[c-nsp] Number of route reflectors, best practice?

[Peter Rathlev]

On Thu, 2011-07-21 at 22:22 +0200, Gert Doering wrote:
> "Just because everybody else does it" is a no-go in my book :-) - we
> currently have a design similar to your current design, that is, all
> "core" routers (8) are full-meshed, and all "edge" routers in a given
> POP use the "core" as RRs.  Edges have only edge-routes plus default,
> so the computational effort on the RRs is not that bad.  And we don't
> need extra boxes...
> 
> If both cores fail in a POP, that POP is down anyway, and I don't need
> to worry about RR reachability either.

Those are compelling arguments, though I'm not sure we're big enough for
that. The network is physically partial mesh, and the dozen PoPs aren't
geographically hierarchical in relation to the main datacenters.

Furthermore, we currently have a very collapsed design where the current
RRs are also terminating "customer" (internal) networks. This has worked
well for us, but it does present some interesting problems regarding
scale and potential DoS.

(Oh, and we were advised by expensive consultants to introduce seperate
RRs. And buy Nexus.)

-- 
Peter