[c-nsp] downlink bgp interconnect best practices

[Nikolay Shopik]

On 31/05/11 16:57, Gert Doering wrote:
> In your case, I'd have the core/border routers do route-reflector functions
> towards the customer-edge routers.  Saves you having to do a full mesh
> between all the customer-edge routers, and saves you from having to add
> two more boxes (two! one RR is going to fail one day, and then your BGP
> is broken).

If you mean saving from installing two additional routers just for RR, 
then I agree core/border should be RR towards customer-edge routers.

>> btw in your scheme there is link between CR1 and CR2, what's point to
>> have such, if my customer is have bgp session with both there is no
>> chance packet could go to "worst path" and thus link between CR1 and CR2
>> redundant. Probably I miss some case scenario?
>
> Well, it depends a bit how the connectivity between CR1 and CR2 is
> built.  If you have two independent switches there, the direct link is
> not strictly needed.
AR1 and AR2 two independent switches at different racks. If I understood 
you correctly. So to replicate that scenario both physical links from 
AR1 and AR2 to CR2 should fail.

> Still, it has the advantage that if these switches
> should fail, CR1 and CR2 always are connected - otherwise you'll run
> into some weird problems.
>
> Imagine:
>
> upstream        upstream
>     |             |
>    CR1           CR2
>     | \
>     |  \
>     |   \
>    AR1  AR2
>
> now what happens if a packet comes in from upstream 2 to CR2?  It will
> have to be dropped ->  black-holing.  So if you have a direct link CR1-CR2,
> you ensure that your network never partitions, even if switch(es) fail.

Perfect explanation, thanks for that case scenario.