Re: [aqm] Question re draft-baker-aqm-recommendations recomendation #2

["Fred Baker (fred)" <[e-mail]>]

On Apr 29, 2013, at 3:28 PM, Linda Dunbar <[e-mail]> wrote:

> 
> 
>>   I don't follow your point here. There are cases where ECN marking
>> may not work very well, but I don't see any correlation to the size
>> of the network. (ECN signals reach a receiving endpoint unless
>> something
>> is broken.)
>> 
> 
> 
> [Linda] Congestion can be transient.

>From my perspective, network congestion is almost always transient; non-transient congestion is an outage or should be. Operators tend to measure on time scales of 300 ms, for reasons related to the tool (MRTG) commonly used to measure it. But congestion on the time scale of a TCP session is measured in RTTs, which are on the order of milliseconds to hundreds of milliseconds.

The direction you are arguing suggests that there is no value in TCP Congestion Control, for the simple reason that in the next RTT a given TCP may have nothing more to send. I hope that's not what you mean, because that doesn't square very well with measurable reality. It is true that sessions are predominantly short, and short sessions add a lot of noise to any measurement. However, the actual traffic is predominantly from longer sessions, and absent congestion control in one form or another would drive the network into congestive collapse.

If you believe that there is value in congestion control, the question on the table is whether there is value in explicit vs implicit signaling.

> For a network with many hops, the congestion might not exist anymore by the time the Source nodes receive the notification from the end points. 
> When many source nodes throttle back traffic (or control ingress rate) for transient congestion, network oscillation can occur, which is not good. 

Hops are irrelevant, in my view. What is relevant is the duration of sessions and of RTTs. I can show you the same behavior in many-hop and few-hop networks.

> Network can perform much better if there are two different types of congestion marking, one for transient congestion, and another one for congestion caused by "hard degradation" (e.g. LAG link member failure, or microwave link bandwidth being reduced due to weather, etc). 

That might be ECN and loss?