All, I appreciate your help here. I'm traveling all this week and next. I'll forward these comments to some members of my team, but I won't be able to test/look at anything specific to the HPC configuration until I get back. I can say that during my troubleshooting, I did determine that MPI_STATUS_SIZE =3 when I compiled with -m64 and -fdefault-integer-8 options. However, the memory corruption issues only resolved when I used MPI_STATUS_IGNORE.
Cheers, --Jim On Tue, Nov 12, 2013 at 3:25 PM, George Bosilca <bosi...@icl.utk.edu> wrote: > > On Nov 12, 2013, at 19:47 , Jeff Squyres (jsquyres) <jsquy...@cisco.com> > wrote: > > > On Nov 12, 2013, at 4:42 AM, George Bosilca <bosi...@icl.utk.edu> wrote: > > > >>> 2. In the 64 bit case, you'll have a difficult time extracting the MPI > status values from the 8-byte INTEGERs in the status array in Fortran > (because the first 2 of 3 each really be 2 4-byte integers). > >> > >> My understanding is that in Fortran explicitly types variables will > retain their expected size. Thus, instead of declaring > >> > >> INTEGER :: status[MPI_STATUS_SIZE] > >> > >> one should go for > >> > >> INTEGER*4 :: status[MPI_STATUS_SIZE] > >> > >> This should make it work right now. > > > > You are correct. That’s a good workaround. > > Not good … temporary ;) > > >> However, it is a non-standard solution, and we should fix the status > handling internally in Open MPI. > >> > >> Looking at the code I think that correctly detecting the type of our > ompi_fortran_integer_t during configure (which should be a breeze if the > correct flags are passed) should solve all issues here as we are protecting > the status conversion between C and Fortran. > > > > Not quite. We do already correctly determine ompi_fortran_integer_t as > a C "int" or "long long" (that's what I saw yesterday when I tested this > myself). > > > > However, the key here is that MPI_STATUS_SIZE is set to be the size of a > ***C*** MPI_Status (but expressed in units of Fortran INTEGER size -- so in > the sizeof(int)==sizeof(INTEGER)==4 case, MPI_STATUS_SIZE is 6. But in the > sizeof(int)==4, sizeof(INTEGER)==8 case, MPI_STATUS_SIZE is 3. > > > > That being said, we *could* change this so that MPI_STATUS_SIZE is > always 6, and have the C<—>Fortran status routines just do the Right Thing > depending on the size/type of ompi_fortran_integer_t. > > Indeed. We can have an Fortran MPI_Status (only in the Fortran interface) > that will be 3 ompi_fortran_integer_t, and alter the translation macros to > do the right thing (translate from C int to the chosen Fortran int). > > > Either way, as you say, it's a nonstandard solution. So I don't know > which way is "more correct". On the one hand, we've had it this way for > *years* (so perhaps there's code out there that uses the George workaround > and is working fine). But OTOH, it’s different than what you would have to > do in the non-dash-i8 case, and so we should make MPI_STATUS_SIZE be 6 and > then Fortran code will work identically (without INTEGER*4) regardless of > whether you used -i8 or not. > > Honestly, I think that most users will expect that an MPI compiled with > -i8 will have the status as a 3 8 bytes integers and not some other weird > combination depending on another layer of the library (compiled in a > language lacking the subtlety of -i8 ;)). > > George. > > > > > > Shrug. > > > >> Jim, can you go in the include directory on your Open MPI installation > and grep for the definition of ompi_fortran_integer_t please. > >> > >> George. > >> > >> > >> _______________________________________________ > >> users mailing list > >> us...@open-mpi.org > >> http://www.open-mpi.org/mailman/listinfo.cgi/users > > > > > > -- > > Jeff Squyres > > jsquy...@cisco.com > > For corporate legal information go to: > http://www.cisco.com/web/about/doing_business/legal/cri/ > > > > _______________________________________________ > > users mailing list > > us...@open-mpi.org > > http://www.open-mpi.org/mailman/listinfo.cgi/users > > _______________________________________________ > users mailing list > us...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/users >
