Distributed Parallel Extreme Event Analysis in Next Generation Simulation Architectures
Embargo until
Date
2017-05-08
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Johns Hopkins University
Abstract
Numerical simulations present challenges as they reach exascale because they generate petabyte-scale data that cannot be saved without interrupting the simulation
due to I/O constraints. Data scientists must be able to reduce, extract, and visualize the data while the simulation is running, which is essential for in transit and
post analysis. Next generation architectures in supercomputing include a burst buffer
technology composed of SSDs primarily for the use of checkpointing the simulation
in case a restart is required. In the case of turbulence simulations, this checkpoint
provides an opportunity to perform analysis on the data without interrupting the
simulation.
First, we present a method of extracting velocity data in high vorticity regions.
This method requires calculating the vorticity of the entire dataset and identifying
regions where the threshold is above a specified value. Next we create a 3D stencil
from values above the threshold and dilate the stencil. Finally we use the stencil to
extract velocity data from the original dataset. The result is a dataset that is over
an order of magnitude smaller and contains all the data required to study extreme events and visualization of vorticity.
The next extraction utilizes the zfp lossy compressor to compress the entire velocity dataset. The compressed representation results in a dataset an order of magnitude
smaller than the raw simulation data. This provides the researcher approximate data
not captured by the velocity extraction. The error introduced is bounded, and results in a dataset that is visually indistinguishable from the original dataset.
Finally we present a modular distributed parallel extraction system. This system allows a data scientist to run the previously mentioned extraction algorithms in a
distributed parallel cluster of burst buffer nodes. The extraction algorithms are built
as modules for the system and run in parallel on burst buffer nodes. A feature extraction coordinator synchronizes the simulation with the extraction process. A data
scientist only needs to write one module that performs the extraction or visualization
on a single subset of data and the system will execute that module at scale on burst
buffers, managing all the communication, synchronization, and parallelism required
to perform the analysis.
Description
Keywords
big data, distributed, parallel, turbulence, extraction