The trust boundaries will certainly be important. If a tenant is
allowed to load their own OS on bare metal, it will probably be outside
the trust zone of the provider's overlay network.

1. Consider a case of an enterprise where it has full control over
   what is loaded onto a server - in this case regardless of whether
   a hypervisor is on the bare-metal or a non-virtualized OS, it
   makes sense to have the NVE in the OS

2. Similarly a cloud sp may use non-virtualized servers to offer
   certain services such as storage, hadoop, queuing etc. It makes
   sense to have the NVE in the OS

3. Lastly a case where a service provide lets a tenant load their
   own OS onto the server - i.e. some sort of managed hosting. In
   that case the servers dedicated to the tenant will be in
   a separate trust zone. They will most likely be on a different
   VLAN altogether - they may even be physically located in
   a different rack in a data center.

   In this case you are looking at more than beyond simple encap/decap
   to provide connectivity between the two trust zones.

regarding the point about using untagged interfaces today, I think
having two different sets of admins - server admins and network
admins - and who owns setting  up VLANs - may have a lot to do with
that as well. Which set of admins will own encap/decap? I think the roles will 
change.

Somesh

> -----Original Message-----
> From: Ivan Pepelnjak [mailto:[email protected]]
> Sent: Tuesday, August 28, 2012 9:51 AM
> To: Stiliadis, Dimitrios (Dimitri)
> Cc: Somesh Gupta; Black, David; [email protected]; Linda Dunbar
> Subject: Re: [nvo3] Let's refocus on real world
> 
> Exactly right. There are good reasons we're using VLANs and untagged
> server interfaces today.
> 
> I wouldn't trust my servers to choose which virtual network they want
> to
> participate in, let alone my customers' servers.
> 
> Ivan
> 
> On 8/28/12 5:13 PM, Stiliadis, Dimitrios (Dimitri) wrote:
> [...]
> 
> >>   This is certainly only today's restriction. If nov3 takes off,
> there
> >>   certainly could be a pseudo-driver in Linux that could implement
> the
> >>   NVE (like a VLAN driver) without much additional overhead.
> >
> >   That doesn't work if you assume that tenants and DC operators are
> > different
> >   entities. The DC operator cannot rely on the tenant to do the right
> >   encapsulation. Different administrative and trust domains. That's
> why
> >   in my original email I was talking about "trust boundaries".
> >
> > Dimitri
> >>
> >
_______________________________________________
nvo3 mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/nvo3

Reply via email to