North American Network Operators Group

Date Prev | Date Next | Date Index | Thread Index | Author Index | Historical

Re: outside the box Was: The Gorgon's Knot.

  • From: Dave O'Shea
  • Date: Fri Sep 28 17:02:00 2001


Joseph T. Klein [mailto:[email protected]] Sez:
> 
> 
> At 21:32 +0200 28-09-2001, Stephane Bortzmeyer wrote:
> >On Friday 28 September 2001, at 18 h 17,
> >"Joseph T. Klein" <[email protected]> wrote:
> >
> >>  would have problems given larger route tables. We all 
> don't have routers
> >>  that can easily chew through a 100,000+ line BGP table.
> 
> OK ... I amend my prior statement 250,000+
> 
> Yet, I don't belive every corner of every network has routers that can
> easily chew through 100,000 route tables.

Not the corners, probably. But certainly the areas more towards the
core.

> 
> >
> >I believe it is a legend. Unless you use Cisco 25xx to have 
> a full BGP feed.
> 
> Yeah right. I suggest you look at real world loaded 7200s. They have
> problems with full routing tables.
> 

I won't say a 7200 is just going to slice through them, but, they manage
as long as you don't toss too much work at them. Can a 7200 be upgraded
to 256mb? I am not using them these days so I forget...

I have a feeling that when looking at 12-18 month growth of the global
BGP table, there may be a point where it becomes fairly important to
move routing analysis and processing out of the actual boxes that are
moving the traffic, and into boxes that are optimized for it. Even with
the current tables, a cisco 120xx box can work up a bit of a sweat when
you "cl ip bgp (insert favorite AS here)" -- a lot more CPU than it
would burn up in forwarding 2-3gbps of traffic.

Even the current router-as-route-reflector technique uses a system
optimized for packet forwarding to do more math than it was really built
for. Routers are designed to beat existing benchmarks -- and the most
common benchmark is "how many packets can I stuff into this thing?".

> >
> >Any Taiwan-made PC can swallow much more. The limit is not 
> clear but is
> >certainly far away from us.
> 
> I want to you to put a couple of channelized DS-3s, an ATM OC12c,
> and a POS OC48c to your backbone plus all the BGP peers you can sign
> up at AADS on a PC.

There's some truth to the idea that a halfway capable gamer's machine
has more actual processing power than most high-dollar routers. I'd
guess that the difference is that a PC is optimized for analyzing and
manipulating data, with I/O viewed just as a way to get more data.

That said, I am *not* giving up my Quake/Unreal machine to serve as a
route reflector until someone figures out how to stuff a GeForce3 into a
GSR. My wife will beat me with a stick if I tell her I need to run 48vdc
into my den. The 3KVA UPS already got the "Martha Stewart would *not*
approve" look.

I guess the model that comes to mind is something like MAE-West, which
at one time was using something like a pair of Sparc-20's to handle
several hundred various BGP sessions, and left the packet shuffling to
hardware that was designed for the purpose. 

> 
> The black and white simplicity expressed by people on this forum is
> unbelievable.

.. and usually based on the assumption that everyone makes their
decisions based upon the same experiences and observations.