Currently we use a red black tree (rb-tree) to track the currently open
inodes of a root (in struct btrfs_root::inode_tree). This however is not
very efficient when the number of inodes is large since rb-trees are
binary trees. For example for 100K open inodes, the tree has a depth of
17. Besides that, inserting into the tree requires navigating through it
and pulling useless cache lines in the process since the red black tree
nodes are embedded within the btrfs inode - on the other hand, by being
embedded, it requires no extra memory allocations.
We can improve this by using an xarray instead, which is efficient when
indices are densely clustered (such as inode numbers), is more cache
friendly and behaves like a resizable array, with a much better search
and insertion complexity than a red black tree. This only has one small
disadvantage which is that insertion will sometimes require allocating
memory for the xarray - which may fail (not that often since it uses a
kmem_cache) - but on the other hand we can reduce the btrfs inode
structure size by 24 bytes (from 1080 down to 1056 bytes) after removing
the embedded red black tree node, which after the next patches will allow
to reduce the size of the structure to 1024 bytes, meaning we will be able
to store 4 inodes per 4K page instead of 3 inodes.
This change does a straightforward change to use an xarray, and results
in a transaction abort if we can't allocate memory for the xarray when
creating an inode - but the next patch changes things so that we don't
need to abort.
Running the following fs_mark test showed some improvements:
$ cat test.sh
#!/bin/bash
DEV=/dev/nullb0
MNT=/mnt/nullb0
MOUNT_OPTIONS="-o ssd"
FILES=100000
THREADS=$(nproc --all)
echo "performance" | \
tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
mkfs.btrfs -f $DEV
mount $MOUNT_OPTIONS $DEV $MNT
OPTS="-S 0 -L 5 -n $FILES -s 0 -t $THREADS -k"
for ((i = 1; i <= $THREADS; i++)); do
OPTS="$OPTS -d $MNT/d$i"
done
fs_mark $OPTS
umount $MNT
Before this patch:
FSUse% Count Size Files/sec App Overhead
10 1200000 0 92081.6 12505547
16 2400000 0 138222.6 13067072
23 3600000 0 148833.1 13290336
43 4800000 0 97864.7 13931248
53 6000000 0 85597.3 14384313
After this patch:
FSUse% Count Size Files/sec App Overhead
10 1200000 0 93225.1 12571078
16 2400000 0 146720.3 12805007
23 3600000 0 160626.4 13073835
46 4800000 0 116286.2 13802927
53 6000000 0 90087.9 14754892
The test was run with a release kernel config (Debian's default config).
Also capturing the insertion times into the rb tree and into the xarray,
that is measuring the duration of the old function inode_tree_add() and
the duration of the new btrfs_add_inode_to_root() function, gave the
following results (in nanoseconds):
Before this patch, inode_tree_add() execution times:
Count: 5000000
Range: 0.000 - 5536887.000; Mean: 775.674; Median: 729.000; Stddev: 4820.961
Percentiles: 90th: 1015.000; 95th: 1139.000; 99th: 1397.000
0.000 - 7.816: 40 |
7.816 - 37.858: 209 |
37.858 - 170.278: 6059 |
170.278 - 753.961: 2754890 #####################################################
753.961 - 3326.728: 2232312 ###########################################
3326.728 - 14667.018: 4366 |
14667.018 - 64652.943: 852 |
64652.943 - 284981.761: 550 |
284981.761 - 1256150.914: 221 |
1256150.914 - 5536887.000: 7 |
After this patch, btrfs_add_inode_to_root() execution times:
Count: 5000000
Range: 0.000 - 2900652.000; Mean: 272.148; Median: 241.000; Stddev: 2873.369
Percentiles: 90th: 342.000; 95th: 432.000; 99th: 572.000
0.000 - 7.264: 104 |
7.264 - 33.145: 352 |
33.145 - 140.081: 109606 #
140.081 - 581.930: 4840090 #####################################################
581.930 - 2407.590: 43532 |
2407.590 - 9950.979: 2245 |
9950.979 - 41119.278: 514 |
41119.278 - 169902.616: 155 |
169902.616 - 702018.539: 47 |
702018.539 - 2900652.000: 9 |
Average, percentiles, standard deviation, etc, are all much better.
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
310b2f5d5a