What is your local machine? Max parallelism is often a function of CPU count, so knowing how many core the box has is useful, and might explain needing to do more tests between 20 and 30.
There’s an old rule of thumb that tasks that alternate between cpu and IO usage, such as for instance compiler jobs, often benefit from running 2 tasks per core as the second task can make about 50% more progress per second. Your results are awfully similar to that rule of thumb so I’m wondering is this an 8 core box? 10? 12?
> If the buffer has reached its maximum size , it blocks writes until the buffer has been flushed, to properly exert backpressure on the application.
I cannot think of a situation where I could convince my team to allow 100k inserts to be pooled up waiting in memory for a process to crash or a deployment to wipe it out, but having a single buffer is also a mistake. io_uring is a refinement of a very old idea of having a read and a write buffer that swap back and forth between two threads or tasks.
You write to one for a while, and then you hand it off to someone to process it and clear it, and you write to another while that is happening. You only block when both are full. I suspect you’re seeing some of the fruits of this in the COPY example, if you’re creating a new file for each batch.