Lou Berger <[email protected]> wrote:
> Martin,
> 
> On 09/02/2015 06:42 AM, Martin Bjorklund wrote:
> > Andy Bierman <[email protected]> wrote:
> >> On Fri, Aug 28, 2015 at 1:31 PM, Lou Berger <[email protected]> wrote:
> >>> Can one of you give an example of how this word work for a device (which
> >>> may be physical or virtual) that allocates done resources, say interfaces
> >>> to one logical entity (router, system, etc) and other resources to a 
> >>> second
> >>> entity? And of course I want to manage all with yang and the first and
> >>> second (sub) entity must be completely independent and ignorant of each
> >>> other.
> > 
> > [...]
> > 
> >> The logical system knows only about itself:
> >>
> >>     /interfaces
> >>     /system
> > 
> > This is important.
> > 
> >> The / node is represented by <config> or <data> or <filter> in the 
> >> protocol.
> >>
> >>    <get-config>
> >>        <source><running/></source>
> >>        <filter>
> >>          <interfaces />
> >>          <system />
> >>       </filter>
> >>    </getconfig>
> >>
> >> Each logical system can have its own "eth0" interface or whatever.
> >> They are mapped to real interfaces in the physical system.
> >>
> >> All operations on the logical system are validated against its own
> >> virtual datastore.  YANG validation does not work on individual array
> >> slices -- it only applies to an entire datastore.
> > 
> > Yes.
> > 
> >> On the physical server there needs to be a data model to manage the
> >> logical servers (as Martin suggested).
> >>
> >>  <config>             <--- root on PHY server
> >>    <interfaces  />   <--------------- contains the real interfaces,
> >> including eth23
> >>    <virtual-servers>
> >>       <virtual-server>
> >>          <name>vs1</name>
> >>          <itf-map>
> >>              <real-itf>eth23</real-itf>
> >>              <vir-itf>eth0</vir-itf>
> >>          <itf-map>
> >>          <more-virtual-server-params ... />
> >>          <root>                   <----------- YANG mount point (virtual
> >> server root)
> >>             <interfaces>
> >>                <interface>
> >>                   <name>eth0</name>
> >>                     ...
> >>                </interface>
> >>             </interfaces>
> >>             <system ... />
> >>          </root>
> >>        </virtual-server>
> >>     </virtual-servers>
> >>   </config>
> > 
> > I like this, but I would actually not use mount here.  I don't think
> > it is necessary.  This would be a model for devices that support
> > multiple 'virtual-servers' / 'logical-network-elements'.  So in this
> > model you configure these logical-network-elements and allocate
> > resources like interfaces etc to them.  For true virtual servers,
> > you'd also configure the NETCONF server and authentication params,
> > meaning that each such virtual server has its own config, which is
> > completely separate from the others.  In this architecture, it would
> > not be correct to mount all the models in the virtual server list.
> > 
> 
> We discussed this in the DT and (I think) agreed there is room / need
> for both approaches based on device owner/client management model (i.e.,
> is the device owner responsible for client config, device owner without
> client view.)  Do you have a reference for a model that can be used to
> support this, or just thinking one is needed?

I don't know if it is possible to find agreement among hypervisor
implementations etc to find a common model.  But from a YANG language
perspective I think it is pretty straightforward.  Since each virtual
server is truly virtual, there is no need for the mount functionality
in such a model.

As for the other model with logical systems within a real system,
where each logical system tries to present its own 'system view' of
the config and state, even though in reality there is just one config,
I think a mechanism like 'mount' would be useful.  If it was done with
'mount' instead of the proposed model in
draft-rtgyangdt-rtgwg-device-model-00, it doesn't cost anything for
the 99% (more?) of all systems that do not have this kind of logical
systems, and data models would not have to augment the
/device/logical-network-elements/logical-network-element path.


/martin

_______________________________________________
netmod mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/netmod

Reply via email to