Re: shaping: dividing bandwidth between router & NAT hosts

Stephan Balmer wrote at 2010-02-13 04:52 -0600:
> > So, Stephan must have been meant ingress shaping (delaying packets?) and I
> > assumed he meant ingress policing. Is there a significant reason to use
> > shaping rather than policing? Yes, policing drops valid packets, but TCP will
> > cause that anyway before backing off sending.
>
> Sorry for the confusion. Yes, I meant shaping. Maybe policing is better than
> nothing, definitely try it out and tell us :-)

Yeah, I plan to try out policing; I'll do some testing sometime hopefully and
see if I notice any difference.

> But there is a reason to buffer packets instead of dropping them:
> Dropping packets kills TCP throughput. Buffering packets gives TCP connections
> a way to figure out the appropriate rate. Don't ask me to explain because I
> don't understand much about it myself. Ask Google about TCP flow rate and
> congestion control to get an idea.
>
> If you use a hash bucket queue or similar for buffering, each connection
> basically gets its own buffer, so that trickling traffic like SSH does not
> get delayed in long buffers. And real-time traffic like phone calls prefer
> 10% dropped packets over 100% delayed packets so they need their own
> mini-queues.

Okay, that makes sense.