On 28/05/15 08:00, Steven Barth wrote:
> Hi everyone,
>
> again a bit of a basic question, but what are the advantages of doing
> ingress shaping in SQM?
>
> To me it wastes a lot of CPU cycles (decreases forwarding performance)
> and you can't really "unsend" any packets from the ISP. What I mean is
> in 99% of cases your internal forwarding capacity is usually (much?)
> bigger than what the ISP sends to at any rate.
>
> What do I miss here? Some effects on TCP rate-limiting?
>
>
As Dave has already said so well, *if* the ISP did sensible 
shaping/limiting then I totally agree with you it's a waste of cycles.  
Unfortunately it's a big if and I've seen some truly horrible behaviour, 
especially on slow links where I think fair q'ing and latency control 
are actually more important.

Your point about 'unsending' packets is well made though, and since the 
packet has made it this far and actually got to us it seems a shame to 
shoot it.  ECN would appear to be the best of both worlds, mark the 
packet/flow that in an ideal world wouldn't have got through and so 
signalling the other end to back off.

Fitting smaller *managed* pipes here & there is counterintuitive but it 
does help at the cost of bandwidth, something I'm more than prepared to 
put up with for the improved latency control.  But ideally fq_codel 
really, really needs to be implemented by the ISP.