[quagga-users 14423] Re: Quagga, BGP and BFD with Cisco

Kurt Jaeger q-u at opsec.eu
Thu Sep 1 08:47:09 BST 2016


> >>  Second, BFD has always been underwhelming on server platforms,
> >> like those that Quagga typically runs on. [...]
> >> I'm not aware of anyone that can do BFD in the control
> >> plane, and support BFD rates less than 100ms.
> > 
> > ICMP round-trip can be far below 100ms (below 1ms is easy), so at least
> > the packet processing for BFD can be fast on a server, as well.
> > 
> > Why should a server have problems with the reaction on a missing
> > link so that this reaction takes over 100ms ?
> Because Unix isn't a realtime OS?  And ICMP echo reply processing
> is handled in the kernel, not user mode.

So you suggest to put BFD in the kernel ? That's a good idea,
because the most important reaction would be
- to mark the interface configured for BFD as down as soon as
  the condition is reached
- and move the RIB to a secondary RIB that was precomputed for the case
  of that link missing.

With that setup fast reaction times should be possible, even on servers,
don't you think that might work ?

> So no context switches, no swapping, etc.  And I said, 100m is
> probably achievable in Quagga, assuming the BFD patches work.  Going
> less than 100ms is where things will get more difficult.

If we keep BFD in user-space, it might become a factor.

>   And ICMP echo reply is only one packet per second during a
> typical test.  Lets say you have a 50ms BFD rate (which is a typical
> rate).  And you have 10 BGP sessions.  That is 200 packets per
> second in BFD traffic.

Our core routers run with quagga and process much more than that, so
I'm not afraid of the packet rate. If we keep it in the kernel 8-}

pi at opsec.eu            +49 171 3101372                         4 years to go !

More information about the Quagga-users mailing list