[c-nsp] Full Routes / iBGP Issue & Question

[Pete Templin]

On 9/25/14 6:40 AM, David Deutsch wrote:
> Hello all,
>
> Please excuse my disorganized posting/question, as this is my first 
> posting to the mailing list.
>
> Currently I am attempting to implement two ASR1002-X routers on my 
> network edge, each with diverse eBGP provider connections receiving 
> full routes. Everything up to a point has gone very well, including 
> reception of full routes, blocking of default routes, local prefs, 
> etc. However I am getting some odd results when iBGP connecting the 
> two together.
>
> I appears that certain routes from each other are being filtered out, 
> so each router is in some cases aware of both routes to a prefix and 
> in some cases aware of only one. Examples:
My advice to "new BGP students" is to slow down, focus on one route at a 
time, and "think like a router". So, pick one route, go onto one of your 
routers, and do 'sh ip bgp a.b.c.d/e'.

What do you see?  Likely an EBGP route, and possibly an IBGP route.
Which one is best? Why? (Hint: dig out the path selection algorithm, and 
identify the exact step why)
Is it advertised to anyone else?

Now, hop on the other router, and do 'sh ip bgp a.b.c.d/e' for the same 
route.

What do you see? Likely an EBGP route, and possibly an IBGP route.
Which one is best? Why?
Is it advertised to anyone else?

I suspect you have a situation where R1 might be getting a really "good" 
path for route X (EBGP), so it advertises it to R2. R2 is getting a 
really "bad" path for route X (EBGP) and the really "good" path from R1 
(IBGP), so it prefers the route from R1 and therefore has no reason to 
advertise its learned EBGP path to R1 (since it isn't 'best').

Meanwhile, for other routes, R1 might be getting a path for route Y 
(EBGP), so it advertises it to R2. R2 is getting an "equally good" path 
for route Y (EBGP), even when comparing to the IBGP route passed over 
from R1, so it advertises its external path to R1. At this point, R1 
recalculates, but it still likes its EBGP route, so R1/R2 "agree to 
disagree" essentially.

Realize that for route X, if R2 happens to learn the EBGP path first, 
it's the only path so it wins, and R2 advertises it to R1. Then when R1 
happens to learn its EBGP path later, that one wins, so R1 picks that as 
best path and advertises it to R2. At that point, R2 recalculates, and 
decides the path from R1 wins, so it sends a withdrawal of the EBGP path 
it was advertising earlier. So sequence is a factor, but the end result 
is the same.

pt