Improve universal compaction sorted-run trigger #12477
Closed
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
cbi42
force-pushed
the
improve-universal-trigger
branch
2 times, most recently
from
2c9f633
to
a7649de
Compare
|
@cbi42 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
ajkr
reviewed
May 16, 2024
There was a problem hiding this comment.
LGTM!
case 2: auto-tune: level0_file_num_compaction_trigger=5, max_read_amp=0
What about even smaller, like level0_file_num_compaction_trigger=1? After all, this is the PR that hopefully saves users from setting level0_file_num_compaction_trigger to magic numbers for reducing write-amp. I guess it will get a little worse, hopefully not too much.
For something structurally different, maybe kCompactionStopStyleSimilarSize and min_merge_width=3. Actually you may also want to adjust the max_read_amp=0 calculation for the kCompactionStopStyleSimilarSize case, or disallow it for now.
| // 0: Let RocksDB auto-tune. Currently, we determine the max number of | ||
| // sorted runs based on the current DB size, size_ratio and | ||
| // write_buffer_size. | ||
| // N (N > 0): limit the number of sorted runs to be at most N. |
There was a problem hiding this comment.
Should we disallow 0 < max_read_amp < level0_file_num_compaction_trigger?
There was a problem hiding this comment.
Yes, added validation in ValidateOptions(). Also disallows max_read_amp < -1.
ajkr
approved these changes
May 16, 2024
cbi42
force-pushed
the
improve-universal-trigger
branch
from
a7649de
to
321262d
Compare
|
@cbi42 has updated the pull request. You must reimport the pull request before landing. |
|
Thanks for the review.
Good point, I added the benchmark result in PR summary. The WA is just slightly worst.
Thanks for catching this. For kCompactionStopStyleSimilarSize, I updated the code to let max_read_amp=0 fallback to old behavior (use level0_file_num_compaction_trigger).
Besides min_merge_width=3, I set max_read_amp=9 and level0_file_num_compaction_trigger=3, the same benchmark gives WA 3.9. |
|
@cbi42 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
I was worried the sizes would be chaotic like that. T189373775 could help as well as a generous value for size ratio. Anyways that's a separate problem. |
cbi42
force-pushed
the
improve-universal-trigger
branch
from
321262d
to
2012f19
Compare
|
@cbi42 has updated the pull request. You must reimport the pull request before landing. |
|
@cbi42 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
konstantinvilin
pushed a commit
to konstantinvilin/rocksdb
that referenced
this pull request
May 28, 2024
Summary: Universal compaction currently uses `level0_file_num_compaction_trigger` for two purposes: 1. the trigger for checking if there is any compaction to do, and 2. the limit on the number of sorted runs. RocksDB will do compaction to keep the number of sorted runs no more than the value of this option. This can make the option inflexible. A value that is too small causes higher write amp: more compactions to reduce the number of sorted runs. A value that is too big delays potential compaction work and causes worse read performance. This PR introduce an option `CompactionOptionsUniversal::max_read_amp` for only the second purpose: to specify the hard limit on the number of sorted runs. For backward compatibility, `max_read_amp = -1` by default, which means to fallback to the current behavior. When `max_read_amp > 0`,`level0_file_num_compaction_trigger` will only serve as a trigger to find potential compaction. When `max_read_amp = 0`, RocksDB will auto-tune the limit on the number of sorted runs. The estimation is based on DB size, write_buffer_size and size_ratio, so it is adaptive to the size change of the DB. See more in `UniversalCompactionBuilder::PickCompaction()`. Alternatively, users now can configure `max_read_amp` to a very big value and keep `level0_file_num_compaction_trigger` small. This will allow `size_ratio` and `max_size_amplification_percent` to control the number of sorted runs. This essentially disables compactions with reason kUniversalSortedRunNum. Pull Request resolved: facebook#12477 Test Plan: * new unit test * existing unit test for default behavior * updated crash test with the new option * benchmark: * Create a DB that is roughly 24GB in the last level. When `max_read_amp = 0`, we estimate that the DB needs 9 levels to avoid excessive compactions to reduce the number of sorted runs. * We then run fillrandom to ingest another 24GB data to compare write amp. * case 1: small level0 trigger: `level0_file_num_compaction_trigger=5, max_read_amp=-1` * write-amp: 4.8 * case 2: auto-tune: `level0_file_num_compaction_trigger=5, max_read_amp=0` * write-amp: 3.6 * case 3: auto-tune with minimal trigger: `level0_file_num_compaction_trigger=1, max_read_amp=0` * write-amp: 3.8 * case 4: hard-code a good value for trigger: `level0_file_num_compaction_trigger=9` * write-amp: 2.8 ``` Case 1: ** Compaction Stats [default] ** Level Files Size Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop Rblob(GB) Wblob(GB) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 0/0 0.00 KB 1.0 0.0 0.0 0.0 22.6 22.6 0.0 1.0 0.0 163.2 141.94 111.10 108 1.314 0 0 0.0 0.0 L45 8/0 1.81 GB 0.0 39.6 11.1 28.5 39.3 10.8 0.0 3.5 209.0 207.3 194.25 191.29 43 4.517 348M 2498K 0.0 0.0 L46 13/0 3.12 GB 0.0 15.3 9.5 5.8 15.0 9.3 0.0 1.6 203.1 199.3 77.13 75.88 16 4.821 134M 2362K 0.0 0.0 L47 19/0 4.68 GB 0.0 15.4 10.5 4.9 14.7 9.8 0.0 1.4 204.0 194.9 77.38 76.15 8 9.673 135M 5920K 0.0 0.0 L48 38/0 9.42 GB 0.0 19.6 11.7 7.9 17.3 9.4 0.0 1.5 206.5 182.3 97.15 95.02 4 24.287 172M 20M 0.0 0.0 L49 91/0 22.70 GB 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.00 0.00 0 0.000 0 0 0.0 0.0 Sum 169/0 41.74 GB 0.0 89.9 42.9 47.0 109.0 61.9 0.0 4.8 156.7 189.8 587.85 549.45 179 3.284 791M 31M 0.0 0.0 Case 2: ** Compaction Stats [default] ** Level Files Size Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop Rblob(GB) Wblob(GB) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 1/0 214.47 MB 1.2 0.0 0.0 0.0 22.6 22.6 0.0 1.0 0.0 164.5 140.81 109.98 108 1.304 0 0 0.0 0.0 L44 0/0 0.00 KB 0.0 1.3 1.3 0.0 1.2 1.2 0.0 1.0 206.1 204.9 6.24 5.98 3 2.081 11M 51K 0.0 0.0 L45 4/0 844.36 MB 0.0 7.1 5.4 1.7 7.0 5.4 0.0 1.3 194.6 192.9 37.41 36.00 13 2.878 62M 489K 0.0 0.0 L46 11/0 2.57 GB 0.0 14.6 9.8 4.8 14.3 9.5 0.0 1.5 193.7 189.8 77.09 73.54 17 4.535 128M 2411K 0.0 0.0 L47 24/0 5.81 GB 0.0 19.8 12.0 7.8 18.8 11.0 0.0 1.6 191.4 181.1 106.19 101.21 9 11.799 174M 9166K 0.0 0.0 L48 38/0 9.42 GB 0.0 19.6 11.8 7.9 17.3 9.4 0.0 1.5 197.3 173.6 101.97 97.23 4 25.491 172M 20M 0.0 0.0 L49 91/0 22.70 GB 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.00 0.00 0 0.000 0 0 0.0 0.0 Sum 169/0 41.54 GB 0.0 62.4 40.3 22.1 81.3 59.2 0.0 3.6 136.1 177.2 469.71 423.94 154 3.050 549M 32M 0.0 0.0 Case 3: ** Compaction Stats [default] ** Level Files Size Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop Rblob(GB) Wblob(GB) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 0/0 0.00 KB 5.0 0.0 0.0 0.0 22.6 22.6 0.0 1.0 0.0 163.8 141.43 111.13 108 1.310 0 0 0.0 0.0 L44 0/0 0.00 KB 0.0 0.8 0.8 0.0 0.8 0.8 0.0 1.0 201.4 200.2 4.26 4.19 2 2.130 7360K 33K 0.0 0.0 L45 4/0 844.38 MB 0.0 6.3 5.0 1.2 6.2 5.0 0.0 1.2 202.0 200.3 31.81 31.50 12 2.651 55M 403K 0.0 0.0 L46 7/0 1.62 GB 0.0 13.3 8.8 4.6 13.1 8.6 0.0 1.5 198.9 195.7 68.72 67.89 17 4.042 117M 1696K 0.0 0.0 L47 24/0 5.81 GB 0.0 21.7 12.9 8.8 20.6 11.8 0.0 1.6 198.5 188.6 112.04 109.97 12 9.336 191M 9352K 0.0 0.0 L48 41/0 10.14 GB 0.0 24.8 13.0 11.8 21.9 10.1 0.0 1.7 198.6 175.6 127.88 125.36 6 21.313 218M 25M 0.0 0.0 L49 91/0 22.70 GB 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.00 0.00 0 0.000 0 0 0.0 0.0 Sum 167/0 41.10 GB 0.0 67.0 40.5 26.4 85.4 58.9 0.0 3.8 141.1 179.8 486.13 450.04 157 3.096 589M 36M 0.0 0.0 Case 4: ** Compaction Stats [default] ** Level Files Size Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop Rblob(GB) Wblob(GB) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 0/0 0.00 KB 0.7 0.0 0.0 0.0 22.6 22.6 0.0 1.0 0.0 158.6 146.02 114.68 108 1.352 0 0 0.0 0.0 L42 0/0 0.00 KB 0.0 1.7 1.7 0.0 1.7 1.7 0.0 1.0 185.4 184.3 9.25 8.96 4 2.314 14M 67K 0.0 0.0 L43 0/0 0.00 KB 0.0 2.5 2.5 0.0 2.5 2.5 0.0 1.0 197.8 195.6 13.01 12.65 4 3.253 22M 202K 0.0 0.0 L44 4/0 844.40 MB 0.0 4.2 4.2 0.0 4.1 4.1 0.0 1.0 188.1 185.1 22.81 21.89 5 4.562 36M 503K 0.0 0.0 L45 13/0 3.12 GB 0.0 7.5 6.5 1.0 7.2 6.2 0.0 1.1 188.7 181.8 40.69 39.32 5 8.138 65M 2282K 0.0 0.0 L46 17/0 4.18 GB 0.0 8.3 7.1 1.2 7.9 6.6 0.0 1.1 192.2 181.8 44.23 43.06 4 11.058 73M 3846K 0.0 0.0 L47 22/0 5.34 GB 0.0 8.9 7.5 1.4 8.2 6.8 0.0 1.1 189.1 174.1 48.12 45.37 3 16.041 78M 6098K 0.0 0.0 L48 27/0 6.58 GB 0.0 9.2 7.6 1.6 8.2 6.6 0.0 1.1 195.2 172.9 48.52 47.11 2 24.262 81M 9217K 0.0 0.0 L49 91/0 22.70 GB 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.00 0.00 0 0.000 0 0 0.0 0.0 Sum 174/0 42.74 GB 0.0 42.3 37.0 5.3 62.4 57.1 0.0 2.8 116.3 171.3 372.66 333.04 135 2.760 372M 22M 0.0 0.0 setup: ./db_bench --benchmarks=fillseq,compactall,waitforcompaction --num=200000000 --compression_type=none --disable_wal=1 --compaction_style=1 --num_levels=50 --target_file_size_base=268435456 --max_compaction_bytes=6710886400 --level0_file_num_compaction_trigger=10 --write_buffer_size=268435456 --seed 1708494134896523 benchmark: ./db_bench --benchmarks=overwrite,waitforcompaction,stats --num=200000000 --compression_type=none --disable_wal=1 --compaction_style=1 --write_buffer_size=268435456 --level0_file_num_compaction_trigger=5 --target_file_size_base=268435456 --use_existing_db=1 --num_levels=50 --writes=200000000 --universal_max_read_amp=-1 --seed=1716488324800233 ``` Reviewed By: ajkr Differential Revision: D55370922 Pulled By: cbi42 fbshipit-source-id: 9be69979126b840d08e93e7059260e76a878bb2a
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary: Universal compaction currently uses
level0_file_num_compaction_triggerfor two purposes:This can make the option inflexible. A value that is too small causes higher write amp: more compactions to reduce the number of sorted runs. A value that is too big delays potential compaction work and causes worse read performance. This PR introduce an option
CompactionOptionsUniversal::max_read_ampfor only the second purpose: to specifythe hard limit on the number of sorted runs.
For backward compatibility,
max_read_amp = -1by default, which means to fallback to the current behavior.When
max_read_amp > 0,level0_file_num_compaction_triggerwill only serve as a trigger to find potential compaction.When
max_read_amp = 0, RocksDB will auto-tune the limit on the number of sorted runs. The estimation is based on DB size, write_buffer_size and size_ratio, so it is adaptive to the size change of the DB. See more inUniversalCompactionBuilder::PickCompaction().Alternatively, users now can configure
max_read_ampto a very big value and keeplevel0_file_num_compaction_triggersmall. This will allowsize_ratioandmax_size_amplification_percentto control the number of sorted runs. This essentially disables compactions with reason kUniversalSortedRunNum.Test plan:
max_read_amp = 0, we estimate that the DB needs 9 levels to avoid excessive compactions to reduce the number of sorted runs.level0_file_num_compaction_trigger=5, max_read_amp=-1level0_file_num_compaction_trigger=5, max_read_amp=0level0_file_num_compaction_trigger=1, max_read_amp=0level0_file_num_compaction_trigger=9