coalescing in polling mode in 4.9
Hello,
Enabling interrupt coalescing in nvme device, kernel 4.9 significantly reduces performance when using polling mode. When I enable coalescing, IOPS drops from 100K to 35K and latency jumps from 7 usec to 25 usec.
Shouldn't we expect performance boost in polling mode when interrupt coalescing enabled?
Device is Intel DC P3600
Coalescing enabled: nvme set-feature /dev/nvme0 -f 8 -v 0x00ffff fio-2.16 file:
[global]
iodepth=1
direct=1
ioengine=pvsync2
hiprigroup_reportingtime_based
blocksize=4k
norandommap=1
[job1]
rw=read
filename=/dev/nvme0n1
name=raw=sequential-read
numjobs=1
runtime=60
Thanks
Appreciate if someone could help me to understand why do we need interrupts in polling mode, and performance drops when interrupts coalescing enabled.
Thanks
Quoted 34 lines · On Fri, 2 Feb 2018 00:10:28 -0800 Alex Nln <[email protected]> wrote:
Hello,
Enabling interrupt coalescing in nvme device, kernel 4.9 significantly reduces performance when using polling mode. When I enable coalescing, IOPS drops from 100K to 35K and latency jumps from 7 usec to 25 usec.
Shouldn't we expect performance boost in polling mode when interrupt coalescing enabled?
Device is Intel DC P3600
Coalescing enabled: nvme set-feature/dev/nvme0-f 8 -v0x00fffffio-2.16 file:
[global]
iodepth=1
direct=1
ioengine=pvsync2
hiprigroup_reportingtime_based
blocksize=4k
norandommap=1[job1]
rw=read
filename=/dev/nvme0n1
name=raw=sequential-read
numjobs=1
runtime=60Thanks
Sagi Grimberg
maintainerQuoted 3 lines
Appreciate if someone could help me to understand why do we need interrupts in polling mode, and performance drops when interrupts coalescing enabled.
I didn't try with this device. I did try on p3700 in the past and this did not happen.
Quoted 10 lines · Alex Nln
Enabling interrupt coalescing in nvme device, kernel 4.9 significantly reduces performance when using polling mode. When I enable coalescing, IOPS drops from 100K to 35K and latency jumps from 7 usec to 25 usec.
Shouldn't we expect performance boost in polling mode when interrupt coalescing enabled?
Device is Intel DC P3600
That's a pretty low latency for this device. Are you running reads to unmapped blocks?
I've seen the phenomenom occur where higher coalescing settings worsens performance. That usually means the the polling thread exceeded its time:need_resched() returns true, so the tasks schedules out and relies on interrupts, which have been throttled. Maybe a hybrid polling would be better for this.
Hi Alex,
I got into similar problem not long ago. With coalescing enabled, some I/Os took very long. Every time need_reshed() returns true, io_schedule() makes task go to sleep as its state is previous set as non-interruptible.I handled this by setting task state as running, and release the cpu. Diff is attached below, you may give it a try.
Hi Keith,
Should below not be the default way to handle classic-polling case i.e. do not rely on interrupts at all? For sleep during polling-path (which reduces cpu utilization) we anyway have hybrid polling path. If this approach seems acceptable, I can further refactor the code and submit a patch. I look forward to your suggestions.
fs/block_dev.c +0 -0
@@ -236,9 +236,13 @@ __blkdev_direct_IO_simple(struct kiocb *iocb,
struct iov_iter *iter,
set_current_state(TASK_UNINTERRUPTIBLE);
if (!READ_ONCE(bio.bi_private))
break;
- if (!(iocb->ki_flags & IOCB_HIPRI) ||
- !blk_poll(bdev_get_queue(bdev), qc))
+ if (!(iocb->ki_flags & IOCB_HIPRI))
io_schedule();+ else if (!blk_poll(bdev_get_queue(bdev), qc)) {
+ if(need_resched())
+ set_current_state(TASK_RUNNING);
+ io_schedule();
+ }
}
__set_current_state(TASK_RUNNING);Quoted 29 lines · Keith Busch
Quoted 10 lines · Alex Nln
Enabling interrupt coalescing in nvme device, kernel 4.9 significantly reduces performance when using polling mode. When I enable coalescing, IOPS drops from 100K to 35K and latency jumps from 7 usec to 25 usec.
Shouldn't we expect performance boost in polling mode when interrupt coalescing enabled?
Device is Intel DC P3600
That's a pretty low latency for this device. Are you running reads to unmapped blocks?
I've seen the phenomenom occur where higher coalescing settings worsens performance. That usually means the the polling thread exceeded its time:
need_resched()returns true, so the tasks schedules out and relies on interrupts, which have been throttled. Maybe a hybrid polling would be better for this._______________________________________________
Linux-nvme mailing list
Linux-nvme at lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme
Quoted 6 lines · On Mon, 5 Feb 2018 08:02:53 -0700 Keith Busch <[email protected]> wrote:
Quoted 5 lines · Alex Nln
Enabling interrupt coalescing in nvme device, kernel 4.9 significantly reduces performance when using polling mode. When I enable coalescing, IOPS drops from 100K to 35K and latency jumps from 7 usec to 25 usec.
Quoted 8 lines
Quoted 5 lines
Shouldn't we expect performance boost in polling mode when interrupt coalescing enabled?
Device is Intel DC P3600
That's a pretty low latency for this device. Are you running reads to unmapped blocks?
These are synchronous sequential reads, random reads' latency is higher.
Quoted 7 lines
I've seen the phenomenom occur where higher coalescing settings worsens performance. That usually means the the polling thread exceeded its time:
need_resched()returns true, so the tasks schedules out and relies on interrupts, which have been throttled. Maybe a hybrid polling would be better for this.
It looks like the eventual call to need_resched() is a result of low performance. When I set coalescing to 0x00FFFF there are about 160 nvme interrupts/sec, but ~35K IOPS, so clearly only a fractions of CQ entries are processed in interrupt context, most of the CQ entries still processed by polling. Somehow device delays posting completion entries when coalescing is enabled. Is it possible?
Thanks for help
Hi Nitesh,
Quoted 7 lines · On Mon, 05 Feb 2018 21:12:12 +0530 Nitesh <[email protected]> wrote:
Hi Alex,
I got into similar problem not long ago. With coalescing enabled, some I/Os took very long. Every timeneed_reshed()returns true,io_schedule()makes task go to sleep as its state is previous set as non-interruptible.I handled this by setting task state as running, and release the cpu. Diff is attached below, you may give it a try.
I tested this patch on kernel 4.9 and it solves the problem. Thanks a lot
Quoted 23 lines
fs/block_dev.c +6
-2@@ -236,9 +236,13 @@ __blkdev_direct_IO_simple(struct kiocb *iocb,struct iov_iter *iter,set_current_state (TASK_UNINTERRUPTIBLE );if (!READ_ONCE (bio.bi_private))break ; -if (!(iocb->ki_flags &IOCB_HIPRI ) || - !blk_poll (bdev_get_queue (bdev), qc)) +if (!(iocb->ki_flags &IOCB_HIPRI ))io_schedule (); +else if (!blk_poll (bdev_get_queue (bdev), qc)) { +if (need_resched ()) +set_current_state (TASK_RUNNING ); +io_schedule (); + } }__set_current_state (TASK_RUNNING );
Quoted 10 lines · Nitesh
fs/block_dev.c +0
-1@@ -236,9 +236,13 @@ __blkdev_direct_IO_simple(struct kiocb *iocb, structiov_iter *iter,set_current_state (TASK_UNINTERRUPTIBLE );if (!READ_ONCE (bio.bi_private))break ; - if (!(iocb->ki_flags & IOCB_HIPRI) ||
fs/block_dev.c __blkdev_direct_IO_simple(struct kiocb *iocb, struct · line 248
Quoted 10 lines
fs/block_dev.c +6
-1- !blk_poll(bdev_get_queue(bdev), qc)) +if (!(iocb->ki_flags &IOCB_HIPRI ))io_schedule (); +else if (!blk_poll (bdev_get_queue (bdev), qc)) { +if (need_resched ()) +set_current_state (TASK_RUNNING ); +io_schedule (); + } }__set_current_state (TASK_RUNNING );
Yah, I think this looks good. Do you want to send this as a proper patch to linux-block mailing list for consideration?
Hi Keith,
I will send proper patch.
Quoted 27 lines · Keith Busch
Quoted 10 lines · Nitesh
fs/block_dev.c +0
-1@@ -236,9 +236,13 @@ __blkdev_direct_IO_simple(struct kiocb *iocb, structiov_iter *iter,set_current_state (TASK_UNINTERRUPTIBLE );if (!READ_ONCE (bio.bi_private))break ; - if (!(iocb->ki_flags & IOCB_HIPRI) ||fs/block_dev.c __blkdev_direct_IO_simple(struct kiocb *iocb, struct · line 248
Quoted 10 lines
fs/block_dev.c +6
-1- !blk_poll(bdev_get_queue(bdev), qc)) +if (!(iocb->ki_flags &IOCB_HIPRI ))io_schedule (); +else if (!blk_poll (bdev_get_queue (bdev), qc)) { +if (need_resched ()) +set_current_state (TASK_RUNNING ); +io_schedule (); + } }__set_current_state (TASK_RUNNING );Yah, I think this looks good. Do you want to send this as a proper patch to linux-block mailing list for consideration?
Hi Alex,
Did you observe any benefit in polling-performance i.e. with interrupt-coalescing vs without interrupt-coalescing. I'd appreciate if you could share the performance/latency data.
Thanks,
Nitesh
Quoted 48 lines · Alex Nln
Hi Nitesh,
Quoted 7 lines · On Mon, 05 Feb 2018 21:12:12 +0530 Nitesh <[email protected]> wrote:
Hi Alex,
I got into similar problem not long ago. With coalescing enabled, some I/Os took very long. Every timeneed_reshed()returns true,io_schedule()makes task go to sleep as its state is previous set as non-interruptible.I handled this by setting task state as running, and release the cpu. Diff is attached below, you may give it a try.I tested this patch on kernel 4.9 and it solves the problem. Thanks a lot
Quoted 13 lines
fs/block_dev.c +0
-1@@ -236,9 +236,13 @@ __blkdev_direct_IO_simple(struct kiocb *iocb,struct iov_iter *iter,set_current_state (TASK_UNINTERRUPTIBLE );if (!READ_ONCE (bio.bi_private))break ; - if (!(iocb->ki_flags & IOCB_HIPRI) ||fs/block_dev.c __blkdev_direct_IO_simple(struct kiocb *iocb, · line 248
Quoted 10 lines
fs/block_dev.c +6
-1- !blk_poll(bdev_get_queue(bdev), qc)) +if (!(iocb->ki_flags &IOCB_HIPRI ))io_schedule (); +else if (!blk_poll (bdev_get_queue (bdev), qc)) { +if (need_resched ()) +set_current_state (TASK_RUNNING ); +io_schedule (); + } }__set_current_state (TASK_RUNNING );_______________________________________________
Linux-nvme mailing list
Linux-nvme at lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme
Hi Nitesh,
Yes, the benefits are clear, in terms of IOPS and latency. See below.
coalescing iops clat intr/s
0x0000 108710 8.65 108711
0xff01 113081 8.28 56541
0xff03 114955 8.14 28738
0xff07 116318 8.04 14539
0xff0f 116470 8.02 7279
0xff1f 116908 8.00 3653
0xff3f 116833 8.00 1825
0xff7f 116718 8.00 911
0xffff 116962 8.00 456It is quite strange model of IO processing: polling with interrupts. Shouldn't these two be mutual exclusive? Why not to disable interrupts at all on a queue while polling?
Thanks,
Alex
device Intel DC P3600, kernel 4.9.76 + Nitesh patch that fixes polling, fio 2.21:
[global]
iodepth=1
direct=1
ioengine=pvsync2
hiprigroup_reportingtime_based
norandommap=1
[job1]
rw=read
filename=/dev/nvme0n1
name=raw=sequential-read
numjobs=1
runtime=60
Quoted 58 lines · On Sat, 10 Feb 2018 04:07:08 +0530 Nitesh Shetty <[email protected]> wrote:
Hi Alex,
Did you observe any benefit in polling-performance i.e. with interrupt-coalescing vs without interrupt-coalescing. I'd appreciate if you could share the performance/latency data.
Thanks,
NiteshQuoted 48 lines · Alex Nln
Hi Nitesh,
Quoted 7 lines · On Mon, 05 Feb 2018 21:12:12 +0530 Nitesh <[email protected]> wrote:
Hi Alex,
I got into similar problem not long ago. With coalescing enabled, some I/Os took very long. Every timeneed_reshed()returns true,io_schedule()makes task go to sleep as its state is previous set as non-interruptible.I handled this by setting task state as running, and release the cpu. Diff is attached below, you may give it a try.I tested this patch on kernel 4.9 and it solves the problem. Thanks a lot
Quoted 13 lines
fs/block_dev.c +0
-1@@ -236,9 +236,13 @@ __blkdev_direct_IO_simple(struct kiocb *iocb,struct iov_iter *iter,set_current_state (TASK_UNINTERRUPTIBLE );if (!READ_ONCE (bio.bi_private))break ; - if (!(iocb->ki_flags & IOCB_HIPRI) ||fs/block_dev.c __blkdev_direct_IO_simple(struct kiocb *iocb, · line 248
Quoted 10 lines
fs/block_dev.c +6
-1- !blk_poll(bdev_get_queue(bdev), qc)) +if (!(iocb->ki_flags &IOCB_HIPRI ))io_schedule (); +else if (!blk_poll (bdev_get_queue (bdev), qc)) { +if (need_resched ()) +set_current_state (TASK_RUNNING ); +io_schedule (); + } }__set_current_state (TASK_RUNNING );_______________________________________________
Linux-nvme mailing list
Linux-nvme at lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme
Quoted 3 lines · Alex Nln
It is quite strange model of IO processing: polling with interrupts. Shouldn't these two be mutual exclusive? Why not to disable interrupts at all on a queue while polling?
That's a valid point. There are some limitations preventing this at the moment. The nvme driver currently does not have a way to know the command it is submitting is intended to be polled, and if it did, we still need to determine a policy for creating special queue or queues that have interrupts disabled and how we register these with the block layer.