I love you
Sounds like you're doing pretty good on your own...
Possibly playing with the more advanced settings on Cake SQM, (read down the 3 layers of SQM docs on the main OpenWrt site) to squeeze out a bit more performance and fairness results. One thing you would probably benefit from would be the ack-filtering on egress, reducing some traffic load on that 1Mbit upload.
Thanks. After some sleuthing, I added option eqdisc_opts 'nat dual-srchost ack-filter'
but I am not sure it made a big difference in dslreports.com
. It was cool to learn why it is important on very asymmetric links.
One thing that I have been wondering is the link layer adaption per packet overhead, but that would require a good bout of measuring ping
times so I'm leaving it at 44
for the time being.
Yep... I'd add nat dual-dsthost ingress on the ingress side, to have the same kind of flow fairness behavior in both directions. I don't know if the ingress keyword is going to help much or not, but it's suggested.
Tuning the shaping speeds, of course, to find the maximums before the bloat starts to rise is the main thing, seeing where your sweet spot is for the link adaptation less so, but still worth it at some point.
You could use tc -s qdisc on the command line, to see a reporting of how Cake is configured and what it's doing, (and that you didn't do a typo in the optional lines, causing it not to set up an interface at all!) including how many acks are filtered.
Yup, that went in along with the egress options I mentioned.
I did check that while setting things up and several hours later, I can see the effect of ack-filter
:
Bulk Best Effort Voice
thresh 59368bit 950Kbit 237496bit
target 307ms 19.2ms 76.7ms
interval 614ms 114ms 172ms
pk_delay 0us 13.6ms 9.4ms
av_delay 0us 7.88ms 1.53ms
sp_delay 0us 1.24ms 95us
backlog 0b 1393b 0b
pkts 0 6259880 52208
bytes 0 677996170 7181345
[...]
ack_drop 0 1155234 0
[...]
It does not look bad to me (~18% of packets!), but it's still a little arcane.
Wow... much higher % rate than I've ever seen!! And, I guess you can say, the packets get more important (losing a packet when you have less total per second), the slower the rate. So, that would be a LOT of functional bandwidth you've gained back in your uplink!
fast.com and dslreports.com/speedtest do not notice a big increase in uplink bandwidth, but as you said, I may need to retune the threshold. I decided to try diffserv8
on both ingress and egress. Here are the egress stats:
Tin 0 Tin 1 Tin 2 Tin 3 Tin 4 Tin 5 Tin 6 Tin 7
thresh 950Kbit 831248bit 727336bit 636416bit 556864bit 487256bit 426344bit 373048bit
target 19.2ms 21.9ms 25ms 28.6ms 32.7ms 37.4ms 42.7ms 48.8ms
interval 114ms 117ms 120ms 124ms 128ms 132ms 138ms 144ms
pk_delay 0us 5.52s 30.9ms 0us 1.27ms 0us 82.7ms 7.33ms
av_delay 0us 550ms 3.12ms 0us 38us 0us 20ms 1.52ms
sp_delay 0us 46us 120us 0us 23us 0us 26us 18us
backlog 0b 0b 0b 0b 0b 0b 0b 0b
pkts 0 2020 4915228 0 56 0 22568 89207
bytes 0 652192 1201575283 0 5040 0 5392798 12372145
way_inds 0 0 150564 0 0 0 0 903
way_miss 0 6 259935 0 56 0 14 18295
way_cols 0 0 0 0 0 0 0 0
drops 0 242 169933 0 0 0 1396 0
marks 0 0 3605 0 0 0 0 0
ack_drop 0 961 1056227 0 0 0 10165 0
sp_flows 0 1 1 0 1 0 1 1
bk_flows 0 0 1 0 0 0 0 0
un_flows 0 0 0 0 0 0 0 0
max_len 0 25738 68970 0 90 0 36336 590
quantum 300 300 300 300 300 300 300 300
There is really a lot of ack_drop
s (~20% of pkts
).
Up to now I'm not unhappy.
Hmm, never tried diffserv8. Seems that it dosen't have class names so it's hard to know what they are. I usually have it set for diffserv4 or the default 3, and they have names for the tins. Looks like what would be the "best effort" tin is empty?
Overall, I'd guess that perhaps even if you don't see a big speed difference with the ack filtering on, more of the available pipeline is composed of YOUR data rather than acks and the time they take, so the responsiveness and speed that files get there is probably improved.
Yes indeed.
It's interesting to see 5 tins being used and I am definitely not doing anything on the DSCP marking front. Also, tried some video, online conferencing, etc and couldn't quite see a direct correlation while looking at these numbers.
For the time being, I'll let good enough be good enough
Hi,
Thanks so much for sharing your builds Catfriend. I installed your 21.02.0 stable build on my V5 yesterday and I'm going to give it a few weeks to assess stability. So far things seem really snappy and the wifi reach is great.
A note that I am installing this build because the stock 21.02.0 gave me video stuttering and drop outs on Google Meet calls both with wifi and hardwired.
I will let you know how it goes in a few weeks.
I just want to mention that my experience is the same. Random dropouts on OpenWrt 21.02.0. So far this build has been OK, will update again if dropouts continue. I'm scheduling a daily reboot, just to be safe.
This worked for me; I am running 6 Archer V2s and one Archer V4. The WiFi stability is great with the ath10k non-ct-drivers; I was having lots of drop outs with the ct-drivers and had resorted to using your scripts. Since switching to the non-ct-drivers, the scripts never run and I've stopped installing them.
Also, thanks for posting your imagebuild.sh script on GitHub. It got me started building my own images which makes upgrading a breeze with the non-ct-drivers.
Many thanks for sharing, just to add some additional bits to disabling ipv6 according to:
opkg remove ip6tables
opkg remove kmod-ip6tables
opkg remove kmod-nf-ipt6
opkg remove kmod-nf-reject6
opkg remove odhcp6c
opkg remove odhcpd-ipv6only
Poster confirms this has provided additional stability.
With Metta
To be honest, while removing ipv6 stuffs might increase the stability, 2.4GHz dies eventually.
I am testing the workaround of @sammo that remove 'ldpc'. 7 days running, hope it is the definite/real cause.
My assumption is that the issue is not related to ipv6 nor ath10k driver (ct or non-ct), but ath9k driver itself which enters in a waiting loop when a client station is far from the AP or the signal is weak and don't acknowledge some handshake. And a scan brings the driver out off that waiting loop.
The removal of ldpc in ath9k might help as error checking is different.
note: I own also a wdr7500 v3 modified 16MB flash fake C7V2.
Keep us posted... I haven't tried the ldpc disable yet, may try it later.
I'm working with someone who was involved with some of this code, and am trying to dig up some data for him. I have disabled my nightly cron of ifconfig wlan1 down and up... so that I can try troubleshooting the problem.
Annoyingly, it took nearly 2 weeks till I had an episode again! Then I had them on a much shorter basis. My point being, you might have to wait a few weeks till you have a "statistically correct" conclusion on something actually having an effect. Hang in there and let us know.
It does feel like there's a special interaction of traffic types, or special condition, that needs to occur to cause this. And, it's always difficult troubleshooting, when it's an intermittent, seemingly random problem.
Hi Jon, glad to hear someone is dealing with the code itself.
our concern is likely a race deadlock (atomic action?) in a multi-threaded condition.
In a multi-threaded real-time program, even a simple printk
may change behavior.
why a simple iw dev wlan1 scan
bring the driver out off deadlock? I think this should be the direction of your investigation.
Just checking in to let you know that Catfriend's 21.02.0 has been rock solid for the past 3 weeks. I'm on an Archer C7 V5. WIFI is working excellent and no problems with running SQM.
Note that I had to move to Catfriend's firmware after installing the regular 21.02.0. After a day with the regular firmware I would get drop outs and stuttering on Zoom calls and Google Meet. Both WAN and WIFI were affected.
I'm following this thread to jump to OpenWRT with my Archer C7 v2.
Is it stable with both band 2.4ghz and 5ghz ?
On c7v2 : 5ghz yes, 2.4 ghz should be ok most of the time but a weekly reboot won't hurt.
On c7v5: yes -yes
Hi! Just registered to the forum just to ask if your build is not affected by the known issue of v21.02? Thanks in advance
Known issues
Some IPv6 packets are dropped when software flow offloading is used: FS#3373
As a workaround, do not activate software flow offloading, it is deactivate by default.