[c-nsp] bgp in the "core"

[Hroi Sigurdsson]

matthew zeier wrote:

> My network, roughly, is like this:
> 
> transit ---- transit --- transit
> 
>    core -- core
> 
> access switches, L2
> 
> The "core" provides L3 connectivity for customer networks and each transit 
> router is connected to each "core" router/switch.  The access switches are 
> plain L2 switches.  Core and transit run OSPF.

One thing to be aware of in the above setup (bgp-less core):
If you run iBGP between loopbacks in the transit layer, which is 
normally considered good practice, there is a possibility of 
loops/blackholes in the core during link-failure inside the transit 
layer. You need to ensure that inter-transitlayer traffic never escapes 
to the core. You can fix this using separate ospf areas (if that's how 
you carry loopbacks). I would also try to connect the transit routers in 
a triangle, if possible.

> The problem I have is that the transit routers are the boxes initiating my BGP 
> routes and I contend that if they become disconnected from the "core", having 
> them continue to announce routes is a Bad Thing.

As long as they can reach the core via another transit router it would 
seem to be OK to announce the networks.

> I believe that the core should do the route origination.

That would seem to be good practice. You can also redistribute static 
routes which point to the core routers. If the core next-hops become 
unreachable, the announcements will be withdrawn.