The aim of this patch is to reduce the fragmentation of block groups
under certain unhappy workloads. It is particularly effective when the
size of extents correlates with their lifetime, which is something we
have observed causing fragmentation in the fleet at Meta.
This patch categorizes extents into size classes:
- x < 128KiB: "small"
- 128KiB < x < 8MiB: "medium"
- x > 8MiB: "large"
and as much as possible reduces allocations of extents into block groups
that don't match the size class. This takes advantage of any (possible)
correlation between size and lifetime and also leaves behind predictable
re-usable gaps when extents are freed; small writes don't gum up bigger
holes.
Size classes are implemented in the following way:
- Mark each new block group with a size class of the first allocation
that goes into it.
- Add two new passes to ffe: "unset size class" and "wrong size class".
First, try only matching block groups, then try unset ones, then allow
allocation of new ones, and finally allow mismatched block groups.
- Filtering is done just by skipping inappropriate ones, there is no
special size class indexing.
Other solutions I considered were:
- A best fit allocator with an rb-tree. This worked well, as small
writes didn't leak big holes from large freed extents, but led to
regressions in ffe and write performance due to lock contention on
the rb-tree with every allocation possibly updating it in parallel.
Perhaps something clever could be done to do the updates in the
background while being "right enough".
- A fixed size "working set". This prevents freeing an extent
drastically changing where writes currently land, and seems like a
good option too. Doesn't take advantage of size in any way.
- The same size class idea, but implemented with xarray marks. This
turned out to be slower than looping the linked list and skipping
wrong block groups, and is also less flexible since we must have only
3 size classes (max #marks). With the current approach we can have as
many as we like.
Performance testing was done via: https://github.com/josefbacik/fsperf
Of particular relevance are the new fragmentation specific tests.
A brief summary of the testing results:
- Neutral results on existing tests. There are some minor regressions
and improvements here and there, but nothing that truly stands out as
notable.
- Improvement on new tests where size class and extent lifetime are
correlated. Fragmentation in these cases is completely eliminated
and write performance is generally a little better. There is also
significant improvement where extent sizes are just a bit larger than
the size class boundaries.
- Regression on one new tests: where the allocations are sized
intentionally a hair under the borders of the size classes. Results
are neutral on the test that intentionally attacks this new scheme by
mixing extent size and lifetime.
The full dump of the performance results can be found here:
https://bur.io/fsperf/size-class-2022-11-15.txt
(there are ANSI escape codes, so best to curl and view in terminal)
Here is a snippet from the full results for a new test which mixes
buffered writes appending to a long lived set of files and large short
lived fallocates:
bufferedappendvsfallocate results
metric baseline current stdev diff
======================================================================================
avg_commit_ms 31.13 29.20 2.67 -6.22%
bg_count 14 15.60 0 11.43%
commits 11.10 12.20 0.32 9.91%
elapsed 27.30 26.40 2.98 -3.30%
end_state_mount_ns 11122551.90 10635118.90 851143.04 -4.38%
end_state_umount_ns 1.36e+09 1.35e+09 12248056.65 -1.07%
find_free_extent_calls 116244.30 114354.30 964.56 -1.63%
find_free_extent_ns_max 599507.20 1047168.20 103337.08 74.67%
find_free_extent_ns_mean 3607.19 3672.11 101.20 1.80%
find_free_extent_ns_min 500 512 6.67 2.40%
find_free_extent_ns_p50 2848 2876 37.65 0.98%
find_free_extent_ns_p95 4916 5000 75.45 1.71%
find_free_extent_ns_p99 20734.49 20920.48 1670.93 0.90%
frag_pct_max 61.67 0 8.05 -100.00%
frag_pct_mean 43.59 0 6.10 -100.00%
frag_pct_min 25.91 0 16.60 -100.00%
frag_pct_p50 42.53 0 7.25 -100.00%
frag_pct_p95 61.67 0 8.05 -100.00%
frag_pct_p99 61.67 0 8.05 -100.00%
fragmented_bg_count 6.10 0 1.45 -100.00%
max_commit_ms 49.80 46 5.37 -7.63%
sys_cpu 2.59 2.62 0.29 1.39%
write_bw_bytes 1.62e+08 1.68e+08 17975843.50 3.23%
write_clat_ns_mean 57426.39 54475.95 2292.72 -5.14%
write_clat_ns_p50 46950.40 42905.60 2101.35 -8.62%
write_clat_ns_p99 148070.40 143769.60 2115.17 -2.90%
write_io_kbytes 4194304 4194304 0 0.00%
write_iops 2476.15 2556.10 274.29 3.23%
write_lat_ns_max 2101667.60 2251129.50 370556.59 7.11%
write_lat_ns_mean 59374.91 55682.00 2523.09 -6.22%
write_lat_ns_min 17353.10 16250 1646.08 -6.36%
There are some mixed improvements/regressions in most metrics along with
an elimination of fragmentation in this workload.
On the balance, the drastic 1->0 improvement in the happy cases seems
worth the mix of regressions and improvements we do observe.
Some considerations for future work:
- Experimenting with more size classes
- More hinting/search ordering work to approximate a best-fit allocator
Signed-off-by: Boris Burkov <boris@bur.io>
Signed-off-by: David Sterba <dsterba@suse.com>
find_free_extent is a complicated function. It consists (at least) of:
- a hint that jumps into the middle of a for loop macro
- a middle loop trying every raid level
- an outer loop ascending through ffe loop levels
- complicated logic for skipping some of those ffe loop levels
- multiple underlying in-bg allocators (zoned, cluster, no cluster)
Which is all to say that more tracing is helpful for debugging its
behavior. Add two new tracepoints: at the entrance to the block_groups
loop (hit for every raid level and every ffe_ctl loop) and at the point
we seriously consider a block_group for allocation. This way we can see
the whole path through the algorithm, including hints, multiple loops,
etc.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Boris Burkov <boris@bur.io>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The allocator tracepoints currently have a pile of values from ffe_ctl.
In modifying the allocator and adding more tracepoints, I found myself
adding to the already long argument list of the tracepoints. It makes it
a lot simpler to just send in the ffe_ctl itself.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Boris Burkov <boris@bur.io>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
These belong in extent-tree.h, they were missed because they were not
grouped with the other extent-tree.c prototypes.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Move all the extent tree related prototypes to extent-tree.h out of
ctree.h, and then go include it everywhere needed so everything
compiles.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
btrfs_delayed_extent_op can be packed in a better way, it's 40 bytes now
and has 8 unused bytes. Reducing the level type to u8 makes it possible
to squeeze it to the padding byte after key. The bitfields were switched
to bool as there's space to store the full byte without increasing the
whole structure, besides that the generated assembly is smaller.
struct btrfs_delayed_extent_op {
struct btrfs_disk_key key; /* 0 17 */
u8 level; /* 17 1 */
bool update_key; /* 18 1 */
bool update_flags; /* 19 1 */
bool is_data; /* 20 1 */
/* XXX 3 bytes hole, try to pack */
u64 flags_to_set; /* 24 8 */
/* size: 32, cachelines: 1, members: 6 */
/* sum members: 29, holes: 1, sum holes: 3 */
/* last cacheline: 32 bytes */
};
The final size is 32 bytes which gives +26 object per slab page.
text data bss dec hex filename
938811 43670 23144 1005625 f5839 fs/btrfs/btrfs.ko.before
938747 43670 23144 1005561 f57f9 fs/btrfs/btrfs.ko.after
Signed-off-by: David Sterba <dsterba@suse.com>
btrfs_qgroup_account_extents().
The new btrfs_qgroup_account_extents() function should be called in
btrfs_commit_transaction() and it will update all the qgroup according
to delayed_ref_root->dirty_extent_root.
The new function can handle both normal operation during
commit_transaction() or in rescan in a unified method with clearer
logic.
Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
