I note that "proof" is very much in the developer's opinion and limited testing base. Actual operational experience, as in a real deployment, with other applications, heavy context switching, or virtualization, might yield better results. There's lots of defaults in the linux kernel that are just swags, the default NAPI and rx/tx ring buffer sizes being two where devs just copy/paste stuff, which either doesn't scale up, or doesn't scale down. This does not mean I oppose your patch! However I have two points I'd like to make regarding bql and dql in general that I have long longed be explored. 0) Me being an advocate of low latency in general, does mean that I have no problem and even prefer, starving the device rather than always keeping it busy. /me hides 1) BQL is MIAD - multiplicative increase, additive decrease. While in practice so far this does not seem to matter much (and also measuring things down to "us" really hard), a stabler algorithm is AIMD. BQL often absorbs a large TSO burst - usually a minimum of 128k is observed on gbit, where a stabler state (without GSO) seemed to be around 40k on many of the chipsets I worked with, back when I was working in this area. (cake's gso-splitting also gets lower bql values in general, if you have enough cpu to run cake) 2) BQL + hardware mq is increasingly an issue in my mind in that, say, you are hitting 64 hw queues, each with 128k stored in there, is additive, where in order to service interrupts properly and keep the media busy might only require 128k total, spread across the active queues and flows. I have often thought that making BQL scale better to multiple hw queues by globally sharing the buffering state(s), would lead to lower latency, but also that probably sharing that state would be too high overhead. Having not worked out a solution to 2), and preferring to start with 1), and not having a whole lot of support for item 0) in the world, I just thought I'd mention it, in the hope someone might give it a go.
