[torqueusers] Re: Epilogue script
garrick at clusterresources.com
Tue Aug 29 12:04:07 MDT 2006
On Tue, Aug 29, 2006 at 07:38:35PM +0200, Ronny T. Lampert alleged:
> > Thus torque should be independent of the mpiexec implementation and
> > if there is a way to improve torque's behaviour with respect to
> > cleaning up after MPI jobs along the lines Dave indicated this would
> > be a huge benefit. The sites that do not want to use that feature
> > do not have to.
> Well - isn't that what "process groups" were initially thought up for?
> If pbs_mom starts a new process group for every job it should be safe to
> terminate any childs in that group (as long as no one calls setsid() which
> will start a new process group for the child).
> Looking at getpgrp(2) and setsid(2) - or am I missing something?
> The SGi-proposed solution looks nice, but I dislike such a "basic" feature
> being dependant on some special kernel module.
> As I am not entirely sure about PAM - is there any possibility to start
> something like a "hard PAM session group" where no process can escape?
> That looks like the perfect solution.
I don'tknow what the SGI thing is yet, but I'll look this afternoon.
I was looking at PAM awhile ago and couldn't come up with a reliably add
the jobid to the process' env.
More information about the torqueusers