(combining two of Priscilla's posts)
At 10:52 PM +0000 1/30/03, Priscilla Oppenheimer wrote:
>MADMAN wrote:
>>
>> Hmmm, IOS imgaes that are approaching, (in some cases
>> exceeding) 20M ;)
>
>I'm not sure what your point it, other than to be funny :-), but I do have
>to say that it doesn't matter that it's a 20 MB file when talking about the
>file travelling across a fraction of an inch within a switch versus the file
>travelling across say a 10-foot cable.
>
>OK, so the first bit would incur maybe an extra 20 nanoseconds of delay. The
>remaining 160,000,000 bits would be right behind the first one and wouldn't
>encounter any extra delay.
>
>>
>> Dave
>>
>> Larry Letterman wrote:
>> > where did the other 1/3 of the speed go ?
>
>Resistance caused by the cable properties. (It should have said 2/3 the
>speed of light in a vacuum).
>
Why worry? Resistance is futile.
>At 5:52 PM -0500 1/30/03, Priscilla wrote:
>>I would think that the L2 headers would still have to be rewritten, for
>>traffic going through the router part of the swouter, (my new name for a
>>cross between a switch and a router.)
>
>I rather like that. If you had chosen to call it a ritch, that is
>something we in the industry are not, these days.
If you're doing L3 decisionmaking, I don't see how you'll get any
performance improvement in L2, assuming you aren't breaking the rules
of routing. An L2 switch, true, can pass the MAC addresses unchanged,
and, in the strict scheme of things, doesn't need to recompute the
FCS.
If you are making decisions at L3, you aren't going to get any
particular benefit given that you need to substitute the router
egress port MAC address for the previous-hop source, and recompute
the FCS. But, since FCS computation is routinely in hardware, I can't
see that as being an issue.
Now, some Cisco switches play games, and associate a MAC address with
an L3 address, and don't do L3 lookup. If you are going
subnet-to-subnet, you introduce several potential issues:
--security: what happens if the MAC address or its mapping changes?
--ARP: how does it resolve if the target subnet thinks it's getting
a frame based on L2 information? This violates the local
versus remote axiom of IP.
Ye canna violate the laws of routin', Kiptin.
>I could believe that it's much more
>efficent on the swouter than on a router, though. For one thing, the swouter
>probably has modern hardware components and a more optimized architecture.
>Anything else you can say on this aspect?
But isn't that a product implementation rather than an architectural
question? On the router designs I worked on, which were
unquestionably architected for L3 decisionmaking, most of the
per-frame processing was on the ingress forwarding card, such as FIB
lookup.
That sort of hardware, indeed, is expensive. It made sense with
multiple OC-192, but wouldn't for a price-optimized SOHO router. But
look even within the pure "router" Cisco product line, and you'll see
all manner of price-performance tradeoffs. Non-modular is cheaper
than modular. Having less memory expansion is cheaper than having
addressing space and card footprint for more.
Frankly, I see very little difference between an L3 capable switch
and a high-performance router -- but you very well may not need the
performance.
In a very-high-end router, the main contributors to delay are
extensive preprocessing or postprocessing (e.g.., QoS, encryption),
and the delay in getting the frame across the fabric (I'm not even
touching multicast). Shared bus architectures run out of steam at
about 2 Gbps, and you need to go to shared memory or crossbar.
Memory speed is a constraint as well, so crossbar has more growth
potential until we can make pure optical decisions.
Message Posted at:
http://www.groupstudy.com/form/read.php?f=7&i=62229&t=62166
--------------------------------------------------
FAQ, list archives, and subscription info: http://www.groupstudy.com/list/cisco.html
Report misconduct and Nondisclosure violations to [EMAIL PROTECTED]
