Hardware Reference
In-Depth Information
with the single-shared-file approach, completed the same I/O in 60 seconds.
However, scaling the application up to 8000 processes showed the limitations
of the single file approach, and therefore aggregation with multiple files was
finally used. With aggregation, the above example was completed in 12 sec-
onds, and scalability of the code was maintained for larger simulations up to
3.7 billion cells.
17.3.2 Analysis
Seismography. Recent advances in high performance computing and numer-
ical techniques have facilitated full 3D simulations of global seismic wave prop-
agation at high resolutions. The spectral-element method (SEM) described by
Tromp et al. [15] has recently gained popularity in seismology and now SEM is
widely used for 3D global simulations. The Theoretical & Computational Seis-
mology research group at Princeton University has developed the open-source
SEM package SPECFEM3D GLOBE; see Carrington et al. [3]. Their goal is
to improve the models of Earth's subsurface and kinematic representations of
earthquakes.
There are about 6000 earthquakes recorded as of today, with several thou-
sand seismograms available for each earthquake. About 50 iterations of a long
processing pipeline for each earthquake will be needed to achieve a highly
accurate subsurface model of the whole Earth. In the pipeline, thousands of
seismograms (both observed and synthetic ones) are processed and produced
at each step for each earthquake while several earthquakes are processed con-
currently. Observational data is provided by the IRIS (Incorporated Research
Institutions for Seismology) organization but it has to be processed (cleaned
and aligned) for the comparative analysis with the synthetic datasets. The
current standard data format is a separate binary file for each seismogram,
resulting in millions of small input files and multiples of that number of files
for intermediate synthetic data. Even the largest parallel file systems cannot
provide access to this many files with reasonable performance. Therefore, a
new seismography data format has been developed that stores all seismograms
of an earthquake in a single ADIOS file, which provides ecient access to all
of them while reducing the total number of files by several orders of magni-
tude. This change alone allows the processing pipeline to execute on Titan, a
supercomputer at the Oak Ridge Leadership Facility, using the allocated time
eciently. Previously, the processing pipeline was stalled by waiting on the
reads and writes of small files.
Climate research. Climate modeling applications address one of the most
pressing issues faced by our society. Numerous surveys show increasing trends
in the average sea-surface and air temperatures. A promising method of pre-
dicting the future effects of climate change is to run large-scale simulations
of the global climate many years into the future. In such simulations, clima-
tologists discretize our entire globe, setting up differential equations to model
 
Search WWH ::




Custom Search