Packet Latency Changes with DFC

igaffine · ‎10-10-2007

We have been performing some latency tests across Cat6509 switches with SUP720-3B and 67xx blades and wondered what the expected improvement (if any) should be after we installed DFC modules on the blades. We used the DFC-3B modules, so that they matched the supervisor hardware.

During the testing, we noted that the installation of a DFC-3B on each 67xx blade provided no noticeable latency improvement, however, at this stage no load is being placed on the 6509.

Without the DFC modules, I expected the traffic to have the following flow. This is a layer III traffic test within the same 6509:

Ethernet Port--(Switch-Fabric)--(SUP720-Central FDB)--Switch-Fabric)--Ethernet Port

This is with the clients being on the same 67xx blade.

With a DFC installed, I expected:

Ethernet Port--(DFC Module)--Ethernet Port

I guess that there would be an initial flow to allow the switch to confirm the destination of the packet flow, but after that it would be processed by the DFC.

No non-67xx blades are installed in the cat6509.

I may have this completely wrong and the DFC will come into play for switching throughput, rather than improving latency.

Any feedback on this would be appreciated.

Kind regards,

Ian

glen.grant · ‎10-10-2007

Myself I think these cards are more suited to the distribution layer where you would be switching large amounts of traffic between subnets as the dfc cards keep their own copy of the cef routing table and the cards send the traffic directly to the switching fabric instead of the supervisor than to a setup where it is mostly client traffic but maybe I am missing something. Obviously they have a much bigger backplane capacity at 720 gbps than older boxes so maybe in a big server farm where a lot of traffic is going ojn just within the 6500 itself it would help a lot .

Joseph W. Doherty · ‎10-10-2007

". . . however, at this stage no load is being placed on the 6509. "

What the DFCs really do is up the pps rate. Sup720 is rated for 30 Mpps for the entire chassis using fabric cards. This would support about 20 Gbps of traffic (with 64 byte packets). With larger packets, much more bandwidth could be pushed.

Only when you really overtax the CFC, will the DFC advantage become useful.