Hello,
I have just committed a change to use MPI_GATHERV instead of MPI_REDUCE
in the MERGE_FORCES subroutine of debug.f. For parallel runs the debug
facility was failing in two different cases.
init /= 3
---------
Here all processes have the forces for all atoms. MPI_REDUCE sums all
of the forces from all of the processes to give the force on the master
process as sandersize*frc.
init == 3
---------
Here each process has the correct forces for its own subset of atoms but
unspecified values for the remaining atoms. MPI_REDUCE adds the
correct forces with the unspecified values, giving incorrect forces on
the master node.
Using MPI_GATHERV fixes both of these cases. If this change has broken
anything, please let me know.
Tyler
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Mon Jul 06 2009 - 10:10:49 PDT
