Re: OpenMPI 4.0.5 segfault with mpi_file_open() [ OpenBSD 6.8 release & current ]

I looked into the IO issue, one hint is found with

export OMPI_MCA_io_base_verbose=40
mpirun -np 1 -H localhost:1 ./mpitest
...
mca: base: components_open: found loaded component ompio
mca: base: components_open: component ompio open function successful
mca: base: components_open: found loaded component romio321
mca: base: components_open: component romio321 open function successful
...

So in fact there are two IO components available: ompio and romio321.
The
first is selected (and fails). If you select the second, mpitest works:

$ mpirun -np 1 -H localhost:1 --mca io romio321 ./mpitest
This is process 1 / 1

You can make this permanent with

export OMPI_MCA_io=romio321

added to your login scripts.

HTH. (OpenMPI is too complicated for it's own good.)


John



On 2021-02-03 09:51, j@bitminer.ca wrote:
> Hi Martin,
>
> I haven't run into your MPI_File_open issue (don't use it), but
> your code does fail for me too in the same way.
>
>> $> mpirun -np 1 -H localhost:1 ./fmpitest
>> fmpitest:/usr/local/lib/libmpi.so.5.0: ./fmpitest : WARNING:
>> symbol(mpi_fortran_statuses_ignore_) size mismatch,
>> relink your program
>> fmpitest:/usr/local/lib/libmpi.so.5.0: ./fmpitest : WARNING:
>> symbol(mpi_fortran_status_ignore_) size mismatch,
>> relink your program
>
> The Fortran symbol error you see is common and I'm not sure of the
> cause. I did look into it at one point and decided all definitions
> were
> in fact identical, so it might be a weird compiler+linker issue.
>
> It's never been symptomatic beyond the warning so I ignore it.
>
>
> --John