old code for GPFS support in romio321 needs stat64=>stat in ad_gpfs_open.c

[walkup]

Some large supercomputing systems have IBM Spectrum Scale (a.k.a GPFS) for the parallel file-system. The code in openmpi-4.1.2 for GPFS support in romio321 has one minor issue. It will not compile unless one changes "stat64" to "stat" in source file ad_gpfs_open.c. To enable GPFS support in the current openmpi's romio321 path, one must pass the following option to the top-level configure script : configure ... --with-io-romio-flags='--with-file-system=gpfs', then "make" will attempt to build support for GPFS. Doing this helps significantly with MPI-IO performance on systems that have GPFS, when adding -mca io romio321 to the mpirun command. My experience has been that romio321 when properly configured provides better performance on GPFS than the current "ompio" layer. I am familiar with this issue because I was involved in getting romio321 support into IBM' Spectrum MPI for the Summit system at Oak Ridge.

I mention that current versions of MPICH have somewhat newer ROMIO code for GPFS, which does not have this issue. I don't think it is necessary to update the GPFS code in openmpi, other than to fix the issue with stat64 => stat in ad_gpfs_open.c .

[gpaulsen]

@walkup we can take a look at this.

[markalle]

Yeah, I checked and it looks like both v4.0.x and v4.1.x are using romio321 with a full copy in OMPI. I know at some point there's a goal to have newer OMPI just pulling romio straight from mpich master, but I don't think v4.0.x / v4.1.x will be doing that. In any case they're not yet doing that. So I believe these trees are still taking one-off fixes into our romio321 copy.

So I made a couple PRs for changing stat64 to stat:
#9854
#9855