[c-nsp] Non-default BGP hold / keepalive timers

[Simon Leinen]

David J Hughes writes:
> This is just a follow-up on an earlier thread.  When I first asked
> this question nobody could provide me with any data on tested values
> etc.  I thought I'd punt this to the list for the sake of general
> interest (and the archives :) just so the next person might have a
> point of reference.

Thanks for that.

> To overcome the problem of silent failure of eBGP peers that are not
> directly attached (eg via an ethernet switch) we are running much
> reduced timers.  Default Cisco is 3 * 60 seconds.  Default Juniper
> is 3 * 30 seconds.

That's also what RFC1771 and its proposed successor,
draft-ietf-idr-bgp4-26.txt, say: 30 seconds KeepAlive, 90 seconds
HoldTime.

> We are running 3 * 5 seconds without a single flap.  No problems
> experienced in over a month with several different upstream
> providers.

As another data point, we have been running 10/30 (3 * 10) on all iBGP
sessions and with a few upstreams for several months with no ill
effects.  Our routers are mostly 7600s.  We carry only about 28000
IPv4 routes.

Before we started changing our BGP timeouts, I had also looked for
advice from others who use non-standard timeouts.  CENIC also use
10/30 and reported no problems with it.

> And, as Bruce Pinsky has said before, this is no substitute for BFD
> or some other layer 2 fault detection mechanism.  But, hey, it may
> help out of BFD et al aren't an available option.

Right, like if you cannot use 12.0(31)S or 12.4(4)T.  Which is
certainly the case for us (on most routers).

Regards,
-- 
Simon.