On +3:

I was able to get rrul to sort of fail on the ath10k also in the same
way. no kernel messages, but it had trouble establishing the control
channel after rrul ran, on rrul_be.

it could be osx going to hell, I can only get it to happen after
pounding it flat with a dozen other tests first. If I just run it
after a period of idle, can't get it to fail.

The ath9k test on the same hardware I just did of your +3 completed
all tests (except for rrul, same as above) and throughput was a bit
better than the last series I'd done (I switched channels to one less
busy at this hour of the day).

No kernel messages.

There is an interesting pattern emerging relative to the number of
flows (see attached), that I guess I'll have to go check against
linux<->linux's tcp... I've always kind of wondered about block acks
vs del tcp acks.... we're not out of cpu at this lower rate, so it
shouldn't be fq_codel's cpu usage at fault...

The quantum is writable?

Cake works, too. While I won't certify this new hardware as ready for
wide testing by others, yet, it really does look quite good for the
basic stuff we care about.

I will fire up a long series of 300 tests on this, move back to the
archer, and see if I can sort through the x86 patchset... and get in a
swim.

Happy. thx. a nice labor day....