All, ran lots of performance tests with GL-MT6000 from my docsis 3.1 cable modem 1Gbit service -
- Raw: 958/37 Mbits - Load: 45/20/2/19%, +18ms, +31ms, Bufferbloat: B rating
- SFO: No change
- SFO+HFO: 960/37 Mbits - Load%: 0/0.7/0/0, +18ms, +32ms, Bufferbloat: B rating
- SFO+HFO+SQM: No change just confirming SFO+HFO needs to be disabled for SQM
- SQM cake/simplest_tbf: 450/30 Mbits: Load%: 40/10/10/10, +0ms, +0ms, A+ rating
- SQM fq_codel: 800/30 Mbits: Load%: 11/37/38/4, +2ms, +0ms, A+ rating
Yes this target will do 800Mbits +0/+0 with fq_codel, but is inconsistent. Often +0 would become +7ms then drop back down. The algorithm is just not as good as Cake.
Interestingly my WRT32X got perfectly consistent +0/+0 at 600Mbits with cake, but with CPU0 at 99% leaving nothing left.
Overall SQM seems inconsistent on the MT6000 (and assuming Filogic as a whole). For now with speeds at or near gigabit I think the HFO + WED combo (which does not work with SQM) is the way to go. Maybe it'll improve with kernel 6.1 I'll try that build soon.
edit: as a bonus here are my Samba test results from that thread: [Samba 4.x package support thread - #547 by phinn]