Open ax3l opened 9 years ago
Tuning HDF5 for Lustre File Systems (2010)
... However, when operating
in a parallel application, this operation resulted in a call
to MPI_File_set_size, which currently has very poor
performance characteristics on Lustre file systems. 1 Because
an HDF5 file’s size is not required to be accurately set until
the file is closed, this operation was removed from H5Fflush
and added to the code for closing a file.
And the good old Parallel HDF5 Hints
@psychocoderHPC found in a
past evaluationpaper (ref?), that the time for an H5Fclose can be reduced by about a factor 3 if the file-consistency check (H5F.c) is skipped in the HDF5 implementation.Consequences for MPI-I/O output could be evaluated. It might be worth to consider adding a hook/option in the HDF5 API and to propose it for upcoming versions of the library.
There are also various close degrees available such as
H5F_CLOSE_WEAK
while the default for MPI-I/O isH5F_CLOSE_SEMI
.