Hao Xu

Hi akseg73, I'm not an expert of io_uring, but I guess Jens and Pavel are busy now, so allow me to answer part of your questions as possible as I...

May I get more details: 1. Is it buffered/direct write 2. Do you observe the cpu usage when it goes up from 4 therads to 32 3. for the synchronous...

Hi @dcolascione , I'm working on something like this, hopefully it can come out before long.

Hi johalun, which version of kernel are you using? Do the io-workers hang there for socket reads/writes?

Did the number of worker threads exceed the registered limit with/without IOSQE_ASYNC now?

Btw @axboe , Have you started working on IOSQE_ASYNC_HYBRID stuff, if not I'd like to give it a shot.

Hi @johalun Do you mind try the below two patches to see if it gets better: (I developed them based on https://git.kernel.dk/linux-block for-5.16/io_uring) Since this IOSQE_ASYNC_HYBRID flag hasn't been merged...

> Hmm, after a clean rebuild of my app it seems to work. Could have been just a temporary weirdness with the macro having wrong value or something. I'll run...

> so for networked reads, IOSQE_ASYNC_HYBRID will always deliver better performance? > IMHO, no. It really depends on how many cpu resources users want to use. Suppose we are doing...

> Hmm, after a clean rebuild of my app it seems to work. Could have been just a temporary weirdness with the macro having wrong value or something. I'll run...