[mvapich-discuss] Hang in MPI_Win_fence

[Dorian Krause]

Hi List,

the attached program (bs-db.cc) uses a combination of onesided
communcication and derived datatypes to collect data from 2 origin
processes on 2 other target processes. The derived datatypes have been
checked to contain no overlap and the target window is large enough.

Unfortunately the program hangs in MPI_Win_fence after the access epoche
(MPI_Put). Two processes hang in MPIDI_CH3I_SMP_read_progress while two
others don't return from MPIDI_CH3I_SMP_writev.

I'm using mpich2-1.1a2 with the intel compiler suite. The OFED version
is 1.1 (old, I know ...). The configure command was

    $ ./configure --prefix=/home/dkrause/mvapich2-1.2p1-icc10
CFLAGS=-DOFED_VERSION_1_1 CC=icc CXX=icpc FC=ifort --enable-romio
--with-file-system=lustre


The program works correctly with mpich2-1.1a2 (it crashes with OpenMPI
though but I think this is a different issue). For my tests I ran 4
instances of the program on the cluster headnode.

The program is extracted from a real application which hangs IFF the
amount of transfered data is too large.

Any help/ideas would be appreciated ...

Thanks + Regards,
Dorian



-------------- next part --------------
A non-text attachment was scrubbed...
Name: mvapichtest.tar.bz2
Type: application/x-bzip
Size: 149632 bytes
Desc: not available
Url : http://mail.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20081229/3a5a09db/mvapichtest.tar-0001.bin