| Message ID | CAA93jw6NJ2cmLmMauz0xAgC2MGbBq6n0ZiZzAdkK0u4b+O2yXg@mail.gmail.com (mailing list archive) |
|---|---|
| State | RFC |
| Delegated to: | Johannes Berg |
| Headers | show |
| Series | a nuking the mac80211 changing codel parameters patch | expand |
I cannot help but wonder what the results have been on people applying this patch over the past 6 months? On Tue, Dec 20, 2022 at 12:24 PM Dave Taht <dave.taht@gmail.com> wrote: > > This is the single, most buggy, piece of code in "my" portion of wifi > today. It is so wrong, yet thus far I cannot get it out of linux or > find an acceptable substitute. It makes it hard to sleep at night > knowing this code has been so wrong... and now in millions , maybe > even 10s of millions, of devices by now.... Since I've been ranting > about the wrongness of this for years, I keep hoping that we can > excise it, especially for wifi6 devices and even more especially on > 6ghz spectrum... but just about everything, somehow, would benefit > hugely if we could somehow do more of the right thing here. > > I'd tried, last time I got this bee in my bonnet, tried to nuke this call here: > > https://forum.openwrt.org/t/reducing-multiplexing-latencies-still-further-in-wifi/133605/ > > As it is, I really encourage folk, especially with mt79 and to some > extent mt76 ac or ath10k, to try out the attached patch, measure tcp > rtts, and throughput, etc. A slightly less aggressive patch might suit > wifi-n.... > > Maybe there's a reason for keeping this code in linux wifi that I do > not understand. But here are my pithy comments as to why this part of > mac80211 is so wrong... > > static void sta_update_codel_params(struct sta_info *sta, u32 thr) > { > - if (thr && thr < STA_SLOW_THRESHOLD * sta->local->num_sta) { > > 1) sta->local->num_sta is the number of associated, rather than > active, stations. "Active" stations in the last 50ms or so, might have > been a better thing to use, but as most people have far more than that > associated, we end up with really lousy codel parameters, all the > time. Mistake numero uno! > > 2) The STA_SLOW_THRESHOLD was completely arbitrary in 2016. > > - sta->cparams.target = MS2TIME(50); > > This, by itself, was probably not too bad. 30ms might have been > better, at the time, when we were battling powersave etc, but 20ms was > enough, > really, to cover most scenarios, even where we had low rate 2Ghz > multicast to cope with. Even then, codel has a hard time finding any > sane drop > rate at all, with a target this high. > > - sta->cparams.interval = MS2TIME(300); > > But this was horrible, a total mistake, that is leading to codel being > completely ineffective in almost any scenario on clients or APS. > 100ms, even 80ms, here, would be vastly better than this insanity. I'm > seeing 5+seconds of delay accumulated in a bunch of otherwise happily > fq-ing APs.... > > 100ms of observed jitter during a flow is enough. Certainly (in 2016) > there were interactions with powersave that I did not understand, and > still don't, but > if you are transmitting in the first place, powersave shouldn't be a > problemmmm..... > > - sta->cparams.ecn = false; > > At the time we were pretty nervous about ecn, I'm kind of sanguine > about it now, and reliably indicating ecn seems better than turning it > off for > any reason. > > - } else { > - sta->cparams.target = MS2TIME(20); > - sta->cparams.interval = MS2TIME(100); > - sta->cparams.ecn = true; > - } > > And if we aint gonna fiddle with these, we don't need these either. > > In production, on p2p wireless, I've had 8ms and 80ms for target and > interval for years now, and it works great. It is obviously too low, > for those that > prize bandwidth over latency (I personally would prefer TXOPs shrink > intelligently as well as bandwidth, as you add stations, some of which > happens naturally by fq-codels scheduling mechanisms, others don't, I > even run with 2ms txops by default on everything myself) > > + return; > > Ideally we could kill this entire call off entirely. > > } > > A pre-thx for anyone actually trying the attached patch and reporting > back on any results. > > https://forum.openwrt.org/t/reducing-multiplexing-latencies-still-further-in-wifi/133605/ > > > -- > This song goes out to all the folk that thought Stadia would work: > https://www.linkedin.com/posts/dtaht_the-mushroom-song-activity-6981366665607352320-FXtz > Dave Täht CEO, TekLibre, LLC
diff --git a/net/mac80211/sta_info.c b/net/mac80211/sta_info.c index 97d24e9..457209a 100644 --- a/net/mac80211/sta_info.c +++ b/net/mac80211/sta_info.c @@ -2766,15 +2766,7 @@ unsigned long ieee80211_sta_last_active(struct sta_info *sta) static void sta_update_codel_params(struct sta_info *sta, u32 thr) { - if (thr && thr < STA_SLOW_THRESHOLD * sta->local->num_sta) { - sta->cparams.target = MS2TIME(50); - sta->cparams.interval = MS2TIME(300); - sta->cparams.ecn = false; - } else { - sta->cparams.target = MS2TIME(20); - sta->cparams.interval = MS2TIME(100); - sta->cparams.ecn = true; - } + return; } void ieee80211_sta_set_expected_throughput(struct ieee80211_sta *pubsta,
This is the single, most buggy, piece of code in "my" portion of wifi today. It is so wrong, yet thus far I cannot get it out of linux or find an acceptable substitute. It makes it hard to sleep at night knowing this code has been so wrong... and now in millions , maybe even 10s of millions, of devices by now.... Since I've been ranting about the wrongness of this for years, I keep hoping that we can excise it, especially for wifi6 devices and even more especially on 6ghz spectrum... but just about everything, somehow, would benefit hugely if we could somehow do more of the right thing here. I'd tried, last time I got this bee in my bonnet, tried to nuke this call here: https://forum.openwrt.org/t/reducing-multiplexing-latencies-still-further-in-wifi/133605/ As it is, I really encourage folk, especially with mt79 and to some extent mt76 ac or ath10k, to try out the attached patch, measure tcp rtts, and throughput, etc. A slightly less aggressive patch might suit wifi-n.... Maybe there's a reason for keeping this code in linux wifi that I do not understand. But here are my pithy comments as to why this part of mac80211 is so wrong... static void sta_update_codel_params(struct sta_info *sta, u32 thr) { - if (thr && thr < STA_SLOW_THRESHOLD * sta->local->num_sta) { 1) sta->local->num_sta is the number of associated, rather than active, stations. "Active" stations in the last 50ms or so, might have been a better thing to use, but as most people have far more than that associated, we end up with really lousy codel parameters, all the time. Mistake numero uno! 2) The STA_SLOW_THRESHOLD was completely arbitrary in 2016. - sta->cparams.target = MS2TIME(50); This, by itself, was probably not too bad. 30ms might have been better, at the time, when we were battling powersave etc, but 20ms was enough, really, to cover most scenarios, even where we had low rate 2Ghz multicast to cope with. Even then, codel has a hard time finding any sane drop rate at all, with a target this high. - sta->cparams.interval = MS2TIME(300); But this was horrible, a total mistake, that is leading to codel being completely ineffective in almost any scenario on clients or APS. 100ms, even 80ms, here, would be vastly better than this insanity. I'm seeing 5+seconds of delay accumulated in a bunch of otherwise happily fq-ing APs.... 100ms of observed jitter during a flow is enough. Certainly (in 2016) there were interactions with powersave that I did not understand, and still don't, but if you are transmitting in the first place, powersave shouldn't be a problemmmm..... - sta->cparams.ecn = false; At the time we were pretty nervous about ecn, I'm kind of sanguine about it now, and reliably indicating ecn seems better than turning it off for any reason. - } else { - sta->cparams.target = MS2TIME(20); - sta->cparams.interval = MS2TIME(100); - sta->cparams.ecn = true; - } And if we aint gonna fiddle with these, we don't need these either. In production, on p2p wireless, I've had 8ms and 80ms for target and interval for years now, and it works great. It is obviously too low, for those that prize bandwidth over latency (I personally would prefer TXOPs shrink intelligently as well as bandwidth, as you add stations, some of which happens naturally by fq-codels scheduling mechanisms, others don't, I even run with 2ms txops by default on everything myself) + return; Ideally we could kill this entire call off entirely. } A pre-thx for anyone actually trying the attached patch and reporting back on any results. https://forum.openwrt.org/t/reducing-multiplexing-latencies-still-further-in-wifi/133605/