Hardware Reference
In-Depth Information
FIGURE 18.3: Ecacy of topology-aware aggregation, subfiling, and compres-
sion on the I/O performance of HACC Cosmology Simulation.
improvements to each of the components, from compute nodes to storage
disks.
18.3.2 Data Staging for FLASH Astrophysics
The FLASH code [3] is an adaptive mesh, parallel hydrodynamics code
developed to simulate astrophysical thermonuclear flashes in two or three di-
mensions, such as Type Ia supernovae, Type I X-ray bursts, and classical
novae. It solves the compressible Euler equations on a block-structured adap-
tive mesh. FLASH provides an AMR grid using a modified version of the
PARA-MESH package and a Uniform Grid (UG) to store Eulerian data.
For this study, GLEAN developers used the Sedov simulation included in
the FLASH distribution. Sedov evolves a blast wave from a delta-function
initial pressure perturbation [2]. The Sedov problem exercises the FLASH in-
frastructure (with AMR and I/O) of FLASH with the minimal use of physics
solvers. It can, therefore, produce representative I/O behavior of FLASH with-
out spending too much time in computations. The application was run in 3D
with 16 3 cells per block. Each block consists of ten mesh variables, and the
problem size is controlled by adjusting the global number of blocks. The study
advanced four timesteps and produced I/O data at every single step so that
most application runtime is spent on I/O. The I/O in each step consisted of the
checkpoint files for restart purposes and plot files for analysis. A checkpoint
file is the output of the complete state of a runtime application, including mesh
data in double precision. A plot file is a user-selected subset of mesh variables
stored in single precision. In these experiments, checkpoint I/O writes all ten
mesh variables. The plot file I/O writes only selected variables of interest (in
these experiments, the first, sixth, and seventh variables).
 
Search WWH ::




Custom Search