From: asadovnikov (asadovnikov@comcast.net)
Date: Wed Jun 02 2004 - 00:40:23 GMT-3
May I ask why do you need to interconnect the core switches in the first
place?
I know it sounds like silly question, but let us look into it... Let us say
such link is not there... I would assume that all distribution switches
have an uplink to each core.
CORE-A CORE-B
| \ __________/ |
| \__/_______ |
| / \ |
| / \ |
Distribution-A Distribution-B
Say Distribution-A has a packet to ship to Distribution-B, and it sends it
to CORE-A (for the sake of example). CORE-A would send it to Distribution-B
directly.
So when traffic hits one of the core switches it would not ever go to
another core switch, except if there was a link failure. Further, if only
one distribution to core link fails, all other distribution switches would
use the core box which continues to have connectivity to this distribution
switch.
CORE-A CORE-B
| __________/ |
| / |
| / |
| / |
Distribution-A Distribution-B
In this case Distribution-A would know that the best route to Distribution-B
is via CORE-B, and will forward all traffic there.
So only possible need for one core to need to forward to another is in
situation of dual link failures, like this
CORE-A CORE-B
| |
| |
| |
| |
Distribution-A Distribution-B
Even then you have then another dozen or so distribution switches (not shown
on the diagram above) which the cores can use to route around dual failure,
even though it will introduce suboptimal routing.
Much I hate any situations of suboptimal routing I have to ask what are the
chances of your network to run with 2 hardware failures on the core. And my
take would be that it is real low, as each failure on the core should be
promptly fixed.
Having said that you may want to interconnect the cores just in case, but
then since we have just demonstrated that such link would be almost never
used, oversubscription question is much irrelevant. If you want to be on
really safe side by all means do provide core crosslink, but I would not
sweat about how large it is... may be 2-4 gig (or if a lot of spare money do
2x 10 gig).
Obviously each network requirements are so unique, it is almost impossible
to have one-size-fits-all answers. So do send additional information.
Best regards,
Alexei
P.S. QOS on the crosslinks/uplinks is a good think, assuming you have
classification/marking on the edge and can actually tell what traffic is
more or less important. For example if you provide voice services, and
voice is marked on the network entry it is real good idea to give it
priority over data, and with today hardware implementations of QOS there is
no performance penalty to pay.
-----Original Message-----
From: nobody@groupstudy.com [mailto:nobody@groupstudy.com] On Behalf Of
Kenneth Wygand
Sent: Tuesday, June 01, 2004 11:48 AM
To: Group Study
Subject: Connecting two core switches / Design
Hello all,
I was thinking about large providers that need to connect core switches at
very fast speeds. For example, say I have two 6506 core switches I wanted to
connect in the backbone. Say each one is terminating 20 - 30 fiber gigabit
speed links. I would think it would be a poor idea to oversubscribe the
link between these switches to a 1GBit or even 10GBit link. If the link was
oversubscribed, logic would think that QoS should be employed here to make
sure critical traffic (voice, video) gets through first, but all design
guides point towards keeping all QoS out of the core to simply switch
packets as fast as possible...
What is recommended and is there any documentation / experience anyone can
contribute?
Thanks!
Kenneth E. Wygand
Systems Engineer, Project Services
CISSP #37102, CCNP, CCDP, ACSP, Cisco IPT Design Specialist, MCP, CNA,
Network+, A+
Custom Computer Specialists, Inc.
"I am not really smart. I just stick with problems longer." -Albert Einstein
Custom Computer Specialists, Inc.
"Celebrating 25 Years of Excellence"
[GroupStudy removed an attachment of type image/gif which had a name of
image001.gif]
This archive was generated by hypermail 2.1.4 : Sat Jul 03 2004 - 19:40:31 GMT-3