[c-nsp] Using MPLS PEs as gateways for access layer

[Ryan L]

Thanks Peter... sounds pretty straightforward, and mostly what I've got
labbed up/working.

Appreciate everyone's feedback on this - MPLS being somewhat new territory
for me, good to hear it seems folks have rolled similar designs out
successfully.

On Wed, Nov 30, 2016 at 2:12 PM, Peter Rathlev <peter at rathlev.dk> wrote:

> On Wed, 2016-11-30 at 11:38 -0500, Ryan L wrote:
> > One other question I have been having with this is, I'll obviously
> > have some L2 adjacency for my FHRP between the two cores, but should
> > I also be running an IGP within each DC between the two on the
> > private VRF? I don't have any other routed devices in these sites
> > except for the PEs, so I wouldn't be using it for anything aside from
> > between the two.
>
> With the "full MPLS L3VPN" model you would at most need just one
> adjacency in the global IGP between the two PE routers. There's
> generally no need to have an IGP running inside the VRFs and the
> redistribution between this IGP and BGP can be a mess.
>
> With VRF Lite I would on the other hand always make sure there's a
> direct physical link between the two PE routers and create an adjacency
> per VRF between them, though preferably in BGP.
>
> > Route information (connected, static) can be shared via VPN amongst
> > all PEs, but not sure if there is any benefit to sharing things like
> > static routes via IGP between the two PEs within a single site
> > (router ospf -> redist static subnets route-map etc.) as opposed to
> > propagating it to core #2 at the same site via VPN/BGP. In my labbing
> > it seems to work fine without IGP, but devil is always in the
> > (production) details. :)
>
> I'd just use MP-BGP for this. It doesn't matter if the PE routers have
> a direct iBGP adjacency or you use route reflectors, they would see
> each other as next-hop and route optimally according to the global IGP.
>
> I would avoid redistributing between an (in-VRF) IGP and BGP unless I
> really had to. It generally "just works" but can lead to hard to guess
> sub-optimal routing paths.
>
> --
> Peter
>
>