Multi-scale data visualization for computational astrophysics and climate dynamics at Oak Ridge National Laboratory

Abstract

Computational astrophysics and climate dynamics are two principal application foci at the Center for Computational Sciences (CCS) at Oak Ridge National Laboratory (ORNL). We identify a dataset frontier that is shared by several SciDAC computational science domains and present an exploration of traditional production visualization techniques enhanced with new enabling research technologies such as advanced parallel occlusion culling and high resolution small multiples statistical analysis. In collaboration with our research partners, these techniques will allow the visual exploration of a new generation of peta-scale datasets that cross this data frontier along all axes. Introduction: A frontier is shared by several SciDAC computational science domains that generally correlates to aggregate dataset size. This frontier is a multidimensional collection of points each representing output from a particular simulation code. The loci of these points are determined by factors including spatial sampling, temporal sampling, field dynamic range, number of independent fields, and spatial coherence. It is this factorization that characterizes a simulation's data and determines the suitable visualization tools and techniques for exposing the underlying science. Many production visualization techniques begin to fail when considering data that cross this frontier. As such, this boundary marks the departure of visualization techniques that are production and requires specialized research efforts. Choosing visualization techniques for a particular science domain is a process of stepwise refinement that must be verifiable, expository, and communicable with regard to the science. The visualization must bear a clear quantitative relationship to the data, deftly expose and support some element of scientific relevancy in the data, and easily facilitate a transfer of knowledge. Both climate and astrophysics simulations are generating data on the order of tens to hundreds of terabytes, and each dataset composition is predisposed to certain techniques. Science drivers and computational models are dynamic, however, and dataset compositions are often in a state of flux. Data analysis techniques must adapt to reflect these changes. We are adapting these techniques for the changing needs of the science communities as they create datasets with increasing size, complexity, and fidelity. We present both production and research techniques that have demonstrated value with respect to scientific discovery within the respected domains. These techniques combine traditional visualization, parallel hardware exploitation, and coupled statistical analysis. It is only through this directed research toward large-scale data understanding that the realm of production visualization may be expanded to manage these datasets. Computational Astrophysics Visualization: The SciDAC-sponsored TeraScale Supernova Initiative (TSI) is one of the principal users of CCS resources. TSI is a multidisciplinary, multi-institution effort involving the integrated efforts of ast...