PBS/Scyld. Re: queue system on Scyld (Thomas Clausen)

[Rick Niles]

> As I understand it, it should be possible to have openPBS running on top
> of Scyld, and PBS can be configured (I think) to be as simple as a
> first come first served queueing discipline.
> 
> I wanted to do this, however one issue was unclear to me.  It appears
> that jobs are "brokered" on individual nodes by the pbs_mom program.
> However under scyld, they get started on the server and then migrated
> via bproc.
> 
> Does that imply that scyld systems get only one pbs_mom on the server?

The pbs_mom does that exact same function as the bpslave and sendstats
daemons that are already running on a Scyld Beowulf slave.  Thus, I believe
the "correct" setup would be to modify only the pbs_server so that it
uses BProc and have zero moms running.

That is our long term plan with OpenPBS.  However, that's going to
take some work with the internals of the PBS server.  So in the mean
time I've been trying to get it to just run with a mom on every slave
node.  (i.e. minimal changes) This is just to get a baseline on which
to improve from.

Also, please understand that PBS will ONLY accept shell scripts as
jobs and that the current Scyld Beowulf will not run shell scripts at
all.  So you've got big trouble right off.  Dan Ridge and I are
working up a method for running shell scripts on the nodes.  We still
think running scripts on dumb nodes it's really "not the right
thing".... but we understand that "the customer is always
right"...right?

> Does anyone have Quick-Install notes for PBS on top of Scyld?

Well, I think you'll see given that last paragraph... it will be
difficult.

BTW, what's this default home path of "/usr/spool/PBS"?  I think
"/var/spool/pbs" is much more sane.  Who's ever heard of "/usr/spool"?

	Rick Niles.
	Scyld Computing Corporation.