[c-nsp] ASR 1k vs 9k as a non-transit BGP router with full tables?

Łukasz Bromirski lukasz at bromirski.net
Thu Aug 3 17:58:55 EDT 2017



> On 3 Aug 2017, at 09:10, Patrick M. Hausen <hausen at punkt.de> wrote:
> 
> The use case is simply "full tables BGP" with currently 4x 1GB/s
> uplinks and possibly 6 in the near future. Upgrade to something
> 10G-ish not planned at the moment. 300-400 Mbit/s aggregate
> traffic across all uplinks currently.

Quite low aggregate bandwidth needs for 6500 :)

> So we are too memory heavy for the C6500 (SUP720-10G) and
> then there's the TCAM limitation ... although our bandwidth requirements
> are rather small. And then the C6500 definitely starts to rot - I wonder
> if I will ever get anything beyond 12.2(33)SXJ10 if (when!) the next
> remote security bug hits.

For that kind of scenario, Sup720-10GE can still do it’s job if
You use Selective Route Download. You don’t need full tables as
Spotify’s SIR project have shown. You’re even better than Spotify,
as You’re end station for the traffic, not transit as I understood.
Just take a look here (and read on):
https://labs.spotify.com/2016/01/26/sdn-internet-router-part-1/ <https://labs.spotify.com/2016/01/26/sdn-internet-router-part-1/>

You should live well with the same traffic engineering accuracy
with anywhere between 30k and 60-70k for IPv4 unicast put into
FIB and then into TCAMs. At the same time, You should be fine
(maybe with *some* filtering) to fit the 4-6 tables in RAM while
doing SRD (that will select specific routes and put only them as
candidates for FIB installation via BGP).

This however assumes, You want to play with CLI and possibly
NetFlow data to correlate telemetry.

Also, try to stick to 15.xS lines. It seems You’re doing quite simple things
and there’s no real value in staying on 12.2(33) line unless some
hardware dependencies.

BTW, you can upgrade RAM on 720-10GE to 2GB. This is of course not
officially supported, but as You’re anyway running on refubrished equipment,
you don’t care that much. Just remember to upgrade both RP and SP 
memory, as in theory with this Sup you wouldn’t need to care anymore
as SP is just a stub, but may actually play buffer allocation tricks
and if there’s disrepancy between RP and SP RAM size, you may
run into trouble (RP loosing SP, stalling and then rebooting on
watchdog - it isn’t pretty and for sure - not predictable).

— 
./


More information about the cisco-nsp mailing list