btrfs quota groups (qgroups) are a compelling feature of btrfs that
allow flexible control for limiting subvolume data and metadata usage.
However, due to btrfs's high level decision to tradeoff snapshot
performance against ref-counting performance, qgroups suffer from
non-trivial performance issues that make them unattractive in certain
workloads. Particularly, frequent backref walking during writes and
during commits can make operations increasingly expensive as the number
of snapshots scales up. For that reason, we have never been able to
commit to using qgroups in production at Meta, despite significant
interest from people running container workloads, where we would benefit
from protecting the rest of the host from a buggy application in a
container running away with disk usage. This patch series introduces a simplified version of qgroups called
simple quotas (squotas) which never computes global reference counts
for extents, and thus has similar performance characteristics to normal,
quotas disabled, btrfs. The "trick" is that in simple quotas mode, we
account all extents permanently to the subvolume in which they were
originally created. That allows us to make all accounting 1:1 with
extent item lifetime, removing the need to walk backrefs. However, this sacrifices the ability to compute shared vs. exclusive usage. It also
results in counter-intuitive, though still predictable and simple,
accounting in the cases where an original extent is removed while a
shared copy still exists. Qgroups is able to detect that case and count
the remaining copy as an exclusive owner, while squotas is not. As a
result, squotas works best when the original extent is immutable and
outlives any clones.
==Format Change==
In order to track the original creating subvolume of a data extent in
the face of reflinks, it is necessary to add additional accounting to
the extent item. To save space, this is done with a new inline ref item.
However, the downside of this approach is that it makes enabling squota
an incompat change, denoted by the new incompat bit SIMPLE_QUOTA. When
this bit is set and quotas are enabled, new extent items get the extra
accounting, and freed extent items check for the accounting to find
their creating subvolume. In addition, 1:1 with this incompat bit,
the quota status item now tracks a "quota enablement generation" needed
for properly handling deleting extents with predate enablement.
==API==
Squotas reuses the api of qgroups. The only difference is that when you
enable quotas via `btrfs quota enable`, you pass the `--simple` flag.
Squotas will always report exclusive == shared for each qgroup. Squotas
deal with extent_item/metadata_item sizes and thus do not do anything
special with compression. Squotas also introduce auto inheritance for
nested subvols. The API is documented more fully in the documentation
patches in btrfs-progs.
==Testing methodology==
Using updated btrfs-progs and fstests (relevant matching patch sets to
be sent ASAP)
btrfs-progs: https://github.com/boryas/btrfs-progs/tree/squota-progs
fstests: https://github.com/boryas/fstests/tree/squota-test
I ran '-g auto' on fstests on the following configurations:
1a) baseline kernel/progs/fstests.
1b) squota kernel baseline progs/fstests.
2a) baseline kernel/progs/fstests. fstests configured to mkfs with quota
2b) squota kernel/progs/fstests. fstests configured to mkfs with squota
I compared 1a against 1b and 2a against 2b and detected no regressions.
2a/2b both exhibit regressions against 1a/1b which are largely issues
with quota reservations in various complicated cases. I intend to run
those down in the future, but they are not simple quota specific, as
they are already broken with plain qgroups.
==Performance Testing==
I measured the performance of the change using fsperf. I ran with 3
configurations using the squota kernel:
- plain mkfs
- qgroup mkfs
- squota mkfs
And added a new performance test which creates 1000 files in a subvol,
creates 100 snapshots of that subvol, then unshares extents in files in
the snapshots. I measured write performance with fio and btrfs commit
critical section performance side effects with bpftrace on
'wait_current_trans'.
The results for the test which measures unshare perf (unshare.py) with
qgroup and squota compared to the baseline:
group test results
unshare results
metric baseline current stdev diff
========================================================================================
avg_commit_ms 162.13 285.75 3.14 76.24%
bg_count 16 16 0 0.00%
commits 378.20 379 1.92 0.21%
elapsed 201.40 270.40 1.34 34.26%
end_state_mount_ns 26036211.60 26004593.60 2281065.40 -0.12%
end_state_umount_ns 2.45e+09 2.55e+09 20740154.41 3.93%
max_commit_ms 425.80 594 53.34 39.50%
sys_cpu 0.10 0.06 0.06 -42.15%
wait_current_trans_calls 2945.60 3405.20 47.08 15.60%
wait_current_trans_ns_max 1.56e+08 3.43e+08 32659393.25 120.07%
wait_current_trans_ns_mean 1974875.35 28588482.55 1557588.84 1347.61%
wait_current_trans_ns_min 232 232 25.88 0.00%
wait_current_trans_ns_p50 718 740 22.80 3.06%
wait_current_trans_ns_p95 7711770.20 2.21e+08 17241032.09 2761.19%
wait_current_trans_ns_p99 67744932.29 2.68e+08 41275815.87 295.16%
write_bw_bytes 653008.80 486344.40 4209.91 -25.52%
write_clat_ns_mean 6251404.78 8406837.89 39779.15 34.48%
write_clat_ns_p50 1656422.40 1643315.20 27415.68 -0.79%
write_clat_ns_p99 1.90e+08 3.20e+08 2097152 68.62%
write_io_kbytes 128000 128000 0 0.00%
write_iops 159.43 118.74 1.03 -25.52%
write_lat_ns_max 7.06e+08 9.80e+08 47324816.61 38.88%
write_lat_ns_mean 6251503.06 8406936.06 39780.83 34.48%
write_lat_ns_min 3354 4648 616.06 38.58%
squota test results
unshare results
metric baseline current stdev diff
========================================================================================
avg_commit_ms 162.13 164.16 3.14 1.25%
bg_count 16 0 0 -100.00%
commits 378.20 380.80 1.92 0.69%
elapsed 201.40 208.20 1.34 3.38%
end_state_mount_ns 26036211.60 25840729.60 2281065.40 -0.75%
end_state_umount_ns 2.45e+09 3.01e+09 20740154.41 22.80%
max_commit_ms 425.80 415.80 53.34 -2.35%
sys_cpu 0.10 0.08 0.06 -23.36%
wait_current_trans_calls 2945.60 2981.60 47.08 1.22%
wait_current_trans_ns_max 1.56e+08 1.12e+08 32659393.25 -27.86%
wait_current_trans_ns_mean 1974875.35 1064734.76 1557588.84 -46.09%
wait_current_trans_ns_min 232 238 25.88 2.59%
wait_current_trans_ns_p50 718 746 22.80 3.90%
wait_current_trans_ns_p95 7711770.20 1567.60 17241032.09 -99.98%
wait_current_trans_ns_p99 67744932.29 49880514.27 41275815.87 -26.37%
write_bw_bytes 653008.80 631256 4209.91 -3.33%
write_clat_ns_mean 6251404.78 6476816.06 39779.15 3.61%
write_clat_ns_p50 1656422.40 1581056 27415.68 -4.55%
write_clat_ns_p99 1.90e+08 1.94e+08 2097152 2.21%
write_io_kbytes 128000 128000 0 0.00%
write_iops 159.43 154.12 1.03 -3.33%
write_lat_ns_max 7.06e+08 7.65e+08 47324816.61 8.38%
write_lat_ns_mean 6251503.06 6476912.76 39780.83 3.61%
write_lat_ns_min 3354 4062 616.06 21.11%
And the same, but only showing results where the deviation was outside
of a 95% confidence interval for the mean (default significance
highlighting in fsperf):
qgroup test results
unshare results
metric baseline current stdev diff
========================================================================================
avg_commit_ms 162.13 285.75 3.14 76.24%
elapsed 201.40 270.40 1.34 34.26%
end_state_umount_ns 2.45e+09 2.55e+09 20740154.41 3.93%
max_commit_ms 425.80 594 53.34 39.50%
wait_current_trans_calls 2945.60 3405.20 47.08 15.60%
wait_current_trans_ns_max 1.56e+08 3.43e+08 32659393.25 120.07%
wait_current_trans_ns_mean 1974875.35 28588482.55 1557588.84 1347.61%
wait_current_trans_ns_p95 7711770.20 2.21e+08 17241032.09 2761.19%
wait_current_trans_ns_p99 67744932.29 2.68e+08 41275815.87 295.16%
write_bw_bytes 653008.80 486344.40 4209.91 -25.52%
write_clat_ns_mean 6251404.78 8406837.89 39779.15 34.48%
write_clat_ns_p99 1.90e+08 3.20e+08 2097152 68.62%
write_iops 159.43 118.74 1.03 -25.52%
write_lat_ns_max 7.06e+08 9.80e+08 47324816.61 38.88%
write_lat_ns_mean 6251503.06 8406936.06 39780.83 34.48%
write_lat_ns_min 3354 4648 616.06 38.58%
squota test results
unshare results
metric baseline current stdev diff
========================================================================================
elapsed 201.40 208.20 1.34 3.38%
end_state_umount_ns 2.45e+09 3.01e+09 20740154.41 22.80%
write_bw_bytes 653008.80 631256 4209.91 -3.33%
write_clat_ns_mean 6251404.78 6476816.06 39779.15 3.61%
write_clat_ns_p50 1656422.40 1581056 27415.68 -4.55%
write_clat_ns_p99 1.90e+08 1.94e+08 2097152 2.21%
write_iops 159.43 154.12 1.03 -3.33%
write_lat_ns_mean 6251503.06 6476912.76 39780.83 3.61%
Particularly noteworthy are the massive regressions to
wait_current_trans in qgroup mode as well as the solid regressions to
bandwidth, iops and write latency. The regressions/improvements in
squotas are modest in comparison in line with the expectation. I am
still investigating the squota umount regression, particularly whether
it is in the umount's final commit and represents a real performance
problem with squotas.
Link: https://github.com/boryas/btrfs-progs/tree/squota-progs
Link: https://github.com/boryas/fstests/tree/squota-test
Link: https://github.com/boryas/fsperf/tree/unshare-victim