[torqueusers] Green computing - Power on/off
Jeff Anderson-Lee
jonah at eecs.berkeley.edu
Tue Aug 10 12:04:59 MDT 2010
On 8/10/2010 2:05 AM, rishi pathak wrote:
> Power aware scheduling should not restricted to node power on and
> off(costly affair), it should also facilitate for fine grained power
> control at node leve(CPU throttling, power to memory banks). Although
> we have not evaluated MOAB for green computing, but by looking at
> documentation, it seems like moab does not still provide for it. Also
> it would be better if previous usage patterns are considered and not
> just the idle time for a node.
Some studies have shown that the biggest/easiest win in the power game
is from shutting machines off, not slowing them down, as around 50% of
the power in many servers is non-CPU related. Shutting down power to
memory banks could be a possible savings, but I'm not aware of
motherboards that make that an easy option. Turning off disks can help
a little in some cases, for some loads, but given a 15W disk versus a
95W processor...
It may seem costly to turn machines on and off, but if your cluster
usage fluctuates broadly at different times of the day it *is* the *low
hanging fruit*.
Jeff Anderson-Lee
> On Mon, Aug 9, 2010 at 9:39 PM, David Beer
> <dbeer at adaptivecomputing.com <mailto:dbeer at adaptivecomputing.com>> wrote:
>
> I know that Moab has a way to do this, but that involves buying Moab.
>
> ----- Original Message -----
> > For the power saving purpose we want to call one script /
> program for
> > each job submission.
> > This script should run only on head node and perform job to turn on
> > extra nodes if free nodes are not sufficient to run a job.
> >
> > Another script/program will turn-off idle nodes .
> > How to do that ...Will appreciate if you share any idea for
> > implementation.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.supercluster.org/pipermail/torqueusers/attachments/20100810/e903eb12/attachment.html
More information about the torqueusers
mailing list