National Energy Research Scientific Computing Center

14 NERSC ANNUAL REPORT 2015
The Dark Energy Spectroscopic
Instrument (DESI) will measure
the effect of dark energy on the
expansion of the universe by
obtaining optical spectra for
tens of millions of galaxies
and quasars. DESI will be
conducted on the Mayall
4-meter telescope at Kitt Peak
National Observatory in Arizona
(shown here) starting in 2018.
Image: NOAO/AURA/NSF
Extreme-Scale and Data-Intensive Platforms
An increasing number of scientific discoveries are being made based on the analysis of large volumes
of data coming from observational science and large-scale simulation. These analyses are beginning to
require resources only available on large-scale computing platforms, and data-intensive workloads are
expected to become a significant component of the Office of Science’s workload in the coming years.
For example, scientists are using telescopes on Earth—such as the Dark Energy Spectroscopic
Instrument (DESI) managed by LBNL—and mounted on satellites to try to “see” the unseen forces,
such as dark matter, that can explain the birth, evolution and fate of our universe. Detailed,
large-scale simulations of how structures form in the universe will play a key role in advancing our
understanding of cosmology and the origins of the universe.
To ensure that future systems will be able to meet the needs of extreme-scale and data-intensive
computing in cosmology, particle physics, climate and other areas of science, NERSC has
undertaken an extensive effort to characterize the data demands of the scientific workflows run by its
users. In 2015 two important image-analysis workflows—astronomy and microtomography—were
characterized to determine the feasibility of adapting these workflows to exascale architectures.
Although both workflows were found to have poor thread scalability in general, we also determined
that adapting these workflows to manycore and exascale node architectures will not be fundamentally
difficult; coarse-grained threading and the use of large private buffers were the principal limits to
scalability and optimization strategies to address these bottlenecks already exist.
The I/O requirements of these workflows were also characterized, and in general their I/O patterns
are good candidates for acceleration on a flash-based storage subsystem. However, these workflows do
not re-read data heavily, and it follows that transparent-caching architectures would not deliver the