Ah, true, sync -f does a syncfs.
sync without -f on an existing file is always fast. But when I strace
vim, fsync is slow (sometimes >10 seconds) because it needs to write
something.
Hopefully a developer can reply - thanks so far!
On 08-Apr-20 00:11, Chris Murphy wrote:
On Tue, Apr 7, 2020 at 2:39 PM kjansen387 <kjansen387@xxxxxxxxx> wrote:
$ grep /export /etc/fstab
UUID=8ce9e167-57ea-4cf8-8678-3049ba028c12 /export btrfs
device=/dev/sde,device=/dev/sdf 0 2
I'd use only noatime for options. There's no advantage of specifying
devices. And also fs_passno can be 0. fsck.btrfs is a no op anyway, so
it doesn't hurt anything to leave it at 2.
I've attached sysstat info of my disks. What's obvious is that 2 disks
have the load (one is written to, the other one is the mirror), and 3
are pretty idle. But, it's 2.4MB per second - that's not much!
Lots of small file writes maybe? What's iostat show for utilization?
Or vmstat for io? Hard drives of course have limited IO per second.
I've just changed the space_cache to v2, but it doesn't seem to help
much. 'sync -f /export/tmp' still takes very long at times (just took 22
seconds!)
I just did a strace on this command and it uses syncfs, not fsync. I'm
pretty sure on Btrfs this is a full filesystem sync, which is
expensive, all data and metadata. So if it's very dirty, yeah it could
take some time to flush everything to disk.
Try it without -f and you'll get fsync.
Any way we can find the cause, before I move everything into subvolumes
? I'd like to avoid that if possible. Sounds a bit overkill for 2.4MB/s
writes, and I think most of it is going to one influxdb database anyway.
From the sysrq...
[937013.794093] mysqld D 0 10400 1 0x00004000
[937013.794240] do_fsync+0x38/0x70
[937013.794253] mysqld D 0 10412 1 0x00004000
[937013.794297] do_fsync+0x38/0x70
[937013.794306] mysqld D 0 10421 1 0x00004000
[937013.794353] do_fsync+0x38/0x70
[937013.794788] WTJourn.Flusher D 0 1186978 1186954 0x00004320
[937013.794894] do_fsync+0x38/0x70
[937013.794903] ftdc D 0 1186980 1186954 0x00000320
[937013.794951] ? btrfs_create+0x200/0x200 [btrfs]
[937013.795022] influxd D 0 2082782 1 0x00004000
[937013.795086] do_fsync+0x38/0x70
[937013.795098] influxd D 0 2082804 1 0x00004000
[937013.795143] do_fsync+0x38/0x70
[937013.795191] vim D 0 2228648 2223845 0x00004000
[937013.795286] do_fsync+0x38/0x70
Quite a lot of fsync all at once.
But I'm not very good at parsing kernel output. Maybe a developer will
have input.
