Just in case you are having trouble finding it here is a link to the algorithm that was used to solve buffer bloat.

https://en.wikipedia.org/wiki/CoDel#Derived_algorithms

Your time to flush is almost exactly the maximum time to transit the queue that CoDel uses.

Packet drops can be replaced by throttling user space when userspace is writing.

What is very nice about that approach is that it is auto-tuning. Working on both fast and slow devices with the same parameters.

I have been wondering for years when I see these conversations about block layer throttling come up if a CoDel style solution might fix things.

Given that CoDel works hop by hop it might even be possible to apply to each layer of the I/O stack independently. Given fs vs block vs network could be a real advantage.