Re: Network layer reqt? [was Re: Transport level multihoming]

[Brian E Carpenter <brian@hursley.ibm.com>]

Joe Abley wrote:
> 
> On Wed, Apr 04, 2001 at 05:16:44PM -0500, Brian E Carpenter wrote:
> > Margaret Wasserman wrote:
> > >
> > > >No. But today, connections with TCP/IPv4 and typical IP multi-homing get
> > > >dropped and data gets lost when the prefered path fails. Traffic is lost
> > > >until the routing reconverges which is usually long enough to back TCP off
> > > >into oblivion.
> > >
> > > Is this true?  Can other operators or admins of large sites
> > > comment on this statement, in particular?
> >
> > Well, we've had many anecdotes of very long (>100 seconds) BGP
> > convergence times- Abha reported on some actual *facts* in the
> > IETF plenary- see http://www.merit.edu/ipma/. TCP doesn't
> > react well.
> 
> I'm not sure there is a direct one-to-one correspondence here.
> 
> Edge networks will often route by default in the absense of completed
> BGP convergence. Using Abha's taxonomy, instability in covered routes
> in the core need not necessarily cause packet loss since packets will
> be forwarded according to shorter-prefix aggregates.
> 
> Hence packets may still be delivered (perhaps over sub-optimal paths)
> even while BGP is taking time to converge.
> 
> In my experience it is usually only severe turbulance which triggers
> an ICMP back to an endpoint that causes sessions to die; multi-minute
> packet loss due to loops in response to re-homing is not altogether
> usual, I think.

No, it isn't usual. My point is that transactional middleware has to deal
with it, even if it only happens once a week. 

   Brian