[Swift-devel] multiple worker.sh in the same job

[Ben Clifford]

On Tue, 1 Jul 2008, Mihael Hategan wrote:

> This whole thing, I think, applies not only to MPI jobs, but also to any
> job requesting more than one node. So I think the solution is not to
> swap mpirun and wrapper.sh, but, along the lines of what Andriy did,
> perform all the relevant wrapper functions in only one instance and have
> a barrier right before running the executable as well as right after.

> How exactly this would be done is a little hazy in my head, but I guess
> that's what makes it interesting.

Pretty much that's what putting mpirun as the job run by the wrapper 
script does, at least as far as PBS on TG-UC seems to behave:

mpirun exposes a single unix executable that before it runs, no app code 
has run, and when it finishes all app code has finished everywhere. The 
start and end of that executable are the barriers above.

PBS as far as I can tell, though it allocates a bunch of nodes, only 
spawns your job on one and informs you of the list of nodes that it has 
allocated to you through the $PBS_NODEFILE environment variable (that 
mpirun then uses, for example).

--