[c-nsp] BGP fast converge

[Oliver Boehmer (oboehmer)]

Mario Velazquez <> wrote on Friday, July 28, 2006 3:27 PM:

> Hi, i am having some problems in my metro network, bgp convergence is
> really a problem, i was looking arround for some fast convergence
> features like BGP Fast Peering Session Deactivation and BGP Next-Hop
> Address Tracking. this is what i have so far:
> 
>
> CE1 and CE3 are in the same site, the same for CE2 and CE4, the CE-PE
> connection is bgp.
> All CE are running ospf with more or less than 6000 routes each.
> the problem is between CE1/PE1 and CE3/PE3 because when if one port
> fails in SW1/SW2 the bgp session will not be dropped until the dead
> timer expires (180s).
> 
> i have this configuration in all CE
> 
> outer bgp 222
> no synchronization
> bgp log-neighbor-changes
> network 20.0.0.0
> neighbor 1.1.1.1 remote-as 13591
> neighbor 1.1.1.1 ebgp-multihop 3
> neighbor 1.1.1.1 fall-over
> neighbor 3.3.3.3 remote-as 13591
> neighbor 3.3.3.3 ebgp-multihop 3
> neighbor 3.3.3.3 fall-over
> no auto-summary
> 
> But nothing seems to change, in the case of a port fail in SW1 or SW2
> bgp stills wait for the peer dead timer.Is there something missing? or
> am i using the wrong feature? or i have a bad config

Is there a specific reason why you enabled ebgp multihop between CE and
PE? Because if you used a direct peering using the link address(es), a
link-down of the CE interface would trigger session-down without any
configuration.

But regarding "fall-over": This mechanism relies on the neighbor
disappering from your RIB (i.e. no more routes towards the neighbor). If
you connect to the PE using a switch and the remote switch port dies,
you will still have the route towards this neighbor in your RIB (as your
side of the link is still up), so FSD will never fire.

To cover these failures in a scalable manner, you need the interaction
between BFD and BGP, but this is brand new and only available in some
releases (BFD for BGP, "neighbor fall-over bfd"). Then BFD will track
your neighbor reachability, and signal BGP when it goes away. 

For the time being: Why not just decrease the hello timer? You do want
to check with your peer as tuning it too low is not really polite as
this doesn't scale..

	oli