[c-nsp] Cisco 3550 counters and QoS

[Boyan Krosnov]

> Also, last time I checked the overhead of dot1q tagging was 
> way lover than
> 20%! ;)
True, it is only 4 bytes per frame. 0.2% to 5% depending on frame size
is more likely.
Also ethernet frame size vs. IP packet size is about 18 bytes (can be
more if you count the FCS and preamble). This is 1% to 25% depending on
packet size.

Have you noticed that your policers are actually _designed_ to drop
traffic and this means that it is normal to observe some differences on
the input and output interface. With policing I've seen drop-rates as
high as 50% with some severely overutilized policers. You can help it a
bit by increasing the bucket size, but you will get nowhere close to
shaper performance.

Cheers,
Boyan Krosnov
just another techie speaking for himself