Re: Cake compiled for the ERL

 Thanks a lot for your feedback. I'll take this over to the Cake mailing list soon, as much of the testing isn't ER-X specific. I like how Cake does so much with very little setup effort, also in the "no-knobs" spirit of fq_codel.

What I really wish is that an AQM algorithm could exist to function with two fewer major knobs: upload and download bandwidth. I've found it difficult to make AQM work with my WiFi uplinks without cutting bandwidth to something like 50% of the sync rate, as instantaneous bandwidth varies too much. I understand that the shaper needs to know what rate to restrict to and has no knowledge of this, but something that dynamically adapts somehow would make things easier. Imagine a world in which one could just flip a switch to "turn AQM on", or better yet not flip the switch at all, since it's already on.

Your info about implementing BQL is eye opening, but yes, we're after getting rid of that inherent latency, if possible. I would really want to understand what happens to the packet counters in all of the code paths, and get the resets right. I've not even gotten it compiled yet, much less gotten to that point. Image may be NSFW.
Clik here to view. In the back of my mind is also the thought that there's no way Ubiquiti would include this support in the kernel unless it's rock solid, which would almost certainly mean a lot of testing and a thorough review from either the original author of the driver or an Ethernet driver expert.

 Thanks also for your help in getting a setup for compiling the ER-X kernel. You're right about the raeth driver using PDMA. These links made for interesting reading:

https://patchwork.kernel.org/patch/9297625/

http://lists.infradead.org/pipermail/linux-mediatek/2016-January/003428.html

Perhaps the authors of those posts could help with either implementation or code reviews of any driver changes? Although, they're referring to a different Mediatek driver codebase, I think.

I've got a Debian VM with the mipsel environment installed without issue. The challenge now is even after I've compiled a kernel for the ER-X, I don't have one available for testing. I can't put mine at home or at work up on blocks, so I'm considering buying another one, but will see. After what Dave wrote, I'm not sure the testing of any changes can be done anywhere but on the native hardware itself. I'm also not sure if I have the time and knowledge to get BQL implemented in the driver!