[c-nsp] Performance hit for Truncated vs Compact

[Richard A Steenbergen]

On Sat, Jul 22, 2006 at 09:06:07PM -0400, Chris Griffin wrote:
> Hi all,
> 	When a 6509/Sup720 populated with all dcef720 fabric cards moves into 
> truncated mode because of the insertion of a non-fabric card, what is 
> the anticipated hit on performance if nearly all traffic is between DCEF 
> cards?  I understand central forwarding takes a hit in terms of PPS, but 
> can you characterize what hit a DCEF linecard forwarding across the 
> switch fabric might take?

None.

Centralized forwarding lookups (to the PFC) travel over the classic bus 
even on fabric cards, which is the root cause of the 30Mpps limitation 
without any DFCs. When you insert a non-fabric card the lookup mode must 
change to something that all the cards on the shared bus can understand, 
aka it switches from 32 byets/lookup (compact) to 64 bytes/lookup 
(truncated), which cuts your PFC performance to 15Mpps. Any actual data 
traffic you put onto the classic bus cuts into this even further, and when 
you run out of classic bus you also run out of PFC lookup capacity. You 
can see the usage with "show catalyst6000 traffic-meter" (on the sup720 
that comes with 3b/3bxl it seems, sup720-3a doesn't support this even 
after you upgrade the pfc to 3bxl.

When you have DFCs, the lookups happen on the DFC rather than being sent 
to the PFC. If you're sending data to another fabric enabled card, there 
would be no traffic put onto the classic bus at all. Not sure how the 
forwarding data is distributed to the DFCs (over fabric or bus), so I'll 
let someone who knows answer that one.

-- 
Richard A Steenbergen <ras at e-gerbil.net>       http://www.e-gerbil.net/ras
GPG Key ID: 0xF8B12CBC (7535 7F59 8204 ED1F CC1C 53AF 4C41 5ECA F8B1 2CBC)