Open Stack

On 11/14/2012 12:11 PM, Julien Danjou wrote:
> On Wed, Nov 14 2012, Russell Bryant wrote:
> 
>> Can someone do a bit of a problem statement for someone like me
>> that isn't completely up to speed on what's going on in
>> Ceilometer in this area?
> 
> Basically, we want to retrieve informations about instances, like
> the CPU time used by each of them, or their network interfaces
> traffic.
> 
> Currently, we have an agent (ceilometer-compute-agent) running on 
> nova-compute nodes, and polling libvirt using nova internal code
> we import. This has a lot of downside, like depending on nova
> internal code which can change, having an agent running on each
> nova-compute node, and forcing us to maintain per-hypervisor
> support (at least for now).
> 
> We'd like to move this to an RPC approach, so nova-compute would
> offer information like CPU time used by an instance over RPC. This
> would avoid running an agent on the compute node, avoid using nova
> internal API code, and avoid maintaining per-hypervisor logic into
> Ceilometer (nova would be in charge of that).

I see the benefit it brings in cutting down on hypervisor specific
code.  However, at least in my opinion, you're still relying on
internal nova code by using the rpc API directly.

Is this something we could implement as an extension to the compute
REST API, instead?

-- 
Russell Bryant