Imaginations unbound in a petascale world

08.23.07 -

The pursuit of petascale computing is well underway. And while part of that quest will be learning how to exploit these machines for science, engineering, and other research efforts, one thing we do know already: The volume of data produced by petascale simulations will lead to rich and highly detailed visualizations.

How will these visualizations be created?

When possible, it will be best to use basic graphics libraries. NCSA's Advanced Visualization Laboratory (AVL) team has developed our capabilities for high-volume data simulations using the open source applications Visualization ToolKit and ParaView, and we will further extend these capabilities to petascale. Visualization applications for petascale machines must support supercomputing applications, including tiled display and distributed parallel processing.

Research has shown that for rendering very large polygonal datasets (for example, molecular dynamics), traditional rasterization methods are not amenable to parallelization. For those datasets, ray tracing techniques are being investigated; while they are more computationally intensive, they are much more scalable and can provide higher quality visualizations. We plan to explore large parallel ray tracing techniques using up to 1,000 cores to perform high-quality software rendering at interactive rates.

For scalar fields, direct volume rendering provides very detailed and informative images. Much work has been done in parallelization of volume renderers, both in the simple case of uniform grids and also for unstructured tetrahedral meshes. At NCSA, we are evaluating the quality and scalability of applications for parallel volume rendering.

An increasing number of simulations are using adaptive mesh refinement (AMR) techniques in order to increase their accuracy. The output from such simulations presents new visualization challenges. NCSA responded to those challenges by developing Amore, an AMR volume visualization pipeline which enables smooth, volume-rendered animations from time-refined AMR datasets. Amore was instrumental in our ability to create high-resolution scientific visualizations for the planetarium dome show "Black Holes: The Other Side of Infinity" that is currently showing in Denver and Chicago.

Petascale science brings new work scenarios

Petascale computing should be able to support three visualization processing scenarios, including concurrent visualization with intermediate processing such as data reduction, feature extraction, statistical aggregates, and analysis; on-the-fly interactive visualization using parallel software techniques or hardware accelerators; and post-simulation visualization.

In concurrent visualization, a running computation passes intermediate results along to a separate parallel visualization task that performs feature extraction and data reduction. We can also make a catalog of images per time step for later review, using multiple camera angles, or make parameterized sweeps through transfer functions or isosurface values.

NCSA is exploring the scalability of several parallel interactive visualization applications that support client/server visualization. The use of graphics accelerators will make image rendering faster by several orders of magnitude and can also assist in image compression to reduce network bandwidth for streaming visualization to remote users.

Scientists will be best served by viewing these visualizations on large-scale displays, such as tiled walls or very high resolution flat panels with 10 to 100 megapixels. Many remote users are likely to acquire high-resolution displays to explore finely resolved physical simulations, and the nation's high-speed networks will be used to transfer image data to their display systems.

The AVL team believes that research shouldn't languish in the lab. Science and engineering insights must be shared with the widest possible audience so others can participate in the thrill of discovery and share the enthusiasm for scientific inquiry. And we anticipate that visualizations from petascale research will indeed be thrilling.