[mpich-discuss] Using MPICH in Python breaks Fortran MPI_IN_PLACE

Thu Jul 2 05:33:18 CDT 2020

Thank you both for the follow up.  I did engage with Lisandro first, which
you can find here:
https://bitbucket.org/mpi4py/mpi4py/issues/162/mpi4py-initialization-breaks-fortran.
The short version is that this isn't an mpi4py issue either.  For my
particular case, mpi4py could possibly fix it by linking to the Fortran
bindings even though it doesn't need or use them, but I think that focusing
on mpi4py misses the larger point.

As the standalone test files he provided show, this issue can definitely
happen without mpi4py in the loop.  It would appear that using any 3rd
party Python/C library that uses MPI (or custom Python bindings to the MPI
C library) would cause later invocation of Fortran code to behave
incorrectly.  Having the Fortran accessible C wrappers testing the buffer
against the address of a variable defined in another library seems odd to
me, but I don't understand your codebase well enough to propose an
alternative.  I would say that even if this is a problem that the end user
needs to fix, it would be nice if there were some user-facing documentation
to explain and provide a remedy for the problem.

-Patrick

On Wed, Jul 1, 2020 at 8:40 PM Jeff Hammond via discuss <discuss at mpich.org>
wrote:

> Has Lisandro provided any feedback on this anywhere?  I agree with Hui.
> This appears to be an artifact of the way MPI uses magic values for
> MPI_IN_PLACE and how those get encoded in binaries.  I naively speculate
> that MPICH could encode the magic values differently to work around this,
> but that would likely break ABI compatibility, at least on the Fortran side.
>
> Jeff
>
> On Wed, Jun 10, 2020 at 3:31 PM Zhou, Hui via discuss <discuss at mpich.org>
> wrote:
>
>> I am not sure this is an issue for mpich. Seems more to be an issue of
>> `mpi4py`. The issue is exactly as you suspected -- `libmpifort.so` need be
>> loaded before `libmpi.so`, or the external symbol in `libmpifort.so` won’t
>> get resolved.
>>
>>
>>
>> >  It works if you load something linked to libmpifort first or load
>> everything with RTLD_GLOBAL.
>>
>> Seems plausible. I don’t have any better idea other than fixing `mpi4py`
>> so it always load `libmpifort.so`  first.
>>
>>
>>
>> --
>> Hui Zhou
>>
>>
>>
>>
>>
>> *From: *Patrick McNally via discuss <discuss at mpich.org>
>> *Reply-To: *"discuss at mpich.org" <discuss at mpich.org>
>> *Date: *Wednesday, June 10, 2020 at 1:09 PM
>> *To: *"discuss at mpich.org" <discuss at mpich.org>
>> *Cc: *Patrick McNally <rpmcnally at gmail.com>
>> *Subject: *Re: [mpich-discuss] Using MPICH in Python breaks Fortran
>> MPI_IN_PLACE
>>
>>
>>
>> I hate to bug, but this is a pretty serious issue.  I suspect it is why
>> we get segfaults trying to use similar variables like MPI_STATUSES_IGNORE.
>> Any insight would be appreciated.
>>
>>
>>
>> Thanks,
>>
>> Patrick
>>
>>
>>
>> On Wed, May 27, 2020 at 10:25 AM Patrick McNally <rpmcnally at gmail.com>
>> wrote:
>>
>> Our application consists primarily of a Python head calling into Fortran
>> routines to do the heavy lifting.  We have never been able to successfully
>> use MPI_IN_PLACE in Fortran but weren't sure why.  Recently, we discovered
>> that it works fine in standalone Fortran code and is only broken when the
>> Fortran code is run through our Python modules.
>>
>> The issue appears to be related to having code that only links to the C
>> libmpi library loaded first and with RTLD_LOCAL, as happens when we load
>> mpi4py.  It works if you load something linked to libmpifort first or load
>> everything with RTLD_GLOBAL.  I'm assuming this has something to do with
>> how MPICH tests the address of MPIR_F08_MPI_IN_PLACE but I don't understand
>> SO loading well enough to fully grasp the issue.  Below is some standalone
>> code to show the issue.  I'd appreciate any insight you can provide into
>> why this is happening.
>>
>> Relevant system details:
>> RHEL 7.8
>> Python 2.7
>> GCC 7.3.0
>> MPICH 3.3.2 (and 3.2)
>>
>> The below files are also available towards the end of the bug report at
>> the following link:
>>
>> https://bitbucket.org/mpi4py/mpi4py/issues/162/mpi4py-initialization-breaks-fortran
>>
>>
>>
>> Thanks,
>>
>> Patrick
>>
>>
>> makefile
>> -----------
>> libs = testc.so testf.so
>> all: $(libs)
>>
>> testc.so: testc.c
>>         mpicc   -shared -fPIC $< -o $@
>>
>> testf.so: testf.f90
>>         mpifort -shared -fPIC $< -o $@
>>
>> clean:
>>         $(RM) $(libs)
>>
>> testc.c
>> ---------
>> #include <stddef.h>
>> #include <stdio.h>
>> #include <mpi.h>
>>
>> extern void initc(void);
>> extern void testc(void);
>>
>> void initc(void)
>> {
>>   MPI_Init(NULL,NULL);
>> }
>>
>> void testc(void)
>> {
>>   int val = 1;
>>   MPI_Allreduce(MPI_IN_PLACE, &val, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
>>   printf("C val: %2d\n",val);
>> }
>>
>> testf.f90
>> -----------
>> subroutine initf() bind(C)
>>   use mpi
>>   integer ierr
>>   call MPI_Init(ierr)
>> end subroutine initf
>>
>> subroutine testf() bind(C)
>>   use mpi
>>   integer ierr
>>   integer val
>>   val = 1
>>   call MPI_Allreduce(MPI_IN_PLACE, val, 1, MPI_INTEGER, MPI_SUM,
>> MPI_COMM_WORLD, ierr)
>>   print '(A,I2)', 'F val: ', val
>> end subroutine testf
>>
>> test.py
>> ---------
>> from ctypes import CDLL, RTLD_LOCAL, RTLD_GLOBAL
>>
>> mode = RTLD_LOCAL
>> cfirst = True
>>
>> if cfirst: # it does not work!
>>     libc = CDLL("./testc.so", mode)
>>     libf = CDLL("./testf.so", mode)
>> else: # it works!
>>     libf = CDLL("./testf.so", mode)
>>     libc = CDLL("./testc.so", mode)
>>
>> libc.initc.restype  = None
>> libc.testc.argtypes = []
>> libf.initf.restype  = None
>> libf.testf.argtypes = []
>>
>> libc.initc()
>> libc.testc()
>> libf.testf()
>>
>> _______________________________________________
>> discuss mailing list     discuss at mpich.org
>> To manage subscription options or unsubscribe:
>> https://lists.mpich.org/mailman/listinfo/discuss
>>
>
>
> --
> Jeff Hammond
> jeff.science at gmail.com
> http://jeffhammond.github.io/
> _______________________________________________
> discuss mailing list     discuss at mpich.org
> To manage subscription options or unsubscribe:
> https://lists.mpich.org/mailman/listinfo/discuss
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpich.org/pipermail/discuss/attachments/20200702/3607025a/attachment-0001.html>