Hi,
I had a look at the ti_decomp_case. It hanged for me as well.
The hang is caused in sander.f, ca line 353:
if (ifsc > 0) then
partner = ieor(masterrank,1)
if (nat == natom) then
nrank = masterrank
else
nrank = partner
end if
call mpi_bcast(jgroup, nat, MPI_INTEGER, nrank, commmaster, ierr)
end if
I did not understand why nrank is computed this way. But in the test case
it comes out as 0 and 1 for the two processes, causing the following
mpi_bcast to hang, since both processes think they should talk and no one
listens (an all too common problem, not only in parallel programming...)
Replacing the nrank with a 0 in the mpi_bcast call fixes the problem (and
makes the preceeding five lines obsolete) and the test case runs through.
However, I do not know if this breaks decomp in other cases, since the
test has the same number of atoms on both prmtops. In a more typical SC
case with different atom numbers, this might work out different, but I
have nothing to compare...
Kind Regards,
Thomas
Dr. Thomas Steinbrecher
BioMaps Institute
Rutgers University
610 Taylor Rd.
Piscataway, NJ 08854
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Tue Apr 13 2010 - 09:00:08 PDT
