BitTorrent peers are collorative, because otherwise turning on BitTorrent anywhere would immediately result in the parent's boogeyman showing up: being flooded by way more traffic than can be handled. That's not an overlooked part of designing a client: it's known that when writing UDP protocols one has to do one's own flow control.
BT would be a lot friendlier for traffic shaping if the tracker played a more active role in flow control, or if otherwise some distributed consensus was negotiated regarding wanting more unsolicited connections.
The first paper we did was: http://perso.telecom-paristech.fr/~drossi/paper/rossi13tma-b...
Which predated the development of fq_codel, which is the fair queue (flow queue) + aqm hybrid now deployed in cerowrt, openwrt, and countless other QoS systems. That paper only explored RED or SFQ, not a hybrid.
What I think the positive effects we are seeing now are several factors. 1) a "torrent" is usually on 6 flows at a time, rotating to a new flow every 15 seconds or so. 2) The additional peers trying to negotiate IS a problem at very low rates (say below 4mbit) but invisible at higher rates. 3) reducing latencies under load to under 5ms has enormous gains in bidirectional throughput, which more than compensate for losing torrent's delay based backoff mechanism, which only backs off at 100ms. Which would you rather have, 100ms latency or 5ms?
See, for example, what the fq_codel system does for verizon and comcast here:
https://www.bufferbloat.net/projects/codel/wiki/RRUL_Rogues_...
We got back 5x free bandwidth on the verizon test. And if you are experiencing 5ms worth of latency does it really matter if you have 6 torrents in the background?
(80, yes! but it's a function of your bandwidth also and... and... more testing is indicated.)
... so I haven't got around to redoing the research and writing a successor paper on it, and probably won't anytime soon. The original authors of that paper have gone off to do several papers of extensive analyses of RED vs ledbat, and I think they've largely gone off into the weeds, not that I mind having that analysis as a basis if ever we (or someone) gets around to analyzing fq_codel with some of those techniques.
Please feel free to test re torrent with a system that uses something like openwrt's qos-scripts or cerowrt's SQM system.
The conclusion of the original paper was that only classification seemed to be an answer to even further deprioritize torrent in an AQM'd and FQ'd world. (which the above systems can do also)
http://tools.ietf.org/html/draft-hoeiland-joergensen-aqm-fq-...
E.g.: The router has some sort of bandwidthd. The Torrent app checks in, asks how much it can use for a low-priority bulk download, and gets a token good for a 30-second lease on 10 megabits down. Before that expires, it tries to renew the token, and maybe it gets only 5 megabits because somebody is watching Netflix. If at some point the connection quality gets bad, the router's owner gets an email saying that the router has noticed a lot of BitTorrent traffic from "Timmy's MacBook" and that maybe they should switch to a BT client that supports bandwidthd. And so dad gets Timmy to upgrade.