[c-nsp] BGP full-mesh vs. RR

[Pete Templin]

Ignacio Vazquez wrote:
> 
>  we actually are going to change from RR to Full Mesh. We have 75 routers
> (GSRs and 7500) with full-routing and two clusters. We will migrate to
> Full-Mesh because we need to have the optimal path to the routes and not
> loose information with the routers-reflectors.

How many POPs do you have?  How dense is your connectivity?

Planning for the future, we used the advice of AOL/ATDN when we rolled 
out RRs eight months ago.  All of our RRs are configured with 'no bgp 
client-to-client reflection', meaning the RR doesn't worry about 
reflecting routes amongst clients, only to/from non-clients.  Each POP 
is then fully meshed internally.  On each customer-attachment router, it 
has a full view of all local exit points along with a view of the best 
exit point at each remote POP (assuming that the remote POP has an exit 
of similar localpref and AS path length).  Works well for us, and 
positions us for (supposedly) stable MED behavior as we add parallel 
connections to external ASes.

Not the MOST optimal as you're aiming for, but I'd campaign to say that 
it's potentially optimal enough for most cases.

pt