[c-nsp] ASIC to switch port mapping

[Nick Hilliard]

On 10/09/2010 22:47, Vincent Aniello wrote:
> Any recommendations on a switch with larger buffers?  I would like to
> stick with the 1U form factor.  
> 
> Also, are you saying that the microbursts cause the switch to exceed 1Gb
> port speed which causes the drops?  Cisco claims that the
> WS-C3560E-24TD-E switch has a 128-Gbps wire rate non-blocking switching
> fabric capacity.  I may be oversimplifying this, but to me, Cisco's
> claim implies that I should be able to run every port on this switch
> simultaneously at a full 1Gb in both directions.

Yes, probably you can inject a constant 2,976,190 x 64 byte frames per
second on all ports and it won't lose any packets.  Thing is, that's not a
very interesting measurement as far as switch performance is concerned,
because normal traffic is very bursty and doesn't behave like lab traffic
injected on a traffic generator.  What's interesting is what happens when
you have line rate going on on a single port, and then you receive frames
on other ports which are destined to the full port.

>From a simplistic point of view, when this happens on a store-n-forward
switch like a 3560, you need buffers to store the incoming packets -
because the switch cannot send them to the destination port which is full.
 If for some reason you run out of buffer space on the input ports, the
switch will drop incoming packets.

Of course, there are several ways of mitigating against this sort of
problem (virtual queues, QoS etc), but they only get you so far.

Cisco are notoriously cagey about exactly how much rx/tx buffer space is
available on the 3560/3750 series boxes and how it's divvied up, but there
is lots of speculation about this on cisco-nsp.  Here's some:

http://puck.nether.net/pipermail/cisco-nsp/2010-March/068810.html

Nick