[aqm] ingress shaping

[Mikael Abrahamsson <[e-mail]>]

Hello.

I'd just like to bring attention to another deployment scenario where 
ingress queuing might make a lot of sense:

ISP deploys ETTH on a 100/100 port (100 megabit/s ethernet full duplex). 
Service is 100/10 (100 megabit/s towards the customer (downstream), 10 
megabit/s customer->Internet (upstream)).

ISP solves the traffic problem by putting in a policer upstream that has a 
2 second burst size, and polices at 12 megabit/s. This means TCP will be 
sawtoothing, but it will pass the "speedtest" websites, because effective 
throughput will be ~10 megabit/s upstream.

Now, I feel the customer is better served if there was some 
shaping/buffering instead of this policer (policer is defined as something 
that has no buffer but instead just drops packet that is over bytes/s 
averaged over a certain time). Since the first real buffering point in 
this solution is the customer CPE, to achieve buffering we need to do this 
ingress towards the CPE.

Downstream device layout:

Internet -> ISP -> DR -> AS1 -> AS2 -> AS3 -> CPE

AS = Access switch
DR = Distribution Router
CPE=Customer Premice Equipment

Policing is done egress on AS3. To smoothe out the flows, CPE needs to do 
ingress shaping to ~10 megabit/s.

Would it make sense to apply fq_codel with an artificial bw shaper ingress 
on the CPE? Since the fq_codel would limit queuing to ~5ms and smoothe out 
flows, I don't see how the 12 megabit/s policer at AS3 egress would even 
be hit (unless it's very bursty flows)?

If there is any AQM document that one can do procurement decisions 
against, I'd like to have this deployment scenario included anyway, so 
that device vendors include possibility of ingress shaping as well.

-- 
Mikael Abrahamsson    email: [e-mail]