>
> Also, you've put these in rejected:
>
>     Lower the default MRAI timer for iBGP peers to 0
>     Lower the default MRAI timer for EBGP peers to zero
>
> There is work, including some interesting academic work, that suggests 0
> is a bad value for MRAI. /Some/ MRAI is advisable to damp down BGP's
> incredible and ever-increasing (exponentially) propensity to send messages
> as MRAI approaches 0. That work, which was quite old, suggested the best
> value was likely under 5s for Internet use - i.e. for eBGP.
>
> A quick summary and links to the primary works are in:
>
>   https://tools.ietf.org/html/draft-jakma-mrai-02
>
> No MRAI is likely ungood, from what's been described (empirically and
> theoretically). However, the current 30s in Quagga is certainly way, way
> too high, we'd surely all agree.
>
> We /should/ lower that dramatically - just not to 0. I'd be happy to have
> the eBGP MRAI set to 5s, and the iBGP one to 1s. Note that some increase in
> messaging and hence CPU is unavoidable, but the convergence benefits should
> make it worth it for most.


FWIW IOS has used 0 for iBGP and 5 for eBGP since ~2004.  We (cisco at the
time) never had any complaints when we changed to these values (they were
both 30s before). The only reason we didn't do 0 for eBGP was because
dampening was still used a fair amount back then (not so much now).

With the MRAIs set to 0 you can get some bursts of path hunting but they
tend to be short lived because bgp ends up converging so fast (sub-second
is typical).  Given the lack of complaints when the iBGP timer was changed
to 0, I don't think increased path hunting is much of a concern. Path
hunting is worse as meshiness increases and an iBGP core tends to be more
meshy and eBGP connections.  I think we could use 0 for both without any
path hunting issues.

On the off chance that your peer is using dampening you could increase the
MRAI for that one peer.

Daniel
_______________________________________________
Quagga-dev mailing list
[email protected]
https://lists.quagga.net/mailman/listinfo/quagga-dev

Reply via email to