[Bloat] Researchers discover major roadblock in alleviating network congestion

[Jonathan Morton]

> On 5 Aug, 2022, at 2:46 am, Daniel Sterling <sterling.daniel at gmail.com> wrote:
> 
> "Flow control power is non-decentralizable" is from -- 1981? So we've
> known for 40 years that TCP streams won't play nicely with each other
> unless you shape them at the slower endpoint-- am I understanding that
> correctly? But we keep trying anyway? :)

More precisely, what was stated in 1981 was:

The specific metric of "network power" (the ratio of throughput to delay, calculated for each flow and globally summed) cannot reliably be maximised solely by the action of individual endpoints, without information from within the network itself.

Current TCPs generally converge not to maximise or even equalise "network power", but to equalise between flows a completely different metric called "RTT fairness", the *product* of throughput and delay.  Adding information from the network via AQMs allows for reductions in delay with little effect on throughput, and thus a general increase in network power, but the theoretical global optimum is still not even approached.

Adding FQ in the network, thus implementing "max-min fairness" instead of "RTT fairness", hence equalising throughput instead of the product of throughput and delay.  This is essentially the geometric mean of RTT-fairness and network power.

I believe it is actually possible to achieve equalisation of network power between flows, which would approach the global optimum of network power, using information from the network to guide endpoint behaviour.  This is *only* possible using explicit information from the network, however, and is not directly compatible with the current congestion-control paradigm of RTT-fairness by default.

 - Jonathan Morton