Bufferbloat, or when do we get fq_codel?

[nagle@animats.com]

The resold AT&T service seems to have bufferbloat problems, and fails bufferbloat tests. The test is simple. Run Speed Test and a ping at the same time; if the ping time shoots way up, there's bufferbloat. VoIP and games will then work badly, and uploads will interfere with downloads.

Here's a screenshot of that: https://image.ibb.co/dzF4jy/speedtest.png. Notice the ping time going from 30ms to around 300ms once the upload test starts. That's bufferbloat in action. Whatever sends the most gets to hog the link.

The fix for this is called fq_codel. Huge performance win. When do we get that?

[danielg4]

You won't. The resold AT&T service is VDSL, and DSP drivers for DSL are written without regard for bufferbloat mitigation and very proprietary. There is no workaround, now or likely ever.

[oddhack]

I was under the impression that bufferbloat mitigation could be done in your own router sitting before the modem. Is that incorrect?

[danielg4]

The non-optional AT&T gateway introduces its own bufferbloat even in DMZ+/IP Passthrough mode, both in its DSP driver and in the transparent mangling it still applies to both layer 3 and layer 4. (!)

Using either cake or fq_codel behind it has marginal impact at best.