As of r323, the default action is to run mtt_ompi_cleanup.pl with &test_np(), meaning that we could be running multiple copies on each node. This is not ''harmful'' -- it creates some harmless race conditions (i.e., which copy exactly will kill the rogue orteds), but it is at least somewhat wasteful. It would be better to somehow figure out how to run only 1 copy of the script per node (a la LAM's "N" scheduling).
Perhaps if OMPI supports LAM's C/N scheduling someday, this would be much easier... :-)
As of r323, the default action is to run mtt_ompi_cleanup.pl with &test_np(), meaning that we could be running multiple copies on each node. This is not ''harmful'' -- it creates some harmless race conditions (i.e., which copy exactly will kill the rogue orteds), but it is at least somewhat wasteful. It would be better to somehow figure out how to run only 1 copy of the script per node (a la LAM's "N" scheduling).
Perhaps if OMPI supports LAM's C/N scheduling someday, this would be much easier... :-)