Good. We just don't know. The potential to have measurements in this forum thread based on settings involving multiple clients is a possibility and not nill. Good that you were fast with your feedback, so we can conclude that multiple client streams handling (especially for RX) could also be improved upon, but even if that were fixed, it most likely will not fix the issues you are having.
My dream is that y'all test latency simultaneously with throughput, especially in the problematic cases with walls in the way. Even taking apart packet captures (via tcptrace -G and xplot.org) of iperf3's RTTs would be preferable to just seeing ya'll just reporting bandwidth figures.
I can note a similar issue with this. I setup a WAX202 for my church in a WWAN configuration. I first set it up parallel to a wall (hoping to present a low profile, save space and move the LEDs from direct view of the congregation). From a client in the same room approximately 5 meters away (sanctuary - no obstructions), I could not maintain connection or Internet. Immediately after rotating the device 90 degrees to be perpendicular to the wall - I had 0 issues.
LEDs are controllable.
Yes I know, but this happened on first boot at the location, lol.
Not sure how that's relevant to what I discovered, though.
Could you please give a simple example for how to do it and how it should look like? tcptrace seems to be a very old program (from 2003) and I am running Windows on my machine, which it apparently was not designed for.
Edit: I will try to set up OWAMP or TWAMP, as suggested by this website: https://kadiska.com/measure-network-latency/
Do the packet capture on windows via wireshark. Wireshark also has a pretty decent Statistics -> Stream graphs -> TCP RTT. What I'm mostly looking for is large gaps or spikes in the RTT, which would be more the radio rate seeking malfunctioning, or uncontrolled growth, like I show below.
Yea, tcptrace -G and xplot.org are ancient tools and only work well on linux, but they in general produce outputs that are more detailed.
Heading deeper, if it's more rate control that's the problem here, actually taking captures of the 802.11 frames might be revealing. Lack of packet aggregation proved to be a problem on a mt76 I took apart recently... you can indirectly see aggregation in action by observing layer three packet arrivals all in a bunch...
Anyway, here's an example of a packet capture of iperf3, using wireshark, running over LTE. Over the course of 10seconds it gradually grows to over 1.5s (with no end in sight!). I've seen wifi be even worse... A bidir test will utterly go to hell here...
irtt is far more modern and fine grained than twamp, and easier to setup... it's in pure go....
I managed to install and configure twamp:
Client ← Wifi 802.11ax 5G ch40 80MHz, 2 Walls (~5 meter) ~ 69dBm → D-Link DAP x1860 OpenWRT snapshot 2022-12-22 ← wifi 802.11ac 5G ch40 80MHz, 1 wall → Fritzbox 7490 OEM ← Gigabit LAN → Server
I tried to use Wireshark and found the "roundtrip" option, but was not sure which pakets to select and how to purely measure specific traffic from within my local network. Using the "just try things" method, I downloaded a large 2 gigabyte file from a more or less reliable website (Linux Mint) and then selected a random packet from this large download, which gave me these results:
Client ← Wifi 802.11ax 5G ch40 80MHz, 2 Walls (~5 meter) ~ 69dBm → D-Link DAP x1860 OpenWRT snapshot 2022-12-22 ← wifi 802.11ac 5G ch40 80MHz, 1 wall → Fritzbox 7490 OEM → Internet
I think this is fair enough. Not the best, but since there are 3 walls, my fritzbox 7490 being quite old hardware and the endpoint being on the internet with how many servers in between, i think acceptable for the fact that I have not tried to optimize my router and wifi configuration in any way.
I was not sure, if i should post these results, since I have not found grave issues, but who knows, maybe you can do more with it than me.
@dtaht How can you find the traffic you conducted via iperf3 in wireshark? Edit: When we found them, which packets should we choose for the roundtrip test?
In the screenshot down below, you can see packets from an iperf3 test with following settings:
iperf3 -c 192.168.178.39 -p 5201 -P 8 -t 60 -i 10 --bidir
- Iperf Client < --- Wifi 802.11ax 5G ch40 80MHz, 2 Walls (~5 meter) ~ 69dBm --- > D-Link DAP x1860 OpenWRT snapshot 2022-12-22 < ---Wifi 802.11ac 5G ch40 80MHz, 1 wall --- > Fritzbox 7490 OEM < --- Gigabit LAN ---> Iperf Server 192.168.178.39.
There are hundreds and thousands of packages to choose from, but which one(s) should we select?
I clicked on one of the packets and had this result:
But when I click on another random packet, results can be slightly different.
The stream graphs start with the identified beginning of the flow and plot from there, so selecting any packet in the flow should more or less give you the same result from the wireshark stream graphs.
See how much larger the RTT variance is on your last plot? spikes as high as 60ms, the vast majority close to zero? (this is essentially a zero length path).
(also selecting a throughput graph on the same data and looking at both at the same time is revealing - part of why I use xplot, but whatever)
From what I see of both plots, you are not aggregating very well at all, min RTT
should have been hovering at at least 4ms, instead it is closer to 1. Good aggregation gives 60% or more better throughput...
You can also see packet loss and retransmits...
There's another sort of packet capture you can take of the 802.11 packets "in the air" to see how many AMPDUs are being sent.
clicking on any random packet within a flow (src,dst,dst port,src port, proto should all be the same) should give you the same graph no matter where in the flow you clicked.
@amteza I kind of burned out for a while. You get anywhere on the NAPI stuff? (merry christmas!)
words I don't understand:
- "zero length path". - Length of what? Do you mean "zero latency path"?
How to do good aggregation? How should good aggregation look like?
"zero base latency" path. I use length, rather than "latency", because it's the speed of light I think about. A nanosecond is about a foot, so...
A full TXOP of aggregated packets can be as big as 5.7ms (and we usually can get to 2 or 3), so 30ns vs 15ms is close enough to "zero". Sorry to be unclear.
There are a lot of things that can prevent good aggregation - notably not negotiating the capability at all, a driver problem, or excessive packet loss. Also the notion of "good" is fungible. 8 minutes of this explains how badly we used to do aggregation: https://www.youtube.com/watch?v=Rb-UnHDw02o&t=1560s - and the mt76 is supposed to be doing this far, far more right that it was then - but given all the complaints in every direction, not just on this, but power, I'm beginning to suspect something else is rather wrong... 
Anyway the theory that you are not aggregating well is just a theory for now. Here's how to capture the wifi management frames on a variety of OSes, which could find the truth, as opposed to a theory..
 and I will have to get on top of a mt76 chip to dig deeper myself.
Happy summer holidays. The last weeks were crazy, and I just returned from a trip with my family, so I've got nowhere. However, before leaving, I left running the latest master build with a new patch to fix NAPI settings to our previously tested value. My devices are usually running with latencies under load below 8 ms. Is it related to the NAPI weight being lowered? Maybe, but I am not sure.
My devices are mt76 (7615e) AC devices, not AX (7915).
For completeness, test running r23234, with the usual patches, at 3 metres from the AP, 866 Mb/s 2x2 MIMO, MCS: 9 and NSS: 2. ISP connection is 1000/50, SQM is 800/42 and WiFi network is an AC network. The network was only in use by a Roblox game, YouTube, and a FaceTime call apart from the Waveform test, of course:
A quick comparison of AX vs AC both on 80MHz channel 64 one floor so a few meters on rt3200 with a 7915 chip on openwrt r21885 with all offloading including wifi downstream on a Macbook Air M2. Even saw some RTT over 1000 ms on a couple of AX runs. iperf server on a raspberry connected with ethernet.
AX [ 7] 0.00-1.00 sec 6.04 MBytes 50.6 Mbits/sec 408336 1.34 MBytes 412ms [ 7] 1.00-2.00 sec 39.6 KBytes 325 Kbits/sec 0 987 KBytes 662ms [ 7] 2.00-3.00 sec 1.60 MBytes 13.4 Mbits/sec 653944 822 KBytes 489ms [ 7] 3.00-4.01 sec 1.53 MBytes 12.8 Mbits/sec 0 1.11 MBytes 596ms [ 7] 4.01-5.01 sec 382 KBytes 3.13 Mbits/sec 0 1.05 MBytes 512ms [ 7] 5.01-6.00 sec 2.91 MBytes 24.6 Mbits/sec 0 877 KBytes 627ms [ 7] 6.00-7.00 sec 2.35 MBytes 19.6 Mbits/sec 0 1.08 MBytes 573ms [ 7] 7.00-8.00 sec 524 KBytes 4.30 Mbits/sec 0 1.00 MBytes 485ms [ 7] 8.00-9.00 sec 2.68 MBytes 22.5 Mbits/sec 0 1.08 MBytes 572ms [ 7] 9.00-10.00 sec 2.09 MBytes 17.5 Mbits/sec 0 1.10 MBytes 663ms - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 7] 0.00-10.00 sec 20.1 MBytes 16.9 Mbits/sec 1062280 sender [ 7] 0.00-10.60 sec 18.7 MBytes 14.8 Mbits/sec receiver AC [ ID] Interval Transfer Bitrate Retr Cwnd RTT [ 7] 0.00-1.00 sec 43.9 MBytes 368 Mbits/sec 4881208 1.48 MBytes 20ms [ 7] 1.00-2.00 sec 50.0 MBytes 420 Mbits/sec 2896000 1.53 MBytes 23ms [ 7] 2.00-3.00 sec 50.4 MBytes 423 Mbits/sec 1448000 1.57 MBytes 16ms [ 7] 3.00-4.00 sec 51.3 MBytes 431 Mbits/sec 0 1.61 MBytes 35ms [ 7] 4.00-5.00 sec 44.5 MBytes 373 Mbits/sec 1531984 2.04 MBytes 25ms [ 7] 5.00-6.00 sec 42.5 MBytes 356 Mbits/sec 1743392 2.05 MBytes 51ms [ 7] 6.00-7.00 sec 45.2 MBytes 380 Mbits/sec 0 2.06 MBytes 38ms [ 7] 7.00-8.00 sec 47.8 MBytes 401 Mbits/sec 1867920 1.56 MBytes 33ms [ 7] 8.00-9.00 sec 46.2 MBytes 386 Mbits/sec 1614520 1.18 MBytes 62ms [ 7] 9.00-10.00 sec 44.7 MBytes 375 Mbits/sec 1669544 1.90 MBytes 35ms - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 7] 0.00-10.00 sec 466 MBytes 391 Mbits/sec 17652568 sender [ 7] 0.00-10.02 sec 464 MBytes 388 Mbits/sec receiver
Try a 160MHz channel as a workaround.
Are you using IPerf2 on the Mac to test it?
Tried that too but didn't help.
iperf3-darwin available on macOS