On Sat, Dec 2, 2017 at 9:59 AM, Dave Taht wrote: > On Sat, Dec 2, 2017 at 9:47 AM, George Amanakis wrote: >> Another run using the CMTS topology. >> ISP limited at 200/10mbit. >> Reference: pfifo limit 1000. >> Cake or HTB/{codel,pie,fq,sfq} limited at 180/9mbit. > > That's a little discouraging. That's sch_fq, not fq_codel? (please add fq_codel) Ah, OK, I get it. (please do add fq_codel. cake (all modes) with ack filtering. A 20ms result would be good too.) The ginormous default buffer size for sch_fq leads to seconds of tcp RTT, but the latency number for the measurement flows is low because it's fq. sfq amazingly is holding it's own (which it didn't used to), with it's mere 128 buffers at this RTT, it girates around 200+ms rtt for tcp.. I imagine that tcp pacing is helping it a lot here, even at these long rtts... And the surprising winner at the 98th percentile is... cake flowblind. There are major differences in the aqm algorithm between codel and cake... (all reasonable news.. along with my perpetual joy at "it doesn't crash") There are a few other puzzling things in this data-set. > >> RTT 100ms. >> rrul_be_nflows, 32 download, 8 upload streams. >> >> George > > > > -- > > Dave Täht > CEO, TekLibre, LLC > http://www.teklibre.com > Tel: 1-669-226-2619 -- Dave Täht CEO, TekLibre, LLC http://www.teklibre.com Tel: 1-669-226-2619