Skip to contentSkip to conversation

engines/io_uring: delete fio_option_is_set() calls when submitting sqes

PATCH In review Xiaoguang Wang 2 participants 2 messages Mar 12, 2020 io_uring

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 engines/io_uring: delete fio_option_is_set() calls when submitting sqes

The fio_option_is_set() call in fio_ioring_prep() is time-consuming, which will reduce sqe's submit rate drastically. To fix this issue, add two new variables to record whether ioprio_class or ioprio_set is set. I use a simple fio job to evaluate the performance:

    fio -name=fiotest -filename=/dev/nvme0n1 -iodepth=4 -thread -rw=read
    -ioengine=io_uring -hipri=0 -sqthread_poll=0 -direct=1 -bs=4k -size=10G
    -numjobs=1 -time_based -runtime=120

Before this patch:
READ: bw=969MiB/s (1016MB/s), 969MiB/s-969MiB/s (1016MB/s-1016MB/s), io=114GiB (122GB), run=120001-120001msec

With this patch:
READ: bw=1259MiB/s (1320MB/s), 1259MiB/s-1259MiB/s (1320MB/s-1320MB/s), io=148GiB (158GB), run=120001-120001msec

---
 engines/io_uring.c | 12 ++++++++++--
 1 file changed, 10 insertions(+), 2 deletions(-)
engines/io_uring.c +10 -2
index 1efc6cf..ac57af8 100644
@@ -63,6 +63,8 @@ struct ioring_data {
 	int queued;
 	int cq_ring_off;
 	unsigned iodepth;
+	bool ioprio_class_set;
+	bool ioprio_set;
 
 	struct ioring_mmap mmap[3];
 };
@@ -233,9 +235,9 @@ static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
 		}
 		if (!td->o.odirect && o->uncached)
 			sqe->rw_flags = RWF_UNCACHED;
-		if (fio_option_is_set(&td->o, ioprio_class))
+		if (ld->ioprio_class_set)
 			sqe->ioprio = td->o.ioprio_class << 13;
-		if (fio_option_is_set(&td->o, ioprio))
+		if (ld->ioprio_set)
 			sqe->ioprio |= td->o.ioprio;
 		sqe->off = io_u->offset;
 	} else if (ddir_sync(io_u->ddir)) {
@@ -685,6 +687,12 @@ static int fio_ioring_init(struct thread_data *td)
 		td_verror(td, EINVAL, "fio_io_uring_init");
 		return 1;
 	}
+
+	if (fio_option_is_set(&td->o, ioprio_class))
+		ld->ioprio_class_set = true;
+	if (fio_option_is_set(&td->o, ioprio))
+		ld->ioprio_set = true;
+
 	return 0;
 }
 
Signature · 1 line
1.8.3.1
maintainer
#1 lore raw
Quoted 8 lines · Xiaoguang Wang

The fio_option_is_set() call in fio_ioring_prep() is time-consuming, which will reduce sqe's submit rate drastically. To fix this issue, add two new variables to record whether ioprio_class or ioprio_set is set. I use a simple fio job to evaluate the performance:

    fio -name=fiotest -filename=/dev/nvme0n1 -iodepth=4 -thread -rw=read
    -ioengine=io_uring -hipri=0 -sqthread_poll=0 -direct=1 -bs=4k -size=10G
    -numjobs=1 -time_based -runtime=120
Quoted 7 lines

Before this patch:
READ: bw=969MiB/s (1016MB/s), 969MiB/s-969MiB/s (1016MB/s-1016MB/s), io=114GiB (122GB), run=120001-120001msec

With this patch:
READ: bw=1259MiB/s (1320MB/s), 1259MiB/s-1259MiB/s (1320MB/s-1320MB/s), io=148GiB (158GB), run=120001-120001msec

Thanks, in hindsight that was pretty dumb!

Signature · 1 line
Jens Axboe