hirlam

Hi all,

Just to share some things from using Torque batch scheduler (might be not the most popular at the moment anyway)

Under Sgi mpt-1.26 (not sure about other releases, but if similar as below...);
within mpiexec_mpt wrapper:
having MPI_BATCH_CMD="pbs_track -j $PBS_JOBID"


using adittional options with the executable in the manner they are used in Harmonie Forecast script say:

"mpiexec_mpt -np 128" somedir/MASTERODB -maladin -v$VERSION -e$CNMEXP -c$NCONF -t$TSTEP -fh$LL -a$ADVEC || exit

Might result in:

.............
MPI: r1i0n1: 0x7faa00004fa1392a: pbs_track: invalid option -- m
MPI: r1i0n1: 0x7faa00004fa1392a: pbs_track: invalid option -- a
MPI: r1i0n1: 0x7faa00004fa1392a: pbs_track: invalid option -- l
MPI: r1i0n1: 0x7faa00004fa1392a: pbs_track: invalid option -- a
MPI: r1i0n1: 0x7faa00004fa1392a: pbs_track: invalid option -- d
MPI: r1i0n1: 0x7faa00004fa1392a: pbs_track: invalid option -- i
........ etc.
MPI: r1i0n1: 0x7faa00004fa1392a: USAGE: pbs_track [-j <JOBID>] [-b] a.out arg1 arg2 ... argN
MPI: could not run executable (case #4)
/usr/bin/mpiexec_mpt: line 52: 4941 Killed $mpicmdline_prefix -f $paramfile

Apart from solving this by fighting mpiexec_mpt or using other possible options, one might find it useful to:
If your using source of Torque release (still saw that in in all 4.0 beta being the latest, though tested on an older version) for installation or building rpm's when in:
./src/cmds/pbs_track.c :

changing (if applicable):
#define GETOPT_ARGS "bj:"
to
#define GETOPT_ARGS "+bj:"

solves the problem of pbs_track not accepting additional options supplied togeter with the executable.

I however recommend testing that all of your submission stuff running on torque is ok after applying this change.