Advice for 2nd cluster installation
waitt at saic.com
Fri Jan 10 12:45:25 PST 2003
Robert G. Brown wrote:
> A lot of folks who do moderate to fine grained computations own their
> own dedicated, specifically engineered systems and don't necessarily
> need a batch queue.
Hmm, well maybe I've got "tighly coupled MPI" tunnel vision, but
even single user platforms can benefit from using resource managers.
It is very convenient to submit multiple mpi jobs and not have to
worry about whether they are going to overlap. And can you
imagine having to figure out what nodes/processors are not being
used, then having to specify those nodes when you run the job? Yikes.
Now think big. ASCI big.
I tend to think of batch queues as
> "mostly" being used by folks for whom resource allocation on a
> one-at-a-time-basis is fine, not "N out of M at a time, starting all N
> at this specific time and holding them until that".
I would agrue the opposite. Can you say "Advanced Reservations"
and "Preemptive scheduling?" Mmmmmm, preemptive scheduling. ;)
It would make my life a living hell trying to maintain
clusters with multiple users without resource managers.
And not really because of all the wild and wonderful
scheduling capabilities these days, but simple things
like having the RM run epilogue scripts to kill all
the running procs that the nasty users leave behind when
their jobs die, for example.
Tim Wait waitt at saic.com
SAIC - Advanced Systems Group
PO Box 41, Sumerduck VA 22742
More information about the Beowulf