[AusNOG] powerful routers in core/edge routing/switching

[Lincoln Dale]

On Wed, Feb 13, 2013 at 11:06 AM, Joshua D'Alton <joshua at railgun.com.au>wrote:

> The problem is routing has lagged far behind switching in terms of port
> density, capacity, throughput etc. Obviously a switching engine is peanuts
> compared to a routing engine, but it is exaggerated by the massive amounts
> of features they put in routing engines.


I gave a talk about this at the last AusNOG. the pdf of the talk should be
on the AusNOG site.

Reality is that some switching silicon is sitting on Moore's Law and you're
getting an approximate 2X improvement in speeds/feeds/density every 2
years. I don't think any "routing" silicon is sitting on that same curve,
for a variety of reasons.

"Routers" generally have 'software' running each port so are infinitely
flexible (to the extent that there's an infinite number of monkeys capable
of writing microcode or software on whatever the NPU or FPGA is behind each
port).

"Switches" tend to bake things into silicon such that they cannot be
changed and it isn't flexible.  But again, Moore's Law (which really
relates to transistor density and not to speed/performance) means that
often "switchports" are gaining new capabilities which are more
'router-like' in capability and gains in things like
flexible-packet-parsing and flexible-packet-rewrite means that there is
some "microcoded-like" stuff in some modern switch silicon which makes it
possible to support new packet formats or new encap methods.

The gap will only continue to widen.


>
> Seems to me we almost need a new breed of edge routers, ones that just
> talk BGP to other providers, and the current edge can stay as they are
> handling fancy things like MPLS which is really more of an internal
> routing, therefore switching, feature. Or not :)
>

on the wire, MPLS is just yet-another-tagging format.  if we take the
company I work for, almost all of the switches we sell have silicon capable
of MPLS.
the harder part of MPLS is the protocol stack that goes with it and the
many hundreds-of-man-years of what that protocol stack requires for a new
vendor to do the 'equivalent' of the full MPLS feature-set that some
vendors have.

Some folks have talked about how they've separated control-plane into a
centralized controller and just programme labels in the 'core'.[
http://opennetsummit.org/talks/ONS2012/hoelzle-tue-openflow.pdf]
The real benefit I see of such a thing is really about cost.  That 'core'
can now be "cheap switch ports" rather than "expensive router ports."
There's plenty of other benefits too in terms of flexibility of what those
devices could be.

Many folks already did this kind of thing in the past.  e.g. used a Cisco
6500 rather than a GSR.
or used a Juniper MX rather than Juniper M or T series.

With regard to you wanting a whole-new-breed-of-routers, who do you think
is going to provide that?
You really think that a C or J selling a port today at $X is going to be
perfectly happy selling that same port at $X/10? :)


cheers,

lincoln.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.ausnog.net/pipermail/ausnog/attachments/20130213/f6842ecc/attachment.html>