[torqueusers] auto loadavg thresholds in MOM

Chris Samuel csamuel at vpac.org
Sun Nov 6 18:18:06 MST 2005


On Sat, 5 Nov 2005 06:53 pm, Garrick Staples wrote:

> Any thoughts on this general feature?

Sounds really interesting, I've never touched that feature before as I've 
always done it in the scheduler (thankfully our clusters are homogenous, at 
the moment) using this (for our dual IA-32 nodes):

NODELOADPOLICY          ADJUSTSTATE
NODEMAXLOAD             1.50

to flag a node as busy if something's gone wrong and a job is using more CPU 
than it should (yes, Materials Studio, I'm looking at you again).

>  Can anyone think of a use for 'c' in this context?  I'm not in any way tied
>  to this form. 

Yup, I'd like to be able to say that a node should be marked busy if the load 
is greater than (c-0.3), i.e. there is no longer really a whole CPU free for 
another job there.

That's not as good an estimate as you hit higher number of CPUs, but for our 
dual and quad boxes it should be "good enough".

Hmm, thinking a bit more broadly - and this is probably more for Maui/Moab 
than for Torque, it would be really handy to be able to say that if a node 
has N CPU's with A allocated and a load average of Y, then if (Y>A) set the 
number of free CPUs to (N-Y) rather than (N-A).

cheers!
Chris
-- 
 Christopher Samuel - (03)9925 4751 - VPAC Deputy Systems Manager
 Victorian Partnership for Advanced Computing http://www.vpac.org/
 Bldg 91, 110 Victoria Street, Carlton South, VIC 3053, Australia

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
Url : http://www.supercluster.org/pipermail/torqueusers/attachments/20051107/5f479355/attachment.bin


More information about the torqueusers mailing list