High-level I/O libraries, like Parallel-NetCDF, take MPI-IO's features (and complexity) and package them under an interface that should feel more natural to computational scientists. Scientific applications operate at a higher level of abstraction typically think about arrays or matricies, not "bytes on disk".
Parallel-NetCDF sees a lot of use in the climate and weather communities. It has a fairly simple data model and programming interface. That simplicity has made it possible for us to experiment with classes of interfaces. One approach that has worked well for us over the years is "operation combining" -- using the non-blocking interface to in a sense pre-declare I/O operations. PnetCDF will, once the application "waits" for all I/O to finish, bundle up all operations into a single more efficient operation.
Instead of using IOR, I've written a small test program (pnetcdf-write.c),
where each process writes five variables to a netCDF dataset in parallel.
Those processes are each writing their own column of an array.
We will compare two approaches:
- "Blocking": the program will collectively write each variable one at a time.
- "non-blocking": the program will use the PnetCDF non-blocking interfaces to post the five writes, then wait for them to complete.
Some details are machine-specific. The polaris/pnetcdf-write-compare.sh
script is what I submitted to the Polaris job scheduler.
Instead of generating plots of the benchmark's numbers, we are going to use Darshan .
