lmb at suse
Jul 12, 2011, 4:38 AM
Post #2 of 2
On 2011-07-11T13:57:26, Florian Haas <florian.haas [at] linbit> wrote:
Re: VirtualDomain patch for utilization (was Re: Two git patchs for resource-agents)
[In reply to]
responding to Florian since he already split up the submission into two
distinct parts, which makes a lot of sense. (John, please also remember
that FATE references are meaningless to non-SUSE folks; a bit more
context is usually helpful. Thanks!)
> > 1. Add a functionality that modifies utilization of resource
> > automatically, please see fate 310117.
> > I added 3 ocf parameters:
> > - dynamic_utilization, if set, the utilization parameter of resource
> > will be reset if there are
> > difference between resource parameters and system parameters when agent
> > monitor.
Can this happen with attrd? That would allow us to dampen the
fluctuation of, say, the CPU requirements as the load changes.
> > Otherwise, the resource parameters will be set once when agent start.
> > - set_utilization_cpu, enable setting cpu of utilization.
> > - set_utilization_memory, enable setting memory of utilization.
And I fear "start" is the wrong place to set this; at that time, it's
basically already too late.
The probe (monitor_0) may be the better time to reset these parameters.
Downside is that it would be called on several nodes; but that could be
avoided by "am I running on the lowest nodeid" or something.
More problematic is that the fs that holds the configuration may not be
mounted yet. Hrm.
If we call it in start, and - on noticing that the utilization settings
were initialized or modified - don't go ahead starting the guest,
instead bail out with OCF_NOT_RUNNING, so that the PE gets a chance to
Alas, not perfect. Do we need a pre-start notification? A "prepare"
command? Or is the former good enough?
> Assuming everything works, I'd still ask Andrew to weigh in on this
> before we merge this. The patch highlights an important issue for which
> we might want a more generic resolution: a resource agent being able to
> report "the resource is currently using this much memory, that much CPU,
> that much network traffic, that much disk space", etc. Pacemaker could
> then use this dynamic information for resource placement, the way it
> already does statically.
Well, this is similar to crm_master (except that there we have the neat
distinction that we _can_ reasonably call it in start, since promote is
a separate operation).
Architect Storage/HA, OPS Engineering, Novell, Inc.
SUSE LINUX Products GmbH, GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer, HRB 21284 (AG Nürnberg)
"Experience is the name everyone gives to their mistakes." -- Oscar Wilde
ha-wg-technical mailing list
ha-wg-technical [at] lists