C3750G-24TS forwarding rate?

I was looking at the performance specs of the Cat 3750 series
multilayer switches just now, trying to figure out how close to
wire speed they get.

Looking at Table 2 - Hardware at
http://www.cisco.com/en/US/products...ducts_data_sheet09186a0080161371.html#wp39660

The C3750G-12S and 3750G-24T appear to be spec'd at full wire rate.
They are 17.8 Mpps and 35.7 Mpps respectively, and
35.7 M packets/second * (64+20) bytes/packet * 8 bits/byte / 24 ports
which is 1000 M bits/port to within roundoff error

When I looked at the C3750G-24TS, the 38.7 Mpps at first appeared
to be too high -- the rate for 24 ports came out as 1083.6 M bits/port.
{Truth to tell, I started with this calculation and got myself royally
confused until I did the -24T calculation and managed to get all my
operators straight.}

What I finally realized was that the -24TS is not 24 ports, but rather
28 ports, as it has 4 SPF ports as well as the 24 x 10/100/1000. If one
calculates the bit rate and divides by 28 ports, then the answer comes
out as 928.8 Mbit/s on each port.

Oddly, though, if one calculate the total bit rate (not per port), it is
38.7 M packets/second * (64+20) bytes/packet * 8 bits/byte
which is 26006.4 Mbits/s total. 6.4 Mbits/s is within roundoff error,
so if one looks at this number, it appears that the C3750G-24TS
is designed to run exactly 26 ports at wire speed, not 28 ports.

The figure is -too- close to wire rate for 26 ports for it to be a
coincidence. One could imagine that the switch can't quite keep up
to full wire rate at 28 ports, but if that were the case, one wouldn't
expect it to lag by an average across all the ports that just *happened*
to be equivilent to two wire-speed ports.

This leads me to wonder whether 2 of the SPF ports overlap two of the
10/100/1000 ports, for a total of 26 simultaneously usable ports [a
design Nortel seems fond of, such as on the Baystack 470-48T or
Baystack 5510-48T]; or whether there are pairs of SPF ports that share
controlling chips and so cannot be simultaneously run at wire rate; or
whether the SPF ports on the 3750G-24TS operate at half speed even
though the ones on the 3750G-12S run at full speed; or ... or if Cisco
published incorrect specs? Was someone at Cisco perhaps confused by the
C3750-24TS (notice no 'G'), which has 24 x 10/100 ports and *two* SPF
ports instead of the *four* present in the C3750G-24TS ?

I've looked through the release notes and installation guides,
and have not found any evidence that any of the ports overlap
any others.

Cisco has gone the shared-ASIC route before, such as on the
C2948G-L3 (4 ports per controller), or the multi-GBIC cards for the
Cat5000. And considering the price of the 3750G-12S, it could be that
they put in 12 ASICs to serve the 12 SPF ports at full speed instead
of putting in 6 2-port ASICs to server 12 SPF ports...

Is there anyone out there who can offer some insight into this?

 

-cnrc.gc.ca (Walter Roberson) wrote in
comp.dcom.sys.cisco:

Just a quick note:

shouldn't you divide by twelve?
I always thought the maximum throughput is when half the ports forward to
the other half. I think that if you put a host on each port, throwing
traffic ad the switch, that's routing everything to null0 doesn't count as
'real world traffic' don't you?

 

:-cnrc.gc.ca (Walter Roberson) wrote in
:comp.dcom.sys.cisco:

:> The C3750G-12S and 3750G-24T appear to be spec'd at full wire rate.
:> They are 17.8 Mpps and 35.7 Mpps respectively, and
:> 35.7 M packets/second * (64+20) bytes/packet * 8 bits/byte / 24 ports
:> which is 1000 M bits/port to within roundoff error

:Just a quick note:

:shouldn't you divide by twelve?
:I always thought the maximum throughput is when half the ports forward to
:the other half.

No, divide by the full number of ports is correct. The ports run
at full duplex, so one could have the data that goes into port 1
going out of port 2, the data going into port 2 going out port 3,
the data going in to port 3 going out port four... the data going
into port 24 going out of port 1.

The half-forwarding-to-half scenario is for half-duplex.