Not disagreeing with Sean, which whilst amusing is often unproductive:
One of the higher clue arguments in favour of maintaining an L2
mesh (and 'routing' protocol) is that it converges quicker than
than L3 as in theory it has less nodes. The normal counter arguments
1. that L1 (SONET/SDH) converges quicker still.
But the reason ISPs adopted NBMA L2 networks in the first place is
because the telcos convinced them that L2/L3 integration was going
to happen (or because the ISPs got a cheap deal in support of the
telcos' agenda of making that happen). The ATM Forum has done a great
job on the IP/ATM integration front (I don't think).
As you say, L1 converges quicker still, which is a very good
argument for running IP on top of Packet Over SONET/SDH and
using MPLS in support of IP virtual-circuits to do QoS routing.
2. as MCI/Worldcom may well have aptly demonstrated, L2 routing
protocols have their own problems.
MCI/Worldcom is just a crappy bureaucracy - that's the real reason.
3. L3 routing can be hierarchicalized (sp?) and tweaked to converge
just as effectively - witness sending your loopback IP's round as
next hops in BGP and running an underlying IGP such as OSPF/ISIS/EIGRP
which cuts out many (though unfortunately not all) recomputations
when the underlying fabric changes.
But L3 routing is not the issue...
4. Implementation of L2 heavy protocols on routers is often poor. Witness
how long it took Cisco to implement OAM management. So just how long
does it take to notice a VC's gone down?
Telcos like nailed-up QoS products because they understand them.
ISPs like best-efforts QoS products because they understand them.
Cisco likes both (and buzzwords like 'integration') because it's
into aquiring market share in both markets.
The Internet is never (IMHO) going to become a fully circuit-switched
QoS network because of all the cooperation that has to happen to prevent
it melting down when operating as a best-efforts network (which is what
it basically still is don't forget).
MPLS (despite what the telco spin-doctors might claim) is not really
about integration with ATM - it's about introducing levels of QoS
appropriate to the services ISPs want to provide over their own
networks (and perhaps, given a year or so, in conjunction with peer
networks as well).
VoIP doesn't need the predictably precise levels of jitter (with
associated protocol overheads) and suchlike that the telco ATM
(reinvented voice-switching) weenies claim will be required to
provide a decent QoS for voice traffic.
A more convincing argument for buying ATM circuits is that they
(are) / (used to be) dirt cheap. Recent experience suggests there may
be a reason for this.
The telcos are flogging a dead horse with IP over ATM QoS integration.
Best pretend that's why you're buying ATM if you want a cheap deal (or
just invoke your MCI/Worldcom contractual QoS guarantees ).
GX Networks (formerly Xara Networks)
I think it's high time you dropped the 'formerly' stuff and learn
to live with the fact that you sold out.