RE: [nsp] Gigabit Ether Channel on Cisco 12000s

From: Chris Whyte (cwhyte@microsoft.com)
Date: Thu Jul 18 2002 - 17:12:05 EDT


>
> On Thu, Jul 18, 2002 at 01:07:06PM -0700, Chris Whyte wrote:
> > It's actually a bit more than that. Remember, the
> interconnect from the
> > LC to the fabric is 5Gbps (including overhead). The
> overhead varies, of
> > course, due to cellification but you can actually get
> around 2.8Gbps to
> > possibly 3.0Gbps (subtracting overhead), if I remember correctly.
>
> You're using cisco math. I intentionally quoted all figures
> as full-duplex, because that tends to be a more useful number for
> capacity planning purposes. I probably should have clarified that
> in the original email.
>
> So:
>
> 5.0 Gbps (which is 2.5 Gbps full-duplex)

Btw, it's 5Gbps full-duplex, *not* 2.5Gbps full-duplex.

>
> A cisco cell is 64 bytes, 16 of which are header and CRC,
> and the remaining 48 bytes are payload. Probably works great for
> ATM (gah). So, 25% of our theoretical capacity is gone before we
> factor in any other overhead.

There's also encoding overhead.

>
> We're down to 3.75 Gbps, or 1.875 full-duplex. You can
> fill this thing without using the third port. And after real
> world overhead and other inefficiencies (CEF updates, etc.)
> I'd call this 1.5G full-duplex or so maximum.
>
> Note that in the installation and configuration guide
> for the 3GE-GBIC-SC, cisco clearly states:
>
> "The 3-Port Gigabit Ethernet Line card supports full line
> rate with two ports in service while the third port is shutdown.
> With three ports turned on, the 3-Port Gigabit Ethernet line card
> throughput is limited to the line card forwarding engine limit of
> 4 million packets per second (4 Mpps)."

The docs don't tell the whole story (as usual). :-)

>
> None of this changes my original assertion that trying
> to do some type of load balancing across multiple ports on this
> card can get you into trouble at high traffic volumes.

I would assert that having a more conservative capacity planning
guideline would be the desireable approach. ;-)

Thanks,

Chris

>
> --msa
>



This archive was generated by hypermail 2b29 : Sun Aug 04 2002 - 04:13:50 EDT