axboe / fio

Flexible I/O Tester
GNU General Public License v2.0
5.24k stars 1.26k forks source link

fio with multiple Uring workers #1825

Open HazyMrf opened 2 weeks ago

HazyMrf commented 2 weeks ago

Is it possible to somehow setup io_uring engine to have multiple "write" workers when using fio? There must be such possibility, but there is nothing about it in the docs:

$ fio --enghelp=io_uring
hipri                   : Use polled IO completions
cmdprio_percentage      : Send high priority I/O this percentage of the time
cmdprio_class           : Set asynchronous IO priority class
cmdprio                 : Set asynchronous IO priority level
cmdprio_bssplit         : Set priority percentages for different block sizes
fixedbufs               : Pre map IO buffers
registerfiles           : Pre-open/register files
sqthread_poll           : Offload submission/completion to kernel thread
sqthread_poll_cpu       : What CPU to run SQ thread polling on
nonvectored             : Use non-vectored read/write commands
uncached                : Use RWF_UNCACHED for buffered read/writes
nowait                  : Use RWF_NOWAIT for reads/writes
force_async             : Set IOSQE_ASYNC every N requests

$ fio --version
fio-3.28

$ uname -a
Linux 6.8.0-1014-aws

I use such setup: taskset -c 20-23 fio --name test --ioengine=io_uring --rw=randwrite --iodepth=64 --bs=256k --size=900G --filename=/mnt/test/fio.data --runtime=600 --time_based and have only one uring worker :(

image
axboe commented 2 weeks ago

io_uring in the kernel will manage those, and create more as needed. Generally only if one blocks, if things don't block, then it will not create more workers. Don't mistake these workers as a way to create parallelism, creating more workers than needed just creates contention.

For your case of buffered writes, generally most IOs will just copy into the page cache and dirty the page(s). Outside of that, some writes will end up having to wait for balancing of dirty pages.

On top of that, since it's a single file, io_uring will also serialize writes to it as buffered writes to the same file are exclusive. This is the primary reason you only see one writer above. O_DIRECT (--direct=1) will not run into that issue.

HazyMrf commented 2 weeks ago

Thank you for such a quick response @axboe. It appeared that I made a mistake in my question. I want not many workers, but many cores for my workers. Maybe I can somehow set up this in fio + uring, I mean things like https://man7.org/linux/man-pages/man3/io_uring_register_iowq_aff.3.html may help IDK.

HazyMrf commented 2 weeks ago

I'm trying to reproduce the uneven write load distribution with fio + io_engine=io_uring from https://github.com/axboe/liburing/issues/1260

axboe commented 2 weeks ago

Just run more jobs? This isn't an io_uring type option, that's a generic fio opion. --numjobs=X would do that. But then you probably want to get rid of the --filename option and just use --directory=/mnt/test/ so that each job will use its own file.

HazyMrf commented 1 week ago

Thank you for your response @axboe. I thought about it, but what I actually want is 1 job, but multiple cores for io_uring workers, in other words somehow set affinity to my io_uring with https://man7.org/linux/man-pages/man3/io_uring_register_iowq_aff.3.html