A nine-patch ath11k proposal makes existing mac80211 queue controls effective for the driver. The striking latency result comes from tightening those controls on one access point. Applying the series with its defaults produces essentially the same reported latency.
Read the three configurations together
Julius Bairaktaris’s August 24 v3 cover letter reports an IPQ8074 access point, one HE station, saturated downstream TCP and a 20 Hz latency probe in the same access category. These are the patch author’s measurements.
| Configuration | TCP throughput | Reported loaded latency |
|---|---|---|
| Stock driver | 94.6 Mbit/s | 155 ms |
| Series, default limit | 94.0 Mbit/s | 157 ms |
| Series, best-effort limit 500/1000 µs | 74.3 Mbit/s | 22 ms |
Calculated from these rows, 155/22 is about 7.05, while throughput falls by about 21.5% relative to stock. The default row does not show the advertised latency gain. The tighter values are airtime limits in microseconds, not byte counts or a new Wi-Fi link rate.
Why the queue's location matters
The proposal adds the accounting and scheduling needed to keep more waiting work under mac80211 control instead of letting it accumulate in hardware transmit rings. The driver documentation identifies ath11k as a mac80211-based Qualcomm 802.11ax driver; support for that family does not make every board equivalent to this test.
A useful analogy is a dispatch desk feeding a conveyor. If every package has already entered the conveyor, the desk cannot meaningfully reprioritize an urgent package behind them. Holding some work at the desk preserves scheduling choices. Making that holding limit too restrictive can also leave the conveyor waiting. This explains the direction of the tradeoff without pretending that packets have identical transmission times.
Measure the experience under load
For a comparison on your own equipment, record the exact driver, firmware, radio conditions, client and queue settings. Measure latency during the transfer and retain throughput alongside it. A quiet-link ping and a separate maximum-speed run do not characterize the same condition as simultaneous traffic.
Compare repeated runs with the same offered load and client position. A successful bulk transfer can coexist with poor responsiveness; conversely, a lower latency number can be purchased by moving less data. Choose according to the actual workload rather than maximizing one column.
The linked document is a patch proposal, not a release notice. It does not establish that a distribution kernel includes the series. No router was flashed or tuned for this article, and the author’s configuration should not be treated as a universal access-point preset.
Use v3 measurements with throughput tradeoff, airtime units and unchanged defaults; do not infer a released kernel from a patch proposal.