Distributed feature extraction

Time varying simulations are common in many scientific domains to study the evolution of phenomena or features. The data produced in these simulations is massive. Instead of just one dataset of 512³ or 1024³ (for regular gridded simulations) there could now be hundreds to thousands of timesteps. For datasets with evolving features, feature analysis and visualization tools are crucial to help interpret all the information. For example, it is usually important to know how many regions are evolving, what are their lifetimes, do they merge with others, how does the volume/mass change, etc. Therefore, feature based approaches, such as feature tracking and feature quantification are needed to follow identified regions over time. In our previous work, we have developed a methodology for analyzing time-varying datasets which tracks 3D amorphous features as they evolve in time. However, the implementation is for single-processor non-adaptive grids and for massive multiresolution datasets this approach needs to be distributed and enhanced. In this paper, we describe extensions to our feature extraction and tracking methodology for distributed AMR simulations. Two different paradigms are described, a "fully distributed" and a "partial-merge" strategy. The benefits and implementations of both are discussed.