News & Announcements

January 4, 2013

"Meshing the Universe"

Media Contacts: Gail Pieper, pieper@mcs.anl.gov

Tessellations are meshes made of cells with no overlapping and no gaps. Scientists use tesselations for representing point data because they convert a sparse discrete sample into a dense continuous field that can be used to compute cell statistics and identify features. Tessellating the output of petascale simulations is a data-intensive process and is best executed in parallel on the same supercomputer as the simulation, in order to save memory. Until recently, however, no large-scale parallel tessellation tools have existed.

Researchers at Argonne National Laboratory, with collaborators at the University of Tennessee Knoxville and Kitware Inc., have now developed a method for computing tessellations in parallel. The new method has been implemented in a prototype library, called tess, and its use demonstrated during simulations with a petascale cosmology code on the IBM Blue Gene/P at Argonne.

Tess produced statistical summaries of volume and density distributions and revealed regions of irregular low-density voids amid clusters of high-density halos. Moreover, since tess can generate tessellations at various points during the simulation, the researchers were able to study the evolution of voids. The results were consistent with the governing theory that predicts the formation of high- and low-density structures over time.

“We knew that tessellations were useful in identifying cosmological features,” said Tom Peterka, an assistant computer scientist in Argonne’s Mathematics and Computer Science Division. “But this is the first time such a technique is being used to analyze results in situ with a full N-body simulation.”

Embedding the analysis in the execution of the simulations naturally raises concerns about both performance and scalability. Benchmark runs with tess on the Blue Gene/P put these concerns to rest, however. The tessellation time was only 1–10% of the total run time, with strong and weak scaling similar to the original simulation code.

The research team is hard at work developing a suite of cosmology tools for visualization and analysis. According to Peterka, tess is the first member of a larger in situ analysis framework that will include merger trees, multistream classification, and halo finding.

This research was supported by the U.S. DOE SciDAC Scalable Data Management, Analysis, and Visualization Institute.