News & Announcements
January 18, 2013
"Argonne researcher contributes to new book on high-performance visualization"
Recent advances in supercomputing have resulted in a tremendous growth in the size and complexity of data produced by simulations. Scientific visualization is rapidly becoming an essential tool in understanding this data and gaining insights that will enable breakthrough science.
The new book High Performance Visualization: Enabling Extreme-Scale Scientific Insight (E. W. Bethel, H. Childs, and C. Hansen, eds., CRC Press, 2012) presents the state of the art in scientific visualization---ranging from analysis tools and algorithms, to open source software packages and applications, to techniques for emerging exascale architectures. The book comprises 21 chapters by leading experts from universities, industry, and national laboratories worldwide.
Tom Peterka, an assistant computer scientist in Argonne’s Mathematics and Computer Science Division, is coauthor of two chapters: Parallel Image Compositing Methods (chapter 5) and Parallel Integral Curves (chapter 6).
Image compositing is one of the fundamental parts of high-performance visualization on large-scale parallel machines. It is also extremely costly because it requires communication among many processes. In Chapter 5, Peterka and his coauthor from UC Davis briefly review image compositing algorithms and optimizations over the past decade. They then focus on recent advances that have enabled strong scaling performance on up to 164,000 processes on supercomputers such as the IBM Blue Gene and Cray XT, compared with legacy algorithms whose performance leveled out at a few thousand processes.
In Chapter 6, Peterka and colleagues from Oak Ridge National Laboratory and the University of Kaiserslautern explore the use of integral curves as a powerful visualization technique for understanding vector fields. The challenge here is that such curves place demands on almost all components of a computational system, including memory, processing, input/output, and communication. Because no single scalable algorithm is suitable for all types of vector fields and data set sizes, several parallel algorithms are presented. Each achieves good scalability for very large datasets on large computing resources.