Researchers working with the Exascale Computing Project (ECP) have demonstrated a novel moment invariant pattern detection algorithm to drive in situ pattern recognition during simulations running on high-performance computing architectures. This development addresses the challenge of optimizing computation and communication (i.e., input/output) speeds by enabling a dataset to comb itself for data patterns of interest that can be stored as a subsample for post-hoc analysis. The researchers have released the algorithm module as part of the open-source Visualization Toolkit, or VKT. Their findings were published in the June 2020 edition of Eurographics Proceedings.
The researchers tested their algorithm with two computational approaches—straightforward integration and fast Fourier transform (FFT)— using datasets resulting from three simulations: an ocean system (MPAS Ocean), an asteroid impact on deep ocean water (yA31), and the evolution of a system of discrete dark matter particles gravitationally coupled to an inviscid ideal fluid in an expanding universe (created from an ECP ExaSky code, Nyx). The researchers compared the methods’ usefulness in optimizing run time and data volume at exascale, demonstrating the algorithm’s performance in the distributed data environment and acceleration of the algorithm using the FFT method. Using the computational fluid dynamics modeling tool, MFiX-Exa, the researchers demonstrated an in situ use case of their algorithm’s feature detection capabilities and data reduction benefits.
While both straightforward integration and FFT approaches successfully ran the algorithm in a distributed setting, FFT was shown to provide superior run time performance with a range of input parameter combinations. FFT, however, requires structured datasets and more storage space than straightforward integration. The researchers concluded that a combination of the two approaches is optimal: using FFT when applicable while providing the option of straightforward integration. The work was funded under 2.3.4.16 ALPINE, a project within the ECP that supports data analysis and visualization for science discovery.
Tsai, Karen C., Roxana Bujack, Berk Geveci, Utkarsh Ayachit, and James Ahrens. 2020. “Approaches for In Situ Computation of Moments in a Data-Parallel Environment.” Eurographics Symposium on Parallel Graphics and Visualization. https://doi.org/10.2312/PGV.20201075.
