RE: [nsp] Fabric Line Cards

From: David Sinn (
Date: Fri May 31 2002 - 14:01:31 EDT

In actuality it is not.

Fabric cards will still send just the compressed headers on the
backplane. If the destination is another fabric enabled card, things
happen as normal.

If the destination is a non-fabric card, the results will come back to
the ingress line-card at which point the ingress line-card will send out
the whole packet in non-compressed format. The PFC will have already
made the decision so it simply sends the results out and the non-fabric
card processes the packet as if it had come from another non-fabric card

If a non-fabric card is the ingress, it follows the normal non-fabric
method regardless of destination.

The upshot of this is that you can see greater then 15Mpps from the
chassis if the bulk of the traffic you have is fabric to fabric. The
negative is that for every fabric to non-fabric packet, you eat an extra
64 bytes over and above the full packet that also needs to transit the
backplane (thus you could see less then 15Mpps of forwarding capacity).

If they had locked the box into non-compressed format then you would
have gotten none of the benefit of the fabric if you put a single
non-fabric card. I guess they decided that would be bad.


-----Original Message-----
From: []
Sent: Friday, May 31, 2002 4:05 AM
Subject: RE: [nsp] Fabric Line Cards

This paper on the 6500 Architecture is good:
except it completely ignores exactly this scenario.

Presumably the header compresssion used to send packet headers over the
switching bus
when switching packets between fabric cards is disabled when mixing with
non-fabric cards?
This would pull performance down to something like Supervisor1/PFC
performance of 15Mpps.

If you use Distributed forwarding on the fabric cards you may of course
achieve higher switching speeds,
when switching between fabric cards.


This archive was generated by hypermail 2b29 : Sun Aug 04 2002 - 04:13:46 EDT