Hardware Reference
In-Depth Information
MC4 is a parallel particle-mesh cold dark matter (CDM) code based on
MC2 [6]. MC4 currently achieves scalable parallel performance for maximum
values of N
p
articles = N
g
rid = 4096
3
using O(10; 000) cores, and uses H5hut
to output several terabytes of particle and field data per timestep. Work is
progressing toward problem sizes of N
p
artcles = N
g
rid = 8192
3
or even larger
using O(100; 000) cores.
The HDF5-FastQuery [4] library enables multi-dimensional indexing
and searching on large H5Part datasets by leveraging an ecient bitmap in-
dexing technology called FastBit [16]. Storing bitmap indices directly in the
H5hut file significantly speeds up access to subsets of multi-dimensional data,
especially through range queries. This has been applied to interactive explo-
ration of simulations of laser wakefield particle accelerators. Data from the
simulation is stored with H5Part and indexed with HDF5-FastQuery. High-
energy particles in the laser wakefield can then be eciently selected by thresh-
olding for large momenta in the direction of the beam [10].
16.4.3
CGNS
16.4.3.1
Design
In 1994, NASA, Boeing, and McDonnell Douglas started a common effort
on computational fluid dynamics data standardization. In 1999 they created
the CGNS Steering Committee, a voluntary public forum made up of inter-
national representatives from government and private industry, with the aim
of gathering the CFD community on a standard for data interoperability.
Today, this committee includes the main actors of CFD from the USA and
Europe; they regularly meet in the framework of the American Institute of
Aeronautics and Astronautics (AIAA) who now holds the standard as one of
its recommended practices.
The CGNS standard [3] defines both a data model, the so-called SIDS
(Standard Interface Data Structure), and its implementations on physical
medium. The data model, a public document, is dedicated to CFD (computa-
tional fluid dynamics) computations but has enough genericity to host other
CFD-coupled simulations such as CFD/CAA (Aeroacoustics) or CFD/CSM
(structural mechanics). The CGNS data model handles few nodes but with a
strongly typed grammar. Figure 16.5 shows a simple program that creates a
CGNS file. It defines a complete set of data for computation, the definition
of the equations, the flow fields and parameters, the meshes and associated
topological information, and other auxiliary data.
The data model is the basis for data representation. Once an entity that
the standard defines is described in the CGNS/SIDS document, it must be
translated to store on a physical medium. There are three CGNS/SIDS map-
pings: ADF, HDF5, and Python. A mapping defines the translation between
the tree data model of CGNS/SIDS and the actual nodes and set of nodes
