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

[Pshem Kowalczyk]

Hi,

On 21 July 2011 03:34, Peter Rathlev <peter at rathlev.dk> wrote:
> We've started acting like grown-ups and have (almost) implemented
> seperate route-reflectors. The old setup was a mix of full mesh (6
> devices) and RR clients (a few dozens) hanging off 3 of the full mesh
> devices.
>
> I'm now contemplating what the "right" number of RRs are. We have three
> main "datacenters" and a dozen PoPs. So far we were planning on having
> two RRs, but I'm wondering if three isn't a better number.

That's a setup very similar to ours. Currently we run two standalone RRs.

> Can anyone think of a reason to not have three RRs? Processing power
> required would still be less than a full mesh configuration, so that's
> somewhat irrelevant.
>
> I have heard a vague comment about an odd numbers of RRs being a
> problem, but I can't figure out why that would be.

I can't see any reasons why 3 RRs would be any different to 2 RRs (I
assume that you'd peer all of your PEs to all 3 RRs), except for the
fact that each PE has to process the same prefix from 3 different
sources. If you have VRFs with singnificant number of prefixes and
relatively underpowered PEs somewhere - that might slow convergence a
bit and potentially add to memory consumption.


kind regards
Pshem