[c-nsp] PPS

[Saku Ytti]

On (2007-04-18 07:16 +0800), Lincoln Dale (ltd) wrote:

> actually, it really does "depend" on the platform. many platforms have
> 'distributed' forwarding - e.g. take a 6500 or 7600.
> you *can* have centralized forwarding (no DFCs installed) or you can
> have 'partially centralized' (DFCs on some linecards, none on others) or
> 'fully distributed' (DFCs on all linecards).

Yup, but we were talking about 3750. Of course you could argue, it's
not lookup speed for whole platform even in 3750 when it's stacked,
especially in 3750E where you have egress purge and local-switching
the combined lookup rate for stack could be higher than maximum
per stack-member.
Just as a curiosity non-DFC platform actually is not limited by
the lookup speed, but by the amount of packets you can send over
the shared-bus for lookup.
It's 62.5MHz, 32bytes per cycle, IP being 64bytes and MPLS 96bytes.

[ytti at ytti.fi ~]% echo "(62.5*1000000)/2/1000000"|bc -l
31.25000000000000000000
[ytti at ytti.fi ~]% echo "(62.5*1000000)/3/1000000"|bc -l
20.83333333333333333333

So theoretical max IP speed would be 31.25Mpps and MPLS speed 20.83Mpps on
CFC platform, but not because SUP PFC couldn't handle more, just you can't
send more packets for lookup towards it from LC's.
Number of course decreases further if you have legacy bus LC's, eg.
WS-X61xx, as they compete for the same resource.

> I'd like to add to your math is that 'wire rate' on 1GbE can actually be
> anywhere between 1487946 and 1488244 packets per second (+/- 149 PPS),
> since oscillators on 1GbE are allowed to be +/- 100PPM.

Thanks, this was news to me.

> if you were (say) generating traffic at 1488095 PPS, the chances are all
> you'll end up showing is ever-increasing latency since you end up
> showing queuing delay due to oscillator differences.  best to always
> test at 99.99% of wire-rate.

Also very interesting, thanks!.

-- 
  ++ytti