[nsp] Re: Catalyst 6500 Hybrid

[Mikael Abrahamsson]

On Sun, 2 May 2004 sthaug at nethelp.no wrote:

> I'm not sure cisco-nsp is a good place for a Cisco - Juniper "slug
> fest". That said, I think it's reasonable to have *some* discussion
> about the advantages and disadvantages of the various platforms.
> Jared?

Defiantely. For me, this isnt a vendor argument, this is a platform 
argument.

In one corner: Cisco 12000 and basically all the Junipers and in 
some aspect perhaps the 7300. Real routers, real hardware LPM routing, 
deep packet buffers.

In the other corner, L2/L3 switches, more or less hardware enhanced
routing, most of the time small packet buffers. Here I'd place the
6500/7600 (even though the OSM cards complicate things). Juniper has
nothing here. On the other hand, Extreme and Foundry does.

A couple of years ago the argument was pretty easy to do, the switches
just didn't do proper routing. With the Sup720, Extremes BD10k, Foundry
has something (I dont know about them) that'll do real routing. They still 
lack deep packet buffers.

I have talked to vendors about the size of buffers. Seems everybody is 
making the argument that deep packet buffers costs HUGE amount of money. 
Even putting 80ms instead of 5ms (no to even mention the 600ms buffers on 
the real routers) is a big price hike.

Now, if we're talking just about the Cisco 7600 versus the 12000:

Ok, I want a chassi, 4 10GE ports and 10 1GE ports. If I go with the 7600 
LAN cards I can do this for less than $100k. If I go with the 12000 I'd 
have a hard time doing it for less than $500k. Doing it with 7600 OSM 
cards puts it somewhere in the middle.

My argument is that deep packet buffers is needed when lines are full.  
Unless you actually reach 95%+ utilization on a 10gig line, you'll never
see any buffering. Here is where the aconomic sense of the 12000 breaks
down. I can buy twice the capacity and build two networks, for less than
half the price than for a single 12000 based network.

I like the 12000 and Junipers, these are real routers and they're 
appropriate core tools, but THEY ARE TOO EXPENSIVE. I'd rather build two 
cores and know I'd never overload it, than to build a single core that'll 
handle occasional overload properly.

Now someone will make the argument that 10GE is not WAN/core. My answer to 
this is that newer DWDM systems have 10GE tribs, and WAN PHY (OC192 
compatible Xenpaks for 10GE) will enable 10GE for the WAN as well.

Now all of a sudden I have interchangable framers and optics, I only have 
to use a single card type (great for spare management) for the platform.

My opinion might be scewed by the fact that I work for an ISP that 
has plenty of own fiber infrastructure and DWDM capacity we produce on 
that fiber ourselves, and the fact that the bandwidth price of internet 
capacity in Sweden is ridiculously low.

-- 
Mikael Abrahamsson    email: swmike at swm.pp.se