Stephen,
I've always stated that "switch" is a marketing term meaning
"fast". Thus a
"L2 switch" is a "fast bridge" and a "L3 switch" is a "fast
router". In this light, the Yankee Group is just now
catching on to something we all knew a decade ago -- slow
(i.e. software) routers are dead.
As you are probably more aware than I, software-based-forwarding routers
will die when people stop running the so-called "desktop protocols", and
even then, most next-gen routers will continue to need functions that can
only be provided economically and perhaps thermodynamically (in terms of
heat dissipation) in the form of sw "services" running on purpose-built
and/or general-purpose CPUs. Examples are VOIP call processors, some FW
ALGs as new protocols emerge, etc. The concept of L2 switching based on L3
information tends to be viable only when one can transparently bridge
between the L2 protocols - otherwise, you are making L3-only decisions, and
doing all sorts of L2 rewrite which many traditional Ethernet switches can't
necessarily do.
Things are getting better, but "L3-switches" pale in comparison to today's
high-end routers on almost all fronts. If you take GigE out of the
equation, modern "L3 Switches" are just as expensive as modern "core
routers" - and routable, "mpls-able" L3 GE ports are _more_ expensive on
"switches" than "routers" (see 4xGE OSM vs 4xGE GSR 'tetra' pricing). Media
diversity, queuing performance, and FIB density is what really
differentiates the two at this point, IMO. I am unaware of a traditional
switch-turned-router (and I use these terms here as most do who draw a
distinction) that can exceed the forwarding capacity of a core router when
the media is largely WAN-based, there are complicated classification and
filtering rules that are very dense, when complex queuing policy needs to be
applied, and when the routing table is huge.
Or perhaps my earlier experience with these switches-trying-to-be-routers
has left me a bit jaded....
There's a more interesting level to the discussion if you
look at what carriers are interested in for their backbone
hardware today; while I'm obviously biased based on my
employer, I've seen a lot more emphasis on $20k-per-10GE-port
"L3 switches" than $200k-per-10GE-port "core routers" in the
current economic climate.
Of course, a "routable" 10GE port does NOT cost $20k - sure you can do MLS
or whatver it is called - but things like label imposition/disposition is
not possible. Also, last I saw, my MLS-enabled MSFCs weren't able to gather
"vlan interface" statistics - they were all embedded in some L2 asic that I
had to glean from the "switch." Further, Ethernet has the worst OAM
capabilities of any modern media. BFD will help detect failures when it is
available, but will never be able to tell me "why". SONET is clearly
superior in the aspect.
So, for enterprise switching, "L3 switches" are mostly fine - barring any
funky bridging requirements (Blue protocols). But for carrier backbones, I
suspect we will continue to see the majority of implementations usng modern
"core routers". And we haven't even begun talking ATM and FR, and what
device better suits these applications. Judging from your company's
position on this front, I suspect that "core routers" may be our best bet
here, given that many who could do switching well were unable to "bolt on" a
usable, stable routing implementation. But that is another religious
discussion for another day!
My .02
chris