* [PATCH v2] Fix a rate limit issue.
@ 2021-01-16 19:52 HongweiQin
2021-01-16 20:10 ` Jens Axboe
0 siblings, 1 reply; 2+ messages in thread
From: HongweiQin @ 2021-01-16 19:52 UTC (permalink / raw)
To: fio; +Cc: axboe, HongweiQin
In the current implementation, should_check_rate() returns false
if ddir_rw_sum(td->bytes_done)==0. Therefore, a thread may violate
the rate if iodepth*bs > rate.
This patch addresses the issue by not checking td->bytes_done in should_check_rate.
An example of the issue:
[root@localhost test]# cat fio_randwrite
[global]
thread
kb_base=1000
direct=1
size=28GiB
group_reporting
io_size=16384
ioengine=libaio
iodepth=2
bs=4096
iodepth_batch_submit=1
iodepth_batch_complete=1
filename=/dev/qblkdev
[fio_randwrite]
rw=randwrite
rate_iops=,1
iodepth_batch_submit=1
thinktime_blocks=1
rate_cycle=1000
thinktime=3s
rate_ignore_thinktime=1
[root@localhost test]# fio fio_randwrite
blktrace output:
259,1 11 1 0.100550729 6135 Q WS 3541608 + 8 [fio]
259,1 11 2 0.100552183 6135 G WS 3541608 + 8 [fio]
259,1 11 3 0.100560373 6135 D WS 3541608 + 8 [fio]
259,1 11 4 0.100570436 6135 C WS 3541608 + 8 [0]
259,1 11 5 0.100599816 6135 Q WS 43470024 + 8 [fio]
259,1 11 6 0.100600513 6135 G WS 43470024 + 8 [fio]
259,1 11 7 0.100601579 6135 D WS 43470024 + 8 [fio]
259,1 11 8 0.100612750 6135 C WS 43470024 + 8 [0]
259,1 11 9 3.101034407 6135 Q WS 49511928 + 8 [fio]
259,1 11 10 3.101036067 6135 G WS 49511928 + 8 [fio]
259,1 11 11 3.101054487 6135 D WS 49511928 + 8 [fio]
259,1 11 12 3.101068699 6135 C WS 49511928 + 8 [0]
259,1 11 13 6.101267480 6135 Q WS 27599368 + 8 [fio]
259,1 11 14 6.101269216 6135 G WS 27599368 + 8 [fio]
259,1 11 15 6.101277050 6135 D WS 27599368 + 8 [fio]
259,1 11 16 6.101287956 6135 C WS 27599368 + 8 [0]
Signed-off-by: HongweiQin <glqinhongwei@gmail.com>
---
backend.c | 4 ++--
fio.h | 10 +---------
2 files changed, 3 insertions(+), 11 deletions(-)
diff --git a/backend.c b/backend.c
index 2e6a377..e20a2e0 100644
--- a/backend.c
+++ b/backend.c
@@ -439,7 +439,7 @@ static int wait_for_completions(struct thread_data *td, struct timespec *time)
if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
min_evts = 1;
- if (time && __should_check_rate(td))
+ if (time && should_check_rate(td))
fio_gettime(time, NULL);
do {
@@ -494,7 +494,7 @@ int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
requeue_io_u(td, &io_u);
} else {
sync_done:
- if (comp_time && __should_check_rate(td))
+ if (comp_time && should_check_rate(td))
fio_gettime(comp_time, NULL);
*ret = io_u_sync_complete(td, io_u);
diff --git a/fio.h b/fio.h
index 4d439d9..ee582a7 100644
--- a/fio.h
+++ b/fio.h
@@ -757,17 +757,9 @@ static inline bool option_check_rate(struct thread_data *td, enum fio_ddir ddir)
return false;
}
-static inline bool __should_check_rate(struct thread_data *td)
-{
- return (td->flags & TD_F_CHECK_RATE) != 0;
-}
-
static inline bool should_check_rate(struct thread_data *td)
{
- if (!__should_check_rate(td))
- return false;
-
- return ddir_rw_sum(td->bytes_done) != 0;
+ return (td->flags & TD_F_CHECK_RATE) != 0;
}
static inline unsigned long long td_max_bs(struct thread_data *td)
--
1.8.3.1
^ permalink raw reply related [flat|nested] 2+ messages in thread
* Re: [PATCH v2] Fix a rate limit issue.
2021-01-16 19:52 [PATCH v2] Fix a rate limit issue HongweiQin
@ 2021-01-16 20:10 ` Jens Axboe
0 siblings, 0 replies; 2+ messages in thread
From: Jens Axboe @ 2021-01-16 20:10 UTC (permalink / raw)
To: HongweiQin, fio
On 1/16/21 12:52 PM, HongweiQin wrote:
> In the current implementation, should_check_rate() returns false
> if ddir_rw_sum(td->bytes_done)==0. Therefore, a thread may violate
> the rate if iodepth*bs > rate.
>
> This patch addresses the issue by not checking td->bytes_done in should_check_rate.
Much better, thanks. I have applied it.
--
Jens Axboe
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2021-01-16 20:10 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-01-16 19:52 [PATCH v2] Fix a rate limit issue HongweiQin
2021-01-16 20:10 ` Jens Axboe
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.