From: Hao Xu <howeyxu@xxxxxxxxxxx>
Hi Jens and all,
This is an early RFC for a new way to do async IO. Currently io_uring
works in a way like:
- issue an IO request in nowait way
here nowait means return error(EAGAIN) to io_uring layer when it would
block in deeper kernel stack.
- issue an IO request in a normal(block) way
io_uring catches the EAGAIN error and create/wakeup a io-worker to
redo the IO request in a block way. The original context turns to
issue other requests. (some type of requests like buffered reads,
leverage task work to wipe out io-workers)
This has two main disadvantages:
- we have to find every block point along the kernel code path and
modify it to support nowait.
e.g. alloc_memory() ----> if (alloc_memory() fails) return -EAGAIN
This hugely adds programming complexisity, especially when the code
path is long and complicated. For example, buffered write, we have
to handle locks, possibly journal part, meta data like extent node
misses.
- By create/wakeup a new worker, we redo a IO request from the very
beginning, which means we re-walk the path from beginning to the
previous block point.
The original context backtracks to the io_uring layer from the block
point to submit other requests. While it's better to directly start
the new submission.
This RFC provides a new way to do it.
- We maintain a worker pool for each io_uring instance and each worker
in it can submit requests. The original task only needs to create the
first worker and return to userspace. Later it doesn't need to call
io_uring_enter.[1]
- the created worker begins to submit requests. When it blocks, just
let it be blocked. Create/wakeup another worker to do the submission
[1] I currently keep these workers until the io_uring context exits. In
other words, a worker does submission, sleep, wake up, but won't
exit. Thus the original task don't need to create/wakeup workers.
I've done some testing:
name: buffered write
fs: xfs
env: qemu box, 4 cpu, 8G mem.
tool: fio
- single file test:
fio ioengine=io_uring, size=10M, bs=1024, direct=0,
thread=1, rw=randwrite, time_based=1, runtime=180
async buffered writes:
iodepth
1 write: IOPS=428k, BW=418MiB/s (438MB/s)(73.5GiB/180000msec);
2 write: IOPS=406k, BW=396MiB/s (416MB/s)(69.7GiB/180002msec);
4 write: IOPS=382k, BW=373MiB/s (391MB/s)(65.6GiB/180000msec);
8 write: IOPS=255k, BW=249MiB/s (261MB/s)(43.7GiB/180001msec);
16 write: IOPS=399k, BW=390MiB/s (409MB/s)(68.5GiB/180000msec);
32 write: IOPS=433k, BW=423MiB/s (443MB/s)(74.3GiB/180000msec);
1 lat (nsec): min=547, max=2929.3k, avg=1074.98, stdev=6498.72
2 lat (nsec): min=607, max=84320k, avg=3619.15, stdev=109104.36
4 lat (nsec): min=891, max=195941k, avg=9062.16, stdev=213600.71
8 lat (nsec): min=684, max=204164k, avg=29308.56, stdev=542490.72
16 lat (nsec): min=1002, max=77279k, avg=38716.65, stdev=461785.55
32 lat (nsec): min=674, max=75279k, avg=72673.91, stdev=588002.49
uringlet:
iodepth
1 write: IOPS=120k, BW=117MiB/s (123MB/s)(20.6GiB/180006msec);
2 write: IOPS=273k, BW=266MiB/s (279MB/s)(46.8GiB/180010msec);
4 write: IOPS=336k, BW=328MiB/s (344MB/s)(57.7GiB/180002msec);
8 write: IOPS=373k, BW=365MiB/s (382MB/s)(64.1GiB/180000msec);
16 write: IOPS=442k, BW=432MiB/s (453MB/s)(75.9GiB/180001msec);
32 write: IOPS=444k, BW=434MiB/s (455MB/s)(76.2GiB/180010msec);
1 lat (nsec): min=684, max=10790k, avg=6781.23, stdev=10000.69
2 lat (nsec): min=650, max=91712k, avg=5690.52, stdev=136818.11
4 lat (nsec): min=785, max=79038k, avg=10297.04, stdev=227375.52
8 lat (nsec): min=862, max=97493k, avg=19804.67, stdev=350809.60
16 lat (nsec): min=823, max=81279k, avg=34681.33, stdev=478427.17
32 lat (usec): min=6, max=105935, avg=70.55, stdev=696.08
uringlet behaves worse on IOPS and lantency in small iodepth. I think
the reason is there are more sleep and wakeup.(not sure about it, I'll
look into it later)
The downside of uringlet:
- it costs more cpu resource, the reason is similar with the sqpoll case: a
uringlet worker keeps checking sqring to reduce latency.[2]
- task->plug is disabled for now since uringlet is buggy with it.
[2] For now, I allow a uringlet worker spin on the empty sqring for some
times.
Any comments are welcome, This early RFC only supports buffered write for
now and if the idea under it is proved to be the right way, I'll change
it to a formal patchset and resolve the detail technical issues and try
to support more io_uring features.
Regards,
Hao
