Btrfs: try harder to avoid btree node splits
authorFilipe David Borba Manana <fdmanana@gmail.com>
Mon, 25 Nov 2013 03:20:46 +0000 (03:20 +0000)
committerChris Mason <clm@fb.com>
Tue, 28 Jan 2014 21:19:46 +0000 (13:19 -0800)
commit5a4267ca20d4c452a97dace4612f1dfc04147fbd
treef3dbf5aaf55c935410943ed6af031641e892b5c1
parent1b8e7e45e57cc98bf6c8ab9b3087d634636ba2e6
Btrfs: try harder to avoid btree node splits

When attempting to move items from our target leaf to its neighbor
leaves (right and left), we only need to free data_size - free_space
bytes from our leaf in order to add the new item (which has size of
data_size bytes). Therefore attempt to move items to the right and
left leaves if they have at least data_size - free_space bytes free,
instead of data_size bytes free.

After 5 runs of the following test, I got a smaller number of btree
node splits overall:

sysbench --test=fileio --file-num=512 --file-total-size=5G \
  --file-test-mode=seqwr --num-threads=512 \
   --file-block-size=8192 --max-requests=100000 --file-io-mode=sync

Before this change:
* 6171 splits (average of 5 test runs)
* 61.508Mb/sec of throughput (average of 5 test runs)

After this change:
* 6036 splits (average of 5 test runs)
* 63.533Mb/sec of throughput (average of 5 test runs)

An ideal test would not just have multiple threads/processes writing
to a file (insertion of file extent items) but also do other operations
that result in insertion of items with varied sizes, like file/directory
creations, creation of links, symlinks, xattrs, etc.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/ctree.c
This page took 0.02459 seconds and 5 git commands to generate.