job steps on front-end system (e.g. Blue Gene).

means the tasks are already gone.

 don't seg fault if data is missing, free the data when the agent is
 finished with it.

This will gracefully recover from an assortment of slurmctld failure
modes.

batch job submission. That srun should be long gone when the allocation
actually occurs.

prolog and epilog. Needed to synchronize slurmctld with user job execution:
Don't start script until partition is ready and clear all jobs upon
termination (offloading work from slurmctld).

they fail to execute

select/bluegene can resolve BGL API symbols.

RM_MESH to avoid multiple definitions for these values (one in slurm/slurm.h
and the other in bgl/rm_api.h.

 Suse Linux.

All work from <takao.hatazaki@hp.com>.

nodes in the allocation.

 where the requeue logic was in a tight loop.

tests were change appropriately. This lets one fake a Linux cluster
without all of the BGL-specific logic.

and FastSchedule=0 and the cluster is heterogeneous.

 uninitialized variables (thanks to valgrind).

listed twice.