Dynalink DL-WRX36 Askey RT5010W IPQ8072A technical discussion

sampson · 2023-02-24T15:24:25.474Z

Will try your script next.

I just tested this:

main router: 1Gb to Internet . 2.5Gb port, part of br-lan. 192.168.8.1/24
2nd router: both WAN and WAN6 remoted. All 5 ports part of br-lan. Static 192.168.8.254/24

Internet <WAN 1Gb> main <br-lan 2.5Gb br-lan> 2nd

Test 1:
main: iperf3 -s
2nd: iperf3 -c 192.168.8.1
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.00 sec 2.42 GBytes 2.08 Gbits/sec 16 sender
[ 5] 0.00-10.00 sec 2.42 GBytes 2.08 Gbits/sec receiver

Test 2:
2nd: iperf3 -s
main: iperf3 -c 192.168.8.254
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.01 sec 1.97 GBytes 1.69 Gbits/sec 4 sender
[ 5] 0.00-10.01 sec 1.97 GBytes 1.69 Gbits/sec receiver

Why the result is not the same?

I applied the same set of local start to both main & 2nd:

#assign 4 rx interrupts to each cores
echo 8 > /proc/irq/50/smp_affinity
echo 4 > /proc/irq/51/smp_affinity
echo 2 > /proc/irq/52/smp_affinity
echo 1 > /proc/irq/53/smp_affinity

#assign 3 tcl completions to 3 CPUs
echo 4 > /proc/irq/73/smp_affinity
echo 2 > /proc/irq/74/smp_affinity
echo 1 > /proc/irq/75/smp_affinity

echo performance > /sys/devices/system/cpu/cpufreq/policy0/scaling_governor

sqrwv · 2023-02-24T15:36:02.508Z

So now it's LAN-LAN speed test.

Could you repeat the test several times and see if result changes?

In some test I noticed sometimes Packet Steering doesn't work well, sometimes it distributes load across cores, other times it doesn't. You can verify if is this is the case looking at htop while you do the tests. At least that was my experience in those tests with AX3600.

You need to look at htop while running the speed tests, you'll see the bottleneck is one core being at 100%. You can also identify what processes do the load.

sampson · 2023-02-24T15:36:19.591Z

main runs iperf -s, 2nd runs iperf -c 192.168.8.1
gro - main: on, 2nd: on
2.06 Gbits/sec

gro - main: on, 2nd: off
1.95 Gbits/sec

gro - main: off, 2nd: on
1.14 Gbits/sec

gro - main: off, 2nd: off
1.04 Gbits/sec

I got the follow output when running ./gro.sh on or off

./gro.sh off
 br-lan: off
   lan1: off
   lan2: off
   lan3: off
   lan4: off
     lo: off
 miireg: off
phy0-ap0: Cannot get device feature names: No such device
Cannot get device feature names: No such device
phy1-ap0: Cannot get device feature names: No such device
Cannot get device feature names: No such device
phy1-ap1: Cannot get device feature names: No such device
Cannot get device feature names: No such device
phy1-ap2: Cannot get device feature names: No such device
Cannot get device feature names: No such device
    wan: off

./gro.sh on
 br-lan:  on
   lan1:  on
   lan2:  on
   lan3:  on
   lan4:  on
     lo:  on
 miireg:  on
phy0-ap0:  on
phy1-ap0: Cannot get device feature names: No such device
Cannot get device feature names: No such device
phy1-ap1: Cannot get device feature names: No such device
Cannot get device feature names: No such device
phy1-ap2: Cannot get device feature names: No such device
Cannot get device feature names: No such device
    wan:  on

sqrwv · 2023-02-24T15:39:52.142Z

It's because I have different interface names, you need to run:

root@D1:~# ls /sys/class/net
br-lan    lan2      lan4      miireg    phy1-ap0  phy1-ap2
lan1      lan3      lo        phy0-ap0  phy1-ap1  wan

To know your interface names.
And change the interface names in my the script, in the for line.

But those phy#-ap# are wireless interfaces, probably you have them disable, it won't affect the test if the script can't turn those on or off.
You can even delete them from the script.
We are using the wired interfaces in the test.

EDIT:
So you get slower speed with GRO off?
Interesting.
Do you have packet steering in both routers?

sampson · 2023-02-24T15:54:57.808Z

I know changing local startup should reboot.

Do I need to reboot after enable/disable packet steering / software flow offloading?

Yes, both with software flow offloading on and packet steering on.

main gro, 2nd gro, result
on, on, 1.83 Gb/sec
on, off, 2.09 Gb/sec
off, off, 1.41 Gb/sec
off, on, 1.79 Gb/sec

This is after gro.sh corrected interface names at both wrx36.

sqrwv · 2023-02-24T15:57:34.726Z

You only need to reboot after you change packet steering, but it's already enable.

So GRO is doing a good job.
Nice, default is on.

I had different results in WAN-LAN tests in AX3600 last year:
https://forum.openwrt.org/t/adding-openwrt-support-for-xiaomi-ax3600-part-1/55049/7887
Maybe something was changed meanwhile.

Your test was LAN-LAN?

sampson · 2023-02-24T16:40:36.804Z

Yes, all previous test with gro is LAN-LAN

For main 2.5Gb LAN to 2nd 2.5Gb WAN
main iperf3 -s
2nd iperf3 -c 192.168.8.1

gro
main, 2nd, result
on, on, 2.13 Gb/s
on, off, 2.35 Gb/s
off, on, 1.54 Gb/s
off, off, 1.63 Gb/s

frxc · 2023-02-24T16:44:41.661Z

PPPoE. With the following startup, I can reach 900 Mbps easily. (Big thanks to @SiXX )

#min frequency
echo 1382400 > /sys/devices/system/cpu/cpu0/cpufreq/scaling_min_freq
echo 1382400 > /sys/devices/system/cpu/cpu1/cpufreq/scaling_min_freq
echo 1382400 > /sys/devices/system/cpu/cpu2/cpufreq/scaling_min_freq
echo 1382400 > /sys/devices/system/cpu/cpu3/cpufreq/scaling_min_freq

#assign 4 rx interrupts to each cores
echo 8 > /proc/irq/50/smp_affinity
echo 4 > /proc/irq/51/smp_affinity
echo 2 > /proc/irq/52/smp_affinity
echo 1 > /proc/irq/53/smp_affinity

#assign 3 tcl completions to 3 CPUs
echo 4 > /proc/irq/73/smp_affinity
echo 2 > /proc/irq/74/smp_affinity
echo 1 > /proc/irq/75/smp_affinity

I can see 940 Mbps with Governor, but I'm not sure if it should always be on.
However, even at speeds of 800-900 Mbps, CPU0 still sees 97%.
I didn't fully understand the GRO part. Would it be sufficient to apply the script you mentioned here?

sqrwv · 2023-02-24T16:46:11.080Z

frxc:

I didn't fully understand the GRO part. Would it be sufficient to apply the script you mentioned here?

Never mind, looks like it's better the default on setting.
There were some test in past that would give better results with GRO off, but I've being running with default settings (GRO=on) and probably a lot changed since that time.

PPPoE uses more cpu than DHCP in wan.

frxc · 2023-02-24T20:32:32.503Z

By the way,
I just tried to upgrade (attended sysupgrade) from version r22127-595509cc78 to r22131-218f425b61. It took about 4 minutes, but after the device restarted, I noticed that it was still on the old version. Has anyone else experienced the same issue?

McGiverGim · 2023-02-24T20:59:15.052Z

I think remember gro was "bad" for pppoe setups only.

SiXX · 2023-02-24T21:01:12.817Z

I use the image builder a lot and yes - I notice I often have to flash it twice for it to work.

frxc · 2023-02-24T21:38:25.266Z

Thanks! I tested it and I'm getting better results with GRO off. "Band steering" and "Software flow offloading" are always on. I'll test it a bit more, but there is already an improvement in speed.

sqrwv · 2023-02-24T22:18:11.778Z

See if you need gro=off only in wan interface or if you need also the lan interfaces.

ka2107 · 2023-02-24T22:31:34.408Z

policy0 path works for me, hence I did not mention that in my post.
policy1, policy2, policy3 paths do not exist.

sqrwv · 2023-02-24T22:38:03.877Z

Problem is know, the workaround is to always stop WPAD before sysupgrade.

Some of us have incorporated the workaround (stopping wpad) in /lib/upgrade/stage2 (the upgrade script) and it's working as expected.

So, you can either stop wpad manually before every upgrade or edit /lib/upgrade/stage2 and add this at line 152:

# When wpad can't be killed by kill_remaining function, syupgrade will fail.
# If we stop it before, it will get killed and upgrade will work.
service wpad stop 2>/dev/null
sleep 6

Note: sleep 6 is working well for me, but if you still get any missing upgrades, increase the sleep value.

Then you must ensure your /lib/upgrade/stage2 file survives the upgrade, by adding it to system->backup/flash-> configuration tab.

Edit:
On a different subject, after installing today snapshot, attended sysupgrade was missing, I had to install luci-app-attendedsysupgrade package.

sampson · 2023-02-25T02:08:11.053Z

Question about the port LED

For the 2.5Gb port:

Both Green & Amber on about 3-5s after powered on, without cable.
Both will stay on until link is up, then Amber will blink while Green still solid on.
Same for PHY speed 1000 / 2500 .

For the 1Gb ports, both LEDs are off without cable.
When link is up, Green will be on and blink.
Amber always off.

How to make the 2.5Gb port LEDs report link status and speed status?

sampson · 2023-02-25T04:40:33.249Z

Some wifi client results:

main: Router to Internet
2nd: Using Client Wifi mode

Finding 1: in Client mode, Operating frequency->Width has no effect. No matter I use 20, 40, 80, 160 in 2nd, it will use the one provided by main

Finding 2:
At main, using AX, Auto, 160 MHz . It operates at 160 MHz . (36 & 100 are two detected and enable 160 MHz so far for me.) But if I use AX, 36 or 100, 160, then main will operate in 80MHz only.

As tested with 2nd, 80 MHz and 160 MHz are show the same iperf3 score: ~ 996M - 1.23G per second.
Only with 40 / 20 MHz, iperf3 score will be lower. ( 800M / 350M respectively)

Finding 3:
With htop running in main, CPU0 is less loaded in WIFI companed to LAN .

Finding 4:
Disable 5G wifi, enable 2.4G only in main .
With AX, Auto, 40, all clients connects in 20 mode only.
Only after Force 40MHz mode, then clients connected in 40 mode.
Tested with iPhone 13, 2nd, Nokia 9 . All the same.

xabolcs · 2023-02-25T07:21:52.319Z

sqrwv:

On a different subject, after installing today snapshot, attended sysupgrade was missing, I had to install luci-app-attendedsysupgrade package.

It's intended: it is removed for SNAPSHOT versions, but will included in RELEASE versions

github.com/openwrt/luci

luci: remove installation of attendedsysupgrade

committed 06:26PM - 20 Feb 23 UTC

aparcar

+0 -1

The package is only useful when running an official build. The ImageBuilders use…d for this service can not contain any special compiler flags etc. used by custom builds. Instead of installing the package whenever `luci` is installed, it should be installed via the buildbots itself, like it's currently done for `luci`. Signed-off-by: Paul Spooren <mail@aparcar.org>

See thread below about the motivations:

Is attendedsysupgrade now a requirement? How can I disable it from my build For Developers

I noticed that some new options are selected when I take my diffconfig and expand it to for building. Some are related to attendedsysupgrade. Can this be disabled? It seems like a hard dependency of luci now. Symbol: PACKAGE_luci-app-attendedsysupgrade [=y] Type : tristate Defined at tmp/.config-package.in:60753 Prompt: luci-app-attendedsysupgrade........ LuCI support for attended sysupgrades Location: -> LuCI -> 3. Applications Selects: PACKAGE_attendedsysupgrade-common [=y] && PACK…

frollic · 2023-02-25T08:32:46.302Z

should this info perhaps be added to the device page ?