For the record, I do not support this draft. This group insanity about ECN and ever more relaxed responses to it, is why I quit the IETF. I'm busy fixing bufferbloat everywhere I can, with running code, running on real linux and real bsd systems, and real results[0] - and the only thing I lose sleep over with that deployment[1] is the thought of y'all making an *already not a big enough reduction* for ecn response in modern tcps, even worse. [2] Having left the room, and having realized that few pay attention anyway to this standards body anymore, *please* feel free to pass it over my objection. I realize it's just an experimental draft. However, if there is a future version of this draft, please take my name off the "thanks" page. My suggestion, for an "experiment", would be to CE mark to .5, return linux cubic to .5, and in the presence of both loss and CE marks in an RTT, back off even harder than that, as you've completely overestimated the RTT usually at that point. And then test the hell out of it on the hundreds of flows/minute typically never exiting IW10 slow start on a your typical edge link, against a few long running tcp flows, not the other way 'round. I've had many other suggestions over the years as to valid tests - (many now in flent) against other forms of traffic, videoconferencing, voip, etc, loads, against new forms of DDOS made possible by ECN, etc, etc, all ignored by those afflicted by the ECN madness. The *deployed* aqms I'm most familiar with, fq_codel, and to some extent pie, in the crazy overbuffered[3] world we live in, respond much better to rfc3168-style behavior, in this, typical environment, where we have to inbound shape 680ms of inherent delay in your typical CMTS down to something sane. I have no idea why you need an experimental draft to do this experiment in the first place. Just do the experiment. On your own network, not mine, please. [0] https://github.com/pfsense/pfsense/pull/3941 [1] https://github.com/systemd/systemd/issues/9748 [2] We already, in the sqm-scripts reference distribution, disable ecn universally for outbound, by default, for fq_codel. Recently, we enabled it universally for sch_cake, and as an experiment that's not going particularly well. Please feel free to try these aqms in your experments as they are widely available in openwrt and the current linux kernel. [3] http://www.dslreports.com/speedtest/results/bufferbloat?up=1 PS Better yet, redteam ECN. Simplest experiment, one that blows up many ecn assumptions, is just hit your aqm of choice with ECT(0), ECT(1), and CE floods across 1000 udp ports. It's a really hostile internet out there... -- Dave Täht CEO, TekLibre, LLC http://www.teklibre.com Tel: 1-669-226-2619
