[nsp] WS-X6148-GE-TX

dre andre at operations.net
Thu May 8 15:46:35 EDT 2003


On Thu, May 08, 2003 at 04:50:55PM -0400, Temkin, David wrote:
> Nope, I was talking about the 10/100/1000 module... I was curious to know
> how it was oversubscribed (ie, is it like the cat4k oversubscribed
> 10/100/1000 module) where there are groups of 3 or 6 ports together in a
> group or is the whole module just entitled to 16Gbps on the backplane?  Etc.
> etc.  - I couldn't find *any* documentation on it.

Ok so for the 10/100/1000 6148 (sorry I realized this *is* what you were
talking about from the subject line, which I noticed after i send the email):

WS-X6148-GE-TX: Classic LC

Two DMA channels, each at 8 Gbps (16 Gbps full-duplex) each connecting to
the regular switching bus.  This is SHARED across all modules.
This means 16 Gbps for the whole chassis (not on any type of port-basis)
in each direction (inbound and outbound), or in marketing speak "32 Gbps
full-duplex".  However, assuming only one module, it can still push 16
Gbps in each direction.

WS-X6548-GE-TX: CEF256 LC

This gets more tricky.  It has the above connection to the regular switching
bus, but it also has a single channel connection to the fabric.

Best explained as -
"Modules which connect to both a 256 Gbps fabric and 32 Gbps Bus. These cards
use the centralized CEF Engine on the PFCx for forwarding up to 30 Mpps."

However, the connection to the 256 Gbps fabric is only 8 Gbps (each way, so
16 Gbps total -- I believe so from what I'm reading).  This means that the
payloads can't exceed 8 Gbps on the whole module, however the headers can
get up to 16 Gbps!@#@!@  This makes absolutely no sense, but if what I'm
reading is correct, that's how it works. ;>  Unlike the regular switching
bus, however, this is not SHARED across all modules, but specific to each.
Actually it is shared, but across a 128 Gbps one-way backplane instead of
a 16 Gbps one-way backplane.  Man I should re-write this paragraph ;<

Now this is the perfect module for a 6513 in slots 1-6 (the Sup720's go into
slots 7 and 8), because the 6513 is limited to a single fabric connection
for the first 8 slots (all other slots, 9-13 can have dual fabric connections).

The 6509 chassis has no such limitation, so these 10/100/1000 modules are kind
of a waste if you are trying to take advantage of the 720 Gbps backplane.

Finally, this module can (and probably should) use the DFC3.

** My suggestion:
1) Live with the limitations that the 6148-GE-TX module has (16 Gbps shared bus).
2) Only use the 6548-GE-TX in slots 1-6 in the 6513 chassis and again live with
   the 6548 limitations (8 Gbps each way).  Definitely use the DFC3 with this
   guy, and make sure to setup box-to-box traffic on the same blade (now you
   have a trade-off in terms of performance vs. reliability/serviceability) when
   possible.
3) Go with Foundry, ExtremeNetworks, Force10Networks, or Timetra switches.
   I don't know if they do any better with 10/100/1000, however.

So the best you can do with Sup720 10/100/1000 that's not box-to-box on the
same blade would be 96 Gbps on a fully stacked 6513 (12 6548's + 1 Sup720).
Buy you're paying for 576 ports and only getting to use 1/6th of them!
For the space/power/cost of all that, it makes for a difficult decision.

dre



More information about the cisco-nsp mailing list