[3/3] btrfs: do not set the full sync flag on the inode during page release
diff mbox series

Message ID 20200722112901.15626-1-fdmanana@kernel.org
State New
Headers show
Series
  • [1/3] btrfs: fix race between page release and a fast fsync
Related show

Commit Message

Filipe Manana July 22, 2020, 11:29 a.m. UTC
From: Filipe Manana <fdmanana@suse.com>

When removing an extent map at try_release_extent_mapping(), called through
the page release callback (btrfs_releasepage()), we always set the full
sync flag on the inode, which forces the next fsync to use a slower code
path.

This hurts performance for workloads that dirty an amount of data that
exceeds or is very close to the system's RAM memory and do frequent fsync
operations (like database servers can for example). In particular if there
are concurrent fsyncs against different files, by falling back to a full
fsync we do a lot more checksum lookups in the checksums btree, as we do
it for all the extents created in the current transaction, instead of only
the new ones since the last fsync. These checksums lookups not only take
some time but, more importantly, they also cause contention on the
checksums btree locks due to the concurrency with checksum insertions in
the btree by ordered extents from other inodes.

We actually don't need to set the full sync flag on the inode, because we
only remove extent maps that are in the list of modified extents if they
were created in a past transaction, in which case an fsync skips them as
it's pointless to log them. So stop setting the full fsync flag on the
inode whenever we remove an extent map.

This patch is part of a patchset that consists of 3 patches, which have
the following subjects:

1/3 btrfs: fix race between page release and a fast fsync
2/3 btrfs: release old extent maps during page release
3/3 btrfs: do not set the full sync flag on the inode during page release

Performance tests were ran against a branch (misc-next) containing the
whole patchset. The test exercises a workload where there are multiple
processes writing to files and fsyncing them (each writing and fsyncing
its own file), and in total the amount of data dirtied ranges from 2x to
4x the system's RAM memory (16Gb), so that the page release callback is
invoked frequently.

The following script, using fio, was used to perform the tests:

  $ cat test-fsync.sh
  #!/bin/bash

  DEV=/dev/sdk
  MNT=/mnt/sdk
  MOUNT_OPTIONS="-o ssd"
  MKFS_OPTIONS="-d single -m single"

  if [ $# -ne 3 ]; then
      echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ"
      exit 1
  fi

  NUM_JOBS=$1
  FILE_SIZE=$2
  FSYNC_FREQ=$3

  cat <<EOF > /tmp/fio-job.ini
  [writers]
  rw=write
  fsync=$FSYNC_FREQ
  fallocate=none
  group_reporting=1
  direct=0
  bs=64k
  ioengine=sync
  size=$FILE_SIZE
  directory=$MNT
  numjobs=$NUM_JOBS
  thread
  EOF

  echo "Using config:"
  echo
  cat /tmp/fio-job.ini
  echo

  mkfs.btrfs -f $MKFS_OPTIONS $DEV &> /dev/null
  mount $MOUNT_OPTIONS $DEV $MNT
  fio /tmp/fio-job.ini
  umount $MNT

The tests were performed for different numbers of jobs, file sizes and
fsync frequency. A qemu VM using kvm was used, with 8 cores (the host has
12 cores, with cpu governance set to performance mode on all cores), 16Gb
of ram (the host has 64Gb) and using a NVMe device directly (without an
intermediary filesystem in the host). While running the tests, the host
was not used for anything else, to avoid disturbing the tests.

The obtained results were the following, and the last line printed by
fio is pasted (includes aggregated throughput and test run time).

    *****************************************************
    ****     1 job, 32Gb file, fsync frequency 1     ****
    *****************************************************

Before patchset:

WRITE: bw=29.1MiB/s (30.5MB/s), 29.1MiB/s-29.1MiB/s (30.5MB/s-30.5MB/s), io=32.0GiB (34.4GB), run=1127557-1127557msec

After patchset:

WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=32.0GiB (34.4GB), run=1119042-1119042msec
(+0.7% throughput, -0.8% run time)

    *****************************************************
    ****     2 jobs, 16Gb files, fsync frequency 1   ****
    *****************************************************

Before patchset:

WRITE: bw=33.5MiB/s (35.1MB/s), 33.5MiB/s-33.5MiB/s (35.1MB/s-35.1MB/s), io=32.0GiB (34.4GB), run=979000-979000msec

After patchset:

WRITE: bw=39.9MiB/s (41.8MB/s), 39.9MiB/s-39.9MiB/s (41.8MB/s-41.8MB/s), io=32.0GiB (34.4GB), run=821283-821283msec
(+19.1% throughput, -16.1% runtime)

    *****************************************************
    ****     4 jobs, 8Gb files, fsync frequency 1    ****
    *****************************************************

Before patchset:

WRITE: bw=52.1MiB/s (54.6MB/s), 52.1MiB/s-52.1MiB/s (54.6MB/s-54.6MB/s), io=32.0GiB (34.4GB), run=629130-629130msec

After patchset:

WRITE: bw=71.8MiB/s (75.3MB/s), 71.8MiB/s-71.8MiB/s (75.3MB/s-75.3MB/s), io=32.0GiB (34.4GB), run=456357-456357msec
(+37.8% throughput, -27.5% runtime)

    *****************************************************
    ****     8 jobs, 4Gb files, fsync frequency 1    ****
    *****************************************************

Before patchset:

WRITE: bw=76.1MiB/s (79.8MB/s), 76.1MiB/s-76.1MiB/s (79.8MB/s-79.8MB/s), io=32.0GiB (34.4GB), run=430708-430708msec

After patchset:

WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=245458-245458msec
(+74.7% throughput, -43.0% run time)

    *****************************************************
    ****    16 jobs, 2Gb files, fsync frequency 1    ****
    *****************************************************

Before patchset:

WRITE: bw=74.7MiB/s (78.3MB/s), 74.7MiB/s-74.7MiB/s (78.3MB/s-78.3MB/s), io=32.0GiB (34.4GB), run=438625-438625msec

After patchset:

WRITE: bw=184MiB/s (193MB/s), 184MiB/s-184MiB/s (193MB/s-193MB/s), io=32.0GiB (34.4GB), run=177864-177864msec
(+146.3% throughput, -59.5% run time)

    *****************************************************
    ****    32 jobs, 2Gb files, fsync frequency 1    ****
    *****************************************************

Before patchset:

WRITE: bw=72.6MiB/s (76.1MB/s), 72.6MiB/s-72.6MiB/s (76.1MB/s-76.1MB/s), io=64.0GiB (68.7GB), run=902615-902615msec

After patchset:

WRITE: bw=227MiB/s (238MB/s), 227MiB/s-227MiB/s (238MB/s-238MB/s), io=64.0GiB (68.7GB), run=288936-288936msec
(+212.7% throughput, -68.0% run time)

    *****************************************************
    ****    64 jobs, 1Gb files, fsync frequency 1    ****
    *****************************************************

Before patchset:

WRITE: bw=98.8MiB/s (104MB/s), 98.8MiB/s-98.8MiB/s (104MB/s-104MB/s), io=64.0GiB (68.7GB), run=663126-663126msec

After patchset:

WRITE: bw=294MiB/s (308MB/s), 294MiB/s-294MiB/s (308MB/s-308MB/s), io=64.0GiB (68.7GB), run=222940-222940msec
(+197.6% throughput, -66.4% run time)

Signed-off-by: Filipe Manana <fdmanana@suse.com>
---
 fs/btrfs/extent_io.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

Comments

David Sterba July 23, 2020, 5:18 p.m. UTC | #1
On Wed, Jul 22, 2020 at 12:29:01PM +0100, fdmanana@kernel.org wrote:
> Before patchset:
> 
> WRITE: bw=98.8MiB/s (104MB/s), 98.8MiB/s-98.8MiB/s (104MB/s-104MB/s), io=64.0GiB (68.7GB), run=663126-663126msec
> 
> After patchset:
> 
> WRITE: bw=294MiB/s (308MB/s), 294MiB/s-294MiB/s (308MB/s-308MB/s), io=64.0GiB (68.7GB), run=222940-222940msec
> (+197.6% throughput, -66.4% run time)

That's really great, thank you very much!

Patch
diff mbox series

diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c
index 5eab129e6eb0..f6837a6fe464 100644
--- a/fs/btrfs/extent_io.c
+++ b/fs/btrfs/extent_io.c
@@ -4530,8 +4530,14 @@  int try_release_extent_mapping(struct page *page, gfp_t mask)
 			if (em->generation >= cur_gen)
 				goto next;
 remove_em:
-			set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
-				&btrfs_inode->runtime_flags);
+			/*
+			 * We only remove extent maps that are not in the list of
+			 * modified extents or that are in the list but with a
+			 * generation lower then the current generation, so there
+			 * is no need to set the full fsync flag on the inode (it
+			 * hurts the fsync performance for workloads with a data
+			 * size that exceeds or is close to the system's memory).
+			 */
 			remove_extent_mapping(map, em);
 			/* once for the rb tree */
 			free_extent_map(em);