive visualization has applications in many areas, including computer-aided design, engineering, entertainment, and training. Traditionally, visualization of large datasets has required expensive high-end graphics workstations. Recently, with the exponential trend of higher performance and lower cost of PC graphics cards, inexpensive PCs are becoming an attractive alternative to high-end machines. But a barrier in exploiting this potential is the small memory size of typical PCs. Our system uses new out-of-core techniques to visualize datasets much larger than main memory. In a preprocessing phase, we build a hierarchical decomposition of the dataset using an octree, precompute coefficients used for visibility determination, and create levels of detail. At runtime, we use multiple threads to overlap visibility computation, cache management, and rasterization. The structure of the octree and the visibility coefficients are kept in main memory. The contents of the octree nodes are loaded on demand from disk into a cache. To find the visible set, we use a fast approximate algorithm followed by a hardware-assisted conservative algorithm. To hide I/O latency, a separate thread prefetches nodes that are likely to become visible. We also describe a sort-first parallel extension of the system that uses a cluster of PCs to drive a high-resolution, multi-tile screen. A client process interacts with the user, and a set of server processes render the screen tiles. To avoid sending the entire dataset from the client to the severs every frame, the servers access the dataset from a shared file system or from a local copy on disk. Putting the 1/O load on the server side makes the network bandwidth requirements low and the architecture scalable. Using a cluster of 16 PCs, the system can generate high resolution images (12 megapixels) of large datasets (4 gigabytes) at interactive frame rates (10 frames per second). Thus, our system is a cost-effective alternative to high-end machines, and can help bring visualization of large datasets to a broader audience.