as of linux 4.12, we’d put in some major enhancements to linux wifi, originally for the ath9k chip, nowadays the ath10k, most of mediatek’s mt76 line, and most of intel’s chipsets, making fq_codel run native (near zero cpu cost), and solving the packet aggregation and airtime fairness problems thoroughly. The relevant paper and test suite is here: https://www.cs.kau.se/tohojo/airtime-fairness/ - 2.5 the capacity at 1/10th the latency on some tests.
My question is - does mikrotik use the factory wifi driver, still? Are there options for ATF, now, at least?
A dream for me, has been for more folk to understand how important getting packet aggregation and ATF were to scaling 802.11ac and later for multiple stations. In 8 minutes, here, I’d explained what everyone had been getting wrong with wifi: https://www.youtube.com/watch?v=Rb-UnHDw02o&t=1558s
Mikrotik uses their own drivers. This is great in some ways like uniform behaviour across chipsets when configuring them and also CAPSMAN (which has to make some internal abstractions so you can do things like local and non-local forwarding) but it means they can be behind in chipset features like WAVE2, WPA3 and spectrum scanning on AC devices.
I have never explored this but you can change the queue type on any interface. Have a look at the interface queues and you can see the default for wireless is an SFQ type but you can change it to any defined queue type you like including CAKE and Codel. Whether this has any effect is open to discussion. The help entry doesn’t expand on it at all.
Tested using default SFQ vs default FQ_Codel queue type on wireless interface for hAP ax2. Couldn’t see any performance benefit whatsoever in iPerf3/latency measurements during peak upload/download.
single threaded like iperf tests, without also monitoring tcp latency (Try a packet capture in wireshark, do a RTT plot) will not, in general, show the underlying benefits of fq_codel or cake. You are measuring a dragster on a single track, with a plain iperf test. What of the ability to steer?
The real world consists of dozens or hundreds of flows, competing, across many machines. If you test for that, than you will see the benefits. One favorite test of mine is the rrul test.
