[c-nsp] BGP convergence in VRF vs. global routing table on 7600router

[Christian Bering]

>From: Oliver Boehmer (oboehmer) [mailto:oboehmer at cisco.com] 

Hi Oliver,

>Can you pls describe which "convergence" you are referring? Time it
>takes after bringing up the session until every other PE has 
>the routes?
>Or the table version on the CE and the receiving PE is the same?

Time it takes for PE1's CPU to fall back to normal utilisation and all
destinations reachable again after PE2 loses an upstream provider.

Given that PE1 is a "normal" PE terminating customers and PE2 a border
router-like PE terminating an upstream provider.

>If you talk about a distant PE receiving the full table from the RR via
>iBGP, the overhead will likely depend on the RD setup. I assume you use
>a common RD for the "internet" VRF on all your PEs, right? 

Correct.

And the RRs (7301s) take very little time to converge. After a minute or
so, the BGP process has fallen to normal levels. The CPUs in SUP720s
aren't overly fast - I am aware of that.

>This could be one reason. CEs advertising > 200k prefixes are 
>not widely seen in 2547bis environments :-)

I know. :-/  It's somewhat of a shame because on a sheet of paper, it
seems like a good approach. In reality, it's not widely used and doesn't
seem to get tested well or have the code optimised for it.

>next to PMTUD, none I could think of without knowing more details.

Already done.
 
>>Any improvements to this behaviour in SRB2 over SXE, SXF or SRA?

>need to know more details..

Alright; hope I supplied them. :)

>From: Rodney Dunn [mailto:rodunn at cisco.com] 

Hi Rodney,

>I need to check this in the lab but if you look at the vpn table are
>we not allocating a VPNV4 label for each prefix when it's in the
>VRF?

On PE2 ("border router"), there's an "in label" for every prefix, yes.
No "out label" for locally learned prefixes (from upstreams).

On PE1 ("normal" PE), there's an "out label" for all prefixes learned
from PE2 and other "border routers". "in label" for locally terminated
prefixes.

As I wrote initially, it's only the BGP process that chews up CPU during
convergence on PE1 and other "normal" PEs.

-- 
Regards
 Christian Bering
 IP engineer, nianet a/s
 Phone: (+45) 7020 8730