Re: Is mpi SYNC ALL supposed to work?

[Tobias Burnus <burnus at net-b dot de>]

Hello,

first, I want to point out that the MPI version of gfortran's coarray 
support is NOT yet WORKING. (Except in very special but rather useless 
cases.) I think I really should tone down the release notes a bit more 
to make this clearer.

Regarding your question:

On 01/13/2012 04:50 PM, VladimÃr Fuka wrote:
> I tried the mpi version of coarrays (library from today's trunk and gfotran
> from trunk from 20. December). I can not get an a simple integral with
> reduction at the end to work, because the SYNC ALL does not seem to work for
> me.

I am not surprised that the *reduction* does not work as stated at 
several places: There is NO communication between images is supported at 
the moment.(*)


However, the SYNC ALL itself should work by calling MPI_Barrier. It also 
seems to work for me with:

print *, "BEFORE on ", this_image()
if (mod(this_image(), 2) == 0) call sleep(1)
SYNC ALL
print *, "AFTER on ", this_image()
end

$ mpicc -c libgfortran/caf/mpi.c
$ mpif90 -fcoarray=lib mpi.o foo.f90
$ mpirun -n 5 ./a.out
 BEFORE on            1
 BEFORE on            4
 BEFORE on            2
 BEFORE on            5
 BEFORE on            3
 AFTER on            3
 AFTER on            2
 AFTER on            1
 AFTER on            5
 AFTER on            4

While if I comment out the SYNC ALL, I get:

 BEFORE on            1
 AFTER on            1
 BEFORE on            2
 BEFORE on            5
 AFTER on            5
 BEFORE on            3
 AFTER on            3
 BEFORE on            4
 AFTER on            2
 AFTER on            4

Thus, it works for me and I do not see any reason why it shouldn't.

> and the print were one after another from different images, without any sign
> of performed barrier, ie. they were not ordered all "mpi bariera 1" first and
> then all "mpi bariera 2", but first 1 then 2 and so on.
> I do not understand it.

Me neither. It seems to be an MPI problem.


Regarding the Coarray status:


a) Coarrays for a single image (-fcoarray=single, but also -fcoarray=lib 
with a single image):

This mostly work correctly - the only issue are CLASS (polymorphic 
coarrays) and some issues with derived types with coarray components


b) Multi-images (-fcoarray=lib with MPI version of libcaf):

Only the very basic support is there: Registering/allocating coarrays, 
deregistering/deallocating coarrays, argument passing, this_image(), 
ucobound etc.

Missing are:
(1) Proper request handling in the libcaf_mpi library, i.e. some 'event' 
processing loop
(2) Implementation of SYNC IMAGES (and also sync all and error stop) 
with that method.
(3) Adding calls for the various data pull/push

Until (1)-(3) are at least partially implemented, no sensible coarray 
program is possible. As especially (3) is a larger job, it will take at 
least until mid-2012.

For 4.7 only some progress on (b)(1) is possible as only minor changes 
are allowed until the GCC 4.7.0 release, which could happen end of March.

For 4.8 the plan is to work on all those items, but it is completely 
unclear how far we will get. Especially as there are also other items 
like polymorphism are on the agenda.

Tobias

PS: Note that gfortran is developed by volunteers. The progress depends 
on the time the developers have - and whether new developers are 
joining. If you are interested ...