Re: [nvo3] Requirements + some non-requirement suggestions

[David Meyer <[e-mail]>]

On Mon, Apr 9, 2012 at 1:51 PM,  <[e-mail]> wrote:
>> Dave McDyson asked about three classes of connectivity:
>>
>> 1) a single data center
>> 2) a set of data centers under control of one administrative domain
>> 3) multiple sets of data centers under control of multiple
>>       administrative domains
>>
>> Which of these do we *need* to address in NVO3?
>
> I agree with a number of other people that we have to start with 1), and then I'd
> suggest addressing as much of 2) as "makes sense" without significantly affecting
> the design and applicability of the result to data centers.  For example, a single
> instance of an overlay that spans data centers "makes sense", at least to me, or
> as Thomas Narten described it:
>
>> two DCs are under the same administrative domain, but are
>> interconnected by some (existing) VPN technology, of which we don't
>> care what the details are. The overlay just tunnels end-to-end and
>> couldn't care less about the existence of a VPN interconnecting parts
>> of the network.
>
> Looking at draft-meyer-loc-id-implications-01:
>
>        http://tools.ietf.org/html/draft-meyer-loc-id-implications-01
>
> I would suggest that for initial nvo3 work, reachability between all NVEs in a
> single overlay instance should be assumed, as there will be an IGP routing protocol
> (e.g., OSPF with ECMP) running on the underlying data center network which will
> handle link failures.

That may be a reasonable assumption, but the fact that an IGP is
running doesn't ease the "locator liveness" problem unless the routing
system is carrying /32s (or /128s) and the corresponding /32 (/128)
isn't injected into the IGP unless the decapsulator is up (that in and
of itself might not be sufficient as we've learned from our
experiences with anycast DNS overlays). In any event what we would be
doing in this case is using the routing system to signal a live path
to the decapsulator. Of course, carrying such long prefixes has its
own set of problems.

> Specifically, for initial nvo3 work I'd suggest assuming
> that the underlying network handles reachability of the NVE at the other side of
> the overlay (other end of the tunnel) that does the decapsulation.  In terms
> of that draft, within a single data center (and hence for the scope of initial
> nvo3 work), I'd suggest that the underlying network be responsible for handling
> the Locator Path Liveness problem.

Not sure what you mean here by underlying network. In the LISP case,
does the underlying network handle this problem? In any event can you
be a bit more explicit in what you mean here?
>
> This suggestion also applies to the Multi-Exit problem, although on a related
> note, I think it's a good idea to make sure that nvo3 doesn't turn any crucial
> NVE into a single point of failure. Techniques like VRRP address this in the
> absence of nvo3, so this could be mostly a matter of paying attention to ensure
> that they're applicable to NVE failure.  Regardless of whether things work out
> that way, I'd suggest that availability concerns be in scope.
>
> Turning to the topic of IGP metrics and "tromboning", I'd suggest that having
> nvo3 add a full IGP routing protocol (or even an IGP metrics infrastructure for
> one that has to be administered) beyond what's already running in the underlying
> network is not a good idea.  It seems like a large portion of the "tromboning"
> concerns could be resolved by techniques that distribute the default gateway in
> a virtual network so that moving a VM (virtual machine) automatically sends
> traffic to the locally-applicable instance of the default gateway for the VM's
> new location, based on the same L2 address. There are multiple examples of this
> sort of approach - OTV and draft-raggarwa-data-center-mobility-02 are among them.
>
> One more - Peter Ashwood-Smith writes:
>
>> Is it a requirement to support different tunnel encapsulation types in the same DC?
>>
>> It would seem that a very large DC could well end up with several different kinds of
>> tunnel encapsulations that would need to somehow be bridged if they terminate VMs in
>> the same subnet.
>
> I'd suggest that the latter scenario be out of scope and that crossing virtual
> networks initially involve routing in preference to bridging, so that an NVE
> receiving an unencapsulated packet can determine the overlay and encapsulation by
> knowing which virtual network the packet belongs to.  An implication is that I'd
> suggest figuring out how to optimize the following structure into a single
> network node later (or at least as a cleanly separable work effort:
>
> ... (Overlay1)---- NVE1 ---- (VLANs) ---- NVE2 ---- (Overlay2) ...
>
> In the above, NVE1 and NVE2 are separate nodes, and the parenthesized terms are
> the means of virtual network separation.
>
> That suggests that a starting point for whether different tunnel encapsulation types
> should be supported in a single data center could be "if they don't have an NVE
> node in common, they can be made to work" and optimizations can be considered later.

Agree with this latter statement.

Dave