LEDE Luci QOS and SQM at same time

An alternative to DSLReports I recommend is the Sourceforge Speed Test

Click on the View Details button at the conclusion of the test.

The sourceforge speed test seems to be pretty bogus and inaccurate, it claims an upload speed >25% faster than my actual line speed (18,2% above my contractual upload speed, 24.1% above my VDSL2 modem's sync rate, ~27% above my effective upload speed) - the download speed seems to be roughly correct (just slightly below the actual figures).

If you're overprovisioned, you can see speeds faster than advertised.

My speeds seem normal.


That would only be possible if sf's speed test would use too small and/ or compressable chunks of data for the upload test. Overprovisioning wouldn't explain values almost a quarter above the actual sync rate of my VDSL2 modem, which is just slightly below the physical maximum of the DSLAM line card and the attainable rate - and in every case more than a quarter above the values I get in real world tests with actual data (both small, medium and several GB in size).

can anyone help me with The FLExible Network Tester? i only have windows 10 on pc and laptop, how can i use this to test bufferbloat?

There are two options:

a) install a VM-system like the free virtualbox and create an ubuntu virtual machine (will work, but might not be as efficient as running on bare metal)

b) try to leverage windows 10's linux subsystem (see https://docs.microsoft.com/en-us/windows/wsl/install-win10), this might or might not work, there have been some recent changes to flent's git repository to work better with windoes, but I do not know whether "better" already means "at all".

After you get flent/netperf installed, please contact me again for actually running a test...

I probably would first try my luck with getting the dslreports speedtest and/or its CLI (commad line interface) up and running. As @JonP indicated, thee is a write-up of configuration recommendations for the dslreports speedtest at https://forum.openwrt.org/t/sqm-qos-recommended-settings-for-the-dslreports-speedtest-bufferbloat-testing/2803 (but these will only help after you got around the error; Question: which browsers did you test?).

i used 60 seconds for both up and down

i use chrome browser.

What to you get just using the DSLReports default settings?

Usually A but I get red spikes on download with default test, not with the above setting, seems to stay in green for entire 60 seconds.

Here's the test with wifi 2.4ghz

Upload speed is improved.

Keep in mind these numbers and "grades" can change minute to minute.

I use servers that are close to me geographically, and use the default settings, which I feel is a more "real-world" test than rigging the test settings to get a desired result.

Are you now using SQM only? QOS only? Both?

Okay, this looks pretty decent. The spikes especially in downloading at around 66ms are still not nice, but they seem to be intermittent so I would ignore them for now.

@Happi, @jwoods might be interested in those results, I certainly am not. By default dslreports, like most other speedtests, will measure for a much shorter duration and that often hides bufferbloat (or rather does make its consequences less obvious).

Compared to the wired test above this indicates that on wifi you have other issues (the observed bandwidth indicates that sqm's shaper is not even exercised that much as wifi seems to be the bottleneck). Also the android browser you used is not able to sustain the requested 16/16 streams and falls back to 3/3 ("00.01s restrict to 3 streams due to slow browser"), the RTT to the servers is also significantly worse than during the wired test.
Could you repeat the wifi test while logging into the router and monitor the output of "top -d 1" (look whether the idle % gets close to zero).

That's fine.

I'm interested in seeing tests that are not rigged to get a desired result.

@jwoods, what in the recommendations under https://forum.openwrt.org/t/sqm-qos-recommended-settings-for-the-dslreports-speedtest-bufferbloat-testing/2803 appears to "rig" the test, and why do you believe the default settings to be more "real-world"?

And I am interested in data supporting your insinuation about rigging.

How? These scripts are shell scripts that will not run on windows (at least not out of the box) and they require a gn/bsdu ping and netperf to be installed and in the path, that is going to be tough for non-unix users. (Sure it is possible to run this on macosx but that will not help a considerable number of users that are windows only). And running netperf on the router itself (which will make the user a linux-user) is certainly not recommended if one intends to test the routers routing performance.

I guess your talking about @richb-hanover's betterspeedtest.sh

We ask users all the time to SSH in to the router and run commands...not difficult.

opkg update

opkg install netperf

Cut and paste the contents of betterspeedtest.sh into an .sh file on the router, or copy it over in its entirety.

Run the shell (example is in /tmp)

root@LEDE:/tmp# sh betterspeedtest.sh -4 -H netperf.bufferbloat.net -t 60

I changed the ping location in the script from gstatic.com to google.com.

To cite myself why I consider the idea to run netperf on the router less than ideal:

The point is that running netperf has a considerable CPU cost and will exercise different parts of the network stack and the firewall than routing. This is still a fine test, but it does not test the typical use-case so I would not recommend that to get an sqm configuration in decent shape.

1 Like

Your opinion.

We mitigate bufferbloat at the router, not endpoints.