"Chris A. Icide" <chris@nap.net> writes:
A clumsy tech or a backhoe takes out a physical path
between two POPs, each of which contains several routers
that form part of this flat network.
What happens now?
Layer 2 heals itself, without the knowledge of layer 3. Layer 3
experiences a latency increase between certain paths.Sorry, had to do it.
*duck*
Well, this is certainly the obvious approach, and is
certainly in line with the generally reasonable philosophy
of "degraded performance rather than no performance"
response to partitions, assuming the routers don't
notice a topology change.
There are some interesting issues with respect to
convergence at the L2 layer (rerouting VCs takes time
too), and I've seen problems where the fast-packet layer
converged slowly enough that IS-IS adjacencies were lost,
causing really neat routing slosh as the VC went and came
back.
(To augment one of my own questions, when you have two
routers in a full mesh (FR, ATM, whatever using VCs) and
the VC between N-1 and N-2 goes out of service, how do N-1
and N-2 decide what path(s) to use to talk to each other?
Feel free to consider a network using any commonly
available IGP and the iBGP hack.)
I also have some anecdotes about VCs improperly travelling
along the same path for long periods of time because of
configuration errors or unnoticed physical
disconnectivities between switches. There have also been
neat tales of VCs going simplex from time to time.
Monitoring two networks simultaneously, fun fun fun... 
I'm sure there are people here with practical experiences
who might be willing to share what they've learned.
however, in this fast paced industry, will we
ever have the "perfect" protocol (okay, ever is a big
word).
IPv6 over ATM/LANE with RSVP will someday change the world.
Sean.
- --
Sean Doran <smd@ab.use.net>
"Boy, you obviously don't have a clue!" -- Chad Skidmore, Data Source L.L.C.
==>> Layer 2 heals itself, without the knowledge of layer 3. Layer 3
==>> experiences a latency increase between certain paths.
==>>
==>> Sorry, had to do it.
==>> *duck*
I've seen a "latency increase" on the order of 12 seconds round-trip from
San Jose to Reston, VA before. Most layer 3 protocols don't like 6 second
one-way trip time for a packet (almost 100x greater than normal).
Granted, this was a rare case, and the particular design at the time
didn't have a large negative effect (other than the reston-sanjose link);
had it been a hub-and-spoke-based network and reston were a hub, I would
rather let the L2 backups do their jobs than re-route me through
Bangladesh and provide > 2s round trip to a lot of my leaf sites.
==>(To augment one of my own questions, when you have two
==>routers in a full mesh (FR, ATM, whatever using VCs) and
==>the VC between N-1 and N-2 goes out of service, how do N-1
==>and N-2 decide what path(s) to use to talk to each other?
==>Feel free to consider a network using any commonly
==>available IGP and the iBGP hack.)
Depends on the engineering... =)
I've purposely built configurations which consisted of 20 subinterfaces in
order to make a fully meshed network appear as a large number of
point-to-point links in order to combat some of the problem mentioned by
Sean in the "anecdotes" list. This was a re-engineer from a multipoint
network due to a provider whose switched network was a bit green yet (no,
I won't mention who it was, so don't ask).
I've long been a fan of keeping layer 2 as the transport and avoiding
"routing issues" in favor of letting layer 3 control them. Layer 2
should provide the pipes, and layer 3 can do the transport.
I can accept that costs make up a big portion of the decision to use L2
switching in a network. Particularly, the difference between using a
chunk of a carrier's fiber in the form of a switched service vs dedicated
services plays a large part in the decisions made for connectivity.
/cah