openstack team mailing list archive

[Doug Hellmann <doug.hellmann@xxxxxxxxxxxxx>]

On Mon, Nov 5, 2012 at 3:32 AM, Zehnder Toni (zehndton) <
zehndton@xxxxxxxxxxxxxxxx> wrote:

> > On Thu, Nov 01 2012, Julien Danjou wrote:
> >> On Thu, Nov 01 2012, Zehnder Toni (zehndton) wrote:
>
> >> My goal is to offer monitored data to the admin and customers. The
> >> admin is interested in the utilization of the physical components and
> >> the virtual machines and the customer is interested to know what his
> VMs do or can do.
> >> It would be nice to get the data from a single point. I thought I can
> >> enhance the Ceilometer compute agent to get this data out. Does this
> >> make sense or is it better to use another monitoring tool for the
> >> physical components?
>
> > I think the pollster implementation can be done. I wouldn't implement
> this in the compute agent, but probably in some hardware agent, because
> it's likely it would be used in different kinds of environment and not only
> on compute node, i.e. you may also want to meter hardware usage for you
> cinder or glance node anyway.
>
> I think also the best way to implement this is to integrate a new
> (hardware) agent. Then we have a clear delineation. I'm very interested in
> helping to develop this.
>

If we make the current compute agent take an option telling it which
pollster namespace to use, then the same framework can load different
pollsters. However, there is a fundamental security issue with
communicating from an agent running inside a tenant's OS image using the
RPC stack. At DreamHost, and I suspect at other providers, that RPC network
is completely isolated from any tenant networks. We would not want a tenant
to be able to listen to the message bus, and definitely would not want it
to be able to write anything to the message bus.

We will need an API endpoint to receive data from agents running inside a
tenant image, examine it, and republish it on the message bus for other
consumers. Heat has something similar, although I think they just store the
data immediately.

Doug


>
> Toni
>
> > About the 10 minutes polling interval Doug mentionned, this can be a
> problem indeed, but it's still solvable later and would be easy to solve if
> this in a different agent, since you could change the periodic interval for
> pollster runs to something like 1 or 5 minutes.
>
> _______________________________________________
> Mailing list: https://launchpad.net/~openstack
> Post to     : openstack@xxxxxxxxxxxxxxxxxxx
> Unsubscribe : https://launchpad.net/~openstack
> More help   : https://help.launchpad.net/ListHelp
>