Hardware Reference
In-Depth Information
devise a parallel data decomposition strategy, no matter which library will be
used. Often, this decomposition strategy is the trickiest part. Once a strategy
is in place, the I/O library used is secondary, and can often be hidden by
an abstraction layer (climate codes, for example, use PIO [2] for just this
purpose). The glib answer to \Which I/O library should I use?" is \It doesn't
matter, as long as you use something."
Parallel-NetCDF still provides the standard for a low-overhead I/O library.
The classic netCDF file format imposes some restrictions, but these restric-
tions mean the file layout is known ahead of time. The library does not need
additional coordination among processes to know where to place data.
Other I/O libraries can borrow ideas proven in Parallel-NetCDF. When
serial netCDF introduced parallel I/O to its API, they could use Parallel-
NetCDF's approach for incorporating parallel I/O parameters such as the MPI
communicator and how to express collective I/O. An upcoming release from
the HDF5 project will contain a \multi-dataset" family of routines following
the Parallel-NetCDF approach.
In the flourishing ecosystem of parallel I/O libraries, Parallel-NetCDF still
represents a compelling option for codes using regular arrays to describe their
data, or are willing to transform their data into such arrays. As scientific
data models increase in complexity, the relative simplicity of Parallel-NetCDF
might one day no longer be appropriate. As we in the I/O library community
develop successor libraries for more sophisticated application data models,
Parallel-NetCDF will remain the standard for a lightweight abstraction layer.
15.7 Additional Resources
More Parallel-NetCDF information can be found at the following places:
www.mcs.anl.gov/parallel-netcdf The Parallel-NetCDF home page con-
tains an overview of the package, instructions for joining the mailing
list, tutorials, documentation, and bug tracker.
cucis.ece.northwestern.edu/projects/PnetCDF Northwestern Univer-
sity's Parallel-NetCDF page contains additional material and documen-
tation.
Parallel-NetCDF requires only an MPI-IO implementation, and so is avail-
able on nearly every parallel computer. Users should consult local documen-
tation for details of any site-specific quirks.
