[Pacemaker] Making Virtual Domain utilization dynamic

[Michael Schwartzkopff]

Am Donnerstag, 2. Mai 2013, 16:56:19 schrieb Andrew Beekhof:
> On 02/05/2013, at 4:43 PM, Michael Schwartzkopff <misch at clusterbau.com> 
wrote:
> > Am Freitag, 26. April 2013, 22:14:59 schrieb Michael Schwartzkopff:
> > > Hi,
> > > 
> > > I picked up my old idea to make resouce utilizations dymanic and wrote
> > > the
> > > little patch for VirtualDomain to check CPU utilization every time the
> > > VirtualDomain is monitored:
> > > 
> > > --- /usr/lib/ocf/resource.d/heartbeat/VirtualDomain     2013-01-06
> > > 23:50:54.000000000 +0100
> > > +++ /usr/lib/ocf/resource.d/misch/VirtualDomain 2013-04-26
> > > 22:09:18.377784406 +0200
> > > @@ -404,6 +404,12 @@
> > > 
> > >             fi
> > >         
> > >         done
> > >      
> > >      fi
> > > 
> > > +
> > > +    if [ -x /usr/bin/virt-top ]; then
> > > +      CPU_TICKS=$(/usr/bin/virt-top -n 2 -d 5 --stream | grep
> > > $DOMAIN_NAME
> > > 
> > > | tail -1 | awk '{printf "%.0f",$7}')
> > > 
> > > +      /usr/sbin/crm_resource -r $OCF_RESOURCE_INSTANCE -z -p ticks -v
> > > $CPU_TICKS
> > > +    fi
> > > +
> > > 
> > >      return ${rc}
> > >  
> > >  }
> > > 
> > > I choose the measure CPU ticks since it is more granular than plain CPU
> > > usage.
> > > 
> > > With this patch pacemaker updates the utilization of the resource every
> > > time it monitors the resource. Together with the
> > > placement-strategy="minimal" and live-migration you could do very cool
> > > things.
> > > 
> > > Especially if you think about an extended fencing system that is able to
> > > switch off unused nodes and turn on node as needed.
> > > 
> > > Comments are welcome.
> > 
> > No comments?
> 
> Seems reasonable to me, though you'd probably want to have it controller by
> a new resource parameter.

Why? I think utilizations are the right thing to control dynamically. Nodes 
provide fix utilizations (i.e. CPUs and RAM) and resources consume it. The 
real demand is measured dynamically and the cluster manager can react on 
different load situations moving resources to "empty" nodes.

As a first test is used CPU_TICKS (i.e. 1/100 of seconds) with this unit you 
can tune the resource consumation more detailed.

The only problem: Can utilizations be changed using the hysteresis of 
attrd_update?

Greetings,

-- 
Dr. Michael Schwartzkopff
Guardinistr. 63
81375 München

Tel: (0163) 172 50 98
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.clusterlabs.org/pipermail/pacemaker/attachments/20130502/59f889ee/attachment-0003.html>