I observed poor performance of io_uring compared to synchronous IO. That
turns out to be caused by deeper CPU idle states entered with io_uring,
due to io_uring using plain schedule(), whereas synchronous IO uses
io_schedule().
The losses due to this are substantial. On my cascade lake workstation,
t/io_uring from the fio repository e.g. yields regressions between 20%
and 40% with the following command:
./t/io_uring -r 5 -X0 -d 1 -s 1 -c 1 -p 0 -S$use_sync -R 0 /mnt/t2/fio/write.0.0
This is repeatable with different filesystems, using raw block devices
and using different block devices.
Use io_schedule_prepare() / io_schedule_finish() in
io_cqring_wait_schedule() to address the difference.
After that using io_uring is on par or surpassing synchronous IO (using
registered files etc makes it reliably win, but arguably is a less fair
comparison).
There are other calls to schedule() in io_uring/, but none immediately
jump out to be similarly situated, so I did not touch them. Similarly,
it's possible that mutex_lock_io() should be used, but it's not clear if
there are cases where that matters.
Cc: Jens Axboe <[email protected]>
Cc: Pavel Begunkov <[email protected]>
Cc: [email protected]
Cc: [email protected]
Signed-off-by: Andres Freund <[email protected]>
---
io_uring/io_uring.c | 16 ++++++++++++++--
1 file changed, 14 insertions(+), 2 deletions(-)
diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
index 3bca7a79efda..4661a39de716 100644
--- a/io_uring/io_uring.c
+++ b/io_uring/io_uring.c
@@ -2575,6 +2575,9 @@ int io_run_task_work_sig(struct io_ring_ctx *ctx)
static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
struct io_wait_queue *iowq)
{
+ int ret;
+ int token;
+
if (unlikely(READ_ONCE(ctx->check_cq)))
return 1;
if (unlikely(!llist_empty(&ctx->work_llist)))
@@ -2585,11 +2588,20 @@ static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
return -EINTR;
if (unlikely(io_should_wake(iowq)))
return 0;
+
+ /*
+ * Use io_schedule_prepare/finish, so cpufreq can take into account
+ * that the task is waiting for IO - turns out to be important for low
+ * QD IO.
+ */
+ token = io_schedule_prepare();
+ ret = 0;
if (iowq->timeout == KTIME_MAX)
schedule();
else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
- return -ETIME;
- return 0;
+ ret = -ETIME;
+ io_schedule_finish(token);
+ return ret;
}
/*
--
2.38.0
On 7/7/23 10:20?AM, Andres Freund wrote:
> I observed poor performance of io_uring compared to synchronous IO. That
> turns out to be caused by deeper CPU idle states entered with io_uring,
> due to io_uring using plain schedule(), whereas synchronous IO uses
> io_schedule().
>
> The losses due to this are substantial. On my cascade lake workstation,
> t/io_uring from the fio repository e.g. yields regressions between 20%
> and 40% with the following command:
> ./t/io_uring -r 5 -X0 -d 1 -s 1 -c 1 -p 0 -S$use_sync -R 0 /mnt/t2/fio/write.0.0
>
> This is repeatable with different filesystems, using raw block devices
> and using different block devices.
>
> Use io_schedule_prepare() / io_schedule_finish() in
> io_cqring_wait_schedule() to address the difference.
>
> After that using io_uring is on par or surpassing synchronous IO (using
> registered files etc makes it reliably win, but arguably is a less fair
> comparison).
>
> There are other calls to schedule() in io_uring/, but none immediately
> jump out to be similarly situated, so I did not touch them. Similarly,
> it's possible that mutex_lock_io() should be used, but it's not clear if
> there are cases where that matters.
This looks good to me, and I also separately tested a similar patch and
it showed good results for me even with a heavily performance oriented
setup:
pread2 io_uring io_uring w/io_sched
QD1 185K 170K 186K
QD2 NA 304K 327K
QD4 NA 630K 640K
QD8 NA 891K 892K
I'll add this, with just one small minor cosmetic edit:
> @@ -2575,6 +2575,9 @@ int io_run_task_work_sig(struct io_ring_ctx *ctx)
> static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
> struct io_wait_queue *iowq)
> {
> + int ret;
> + int token;
Should just be a single line.
And I'll mark this for stable as well. Thanks!
--
Jens Axboe
On Fri, 07 Jul 2023 09:20:07 -0700, Andres Freund wrote:
> I observed poor performance of io_uring compared to synchronous IO. That
> turns out to be caused by deeper CPU idle states entered with io_uring,
> due to io_uring using plain schedule(), whereas synchronous IO uses
> io_schedule().
>
> The losses due to this are substantial. On my cascade lake workstation,
> t/io_uring from the fio repository e.g. yields regressions between 20%
> and 40% with the following command:
> ./t/io_uring -r 5 -X0 -d 1 -s 1 -c 1 -p 0 -S$use_sync -R 0 /mnt/t2/fio/write.0.0
>
> [...]
Applied, thanks!
[1/1] io_uring: Use io_schedule* in cqring wait
(no commit info)
Best regards,
--
Jens Axboe