[e2e] TCP un-friendly congestion control

[Jon Crowcroft]

to clarify somethign i said before - if we deploy ecn or any aqm
driven loss feedback, then the queues are gonna stay short, so an rtt
based scheme wont work (at least not well) as the difference in delay
for loaded and unloaded queue will be only 1 or 2 packet's serialisation
time, so rtt reactive schemes (and all _implicit_ congestion signal
schemes) have a limited shelf life...imho

actually, david reed's comments about little's law have some rel. to
this too if you look at the ratio of access link to core link speed,
and the number of connections sharing a core congested v. an edge
congested router, then the information in the rtt or loss is massively
ambigious as we deploy more (but not huge numbers) of Gbps access
links....so i dont hold out hope for either vegas++ or VJ++ schemes in
the really long run - i'd rather build a world that has a self
sustaining/self consistent story for any approach, hence my rant in
favour of ecn...

 cheers

   jon