Author: Kwan-­Liu Ma, University of California, Davis

We are in the era of petascale computing, and exascale computing is around the corner. The incredible processing power and capacity of today’s supercomputers enables scientists to answer some of the most difficult questions in sciences using large-scale, high fidelity simulations. To gain new insights into the simulations and foster new discoveries, scientists must be able to make sense of the full extent of the simulation output at the highest possible resolution. For this purpose, visualization has become an indispensible tool. However, today each run of a petascale simulation typically outputs several hundred terabytes of data to disk. Transferring data at this scale over wide-area networks to the scientist’s laboratory for post-processing visualization and analysis is not an option. Even if the data files can be moved to the scientists, existing desktop data analysis and visualization tools cannot effectively handle such large-scale data. A viable solution is not to move the raw data; instead, scientists will want to make use of the high-performance visualization facilities, high-speed networks, and parallel file systems at the supercomputing center to transform the data into compact images or geometric objects for fast delivery to the scientist’s desktop for viewing and manipulation. Such distance, remote visualization will become common practice in the near future, though adequate technologies from data transformation methods to remote visualization UI remain to be developed.

There are three ways to do distance visualization: move data, move extracts, or move images. Moving raw data is not feasible for obvious reasons. Moving extracts, a small fraction of the data characterizing some features of interest in the data, can significantly reduce the amount of data that must be transferred over networks, but rendering the extracts at interactive rates could demand a powerful computer and a sophisticated user interface. Moving images can support two different modes of visualization: imagebased rendering, which could require a large number of images, and video viewing, which is browsing images in sequential order. The former potentially incurs high cost in computing and transferring the images while the latter only allows very limited explorability, i.e., playing forward and backward.

I would like to introduce the concept of explorable image. Explorable image, a compact intermediate representation of the data for deferred interaction, as a media for distance visualization can allow scientists to visualize their data anywhere anytime on any display device, which may not have a powerful processor and a large memory space, like an iPad. Our preliminary work [1,2,3] shows that explorable image allows interactive exploration in the:

• Spatial and temporal domains of the data
• Transfer function space
• Rendering parameter space without re-rendering and access to the original data.

Without using compression, an explorable image is exactly of the size of 16 ordinary images, much smaller than the original volume data. A conventional image-based method would require hundreds of images to achieve the same level of explorability. Explorable image is without limitation. The resulting visualization is an approximation to the regularly rendered one. Furthermore, the degree of explorability has a limit, but it already proves useful (see Figure 1).

Distance visualization is not a new topic. Previous techniques and settings [4,5] need to be re-evaluated for their scalability and usability. As portable display devices and wireless Internet are commonplace, distance visualization with explorable image becomes very attractive, suggesting scientists to rethink how they do their routine data analysis and visualization. Besides being a media for remote visualization, explorable image would also be a cost-effective previewing method for scientists to find the optimal view points, transfer functions, etc. for making a high-resolution animation of the whole dataset. If we consider the inevitable subsampling of the large data generated by the simulation at extreme scale, explorable image provides a way to capture some essential aspects of the data that must be discarded due to scarce storage space. Visualizing at exascale demands new thinking and explorable image points in a promising direction.



Figure 1. Visualization of an argon bubble data set using explorable image. (Images were created by Anna Tikhonova. The dataset was provided by Center of Computational Science of the Lawrence Berkeley National Laboratory.)

References:

1. Tikhonova, A., Correa, D. D., and Ma, K.-L. 2010. Explorable Images for Visualizing Volume Data.
In Proceedings of IEEE Pacific Visualization Symposium, pp. 177–184.
2. Tikhonova, A., Correa, D. D., and Ma, K.-L. 2010. An Exploratory Technique for Coherent
Visualization of Time-varying Volume Data. Computer Graphics Forum (also EuroVis 2010
Proceedings), Volume 29, Number 3, pp. 783-792.
3. Tikhonova, A., Correa, C. D. , and Ma, K.-L. 2010. Visualization by Proxy: A Novel Framework for Deferred Interaction with Volume Data. IEEE Transactions on Visualization and Computer Graphics (also Visualization 2010 Conference Proceedings), Volume 16, Number 6.
4. Foster, I., Insley, J., von Laszeqski, G., Kesselman, C, and Thiebaux, M. 1999. Distance
Visualization: Data Exploration on the Grid. Computer, Volume 32, Number 12, pp. 36-43.
5. Cedilnik, A., Geveci, B., Moreland, K., Ahrens, J., and Favre, J. 2006. Remote Large Data
Visualization in the ParaView Framework. In Proceedings of Eurographics Symposium on Parallel Graphics and Visualization, pp. 163-170.

About the author:

Kwan-Liu Ma

is a professor of computer science and Chair of the Graduate Groupin Computer Science at the University of California, Davis. He directs the Visualization and Interface Design Innovation research group and the SciDAC Ultra-Scale Visualization Institute. He received his Ph.D. in computer science from the University of Utah in 1993. His research interests include computer graphics, visualization, user interface design, and high-performance computing.

Professor Ma can be reached via email: ma@cs.ucdavis.edu.