Summary:
There's no technological impediment to allowing the Bloom
filter bits/key to be non-integer (fractional/decimal) values, and it
provides finer control over the memory vs. accuracy trade-off. This is
especially handy in using the format_version=5 Bloom filter in place
of the old one, because bits_per_key=9.55 provides the same accuracy as
the old bits_per_key=10.
This change not only requires refining the logic for choosing the best
num_probes for a given bits/key setting, it revealed a flaw in that logic.
As bits/key gets higher, the best num_probes for a cache-local Bloom
filter is closer to bpk / 2 than to bpk * 0.69, the best choice for a
standard Bloom filter. For example, at 16 bits per key, the best
num_probes is 9 (FP rate = 0.0843%) not 11 (FP rate = 0.0884%).
This change fixes and refines that logic (for the format_version=5
Bloom filter only, just in case) based on empirical tests to find
accuracy inflection points between each num_probes.
Although bits_per_key is now specified as a double, the new Bloom
filter converts/rounds this to "millibits / key" for predictable/precise
internal computations. Just in case of unforeseen compatibility
issues, we round to the nearest whole number bits / key for the
legacy Bloom filter, so as not to unlock new behaviors for it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6092
Test Plan: unit tests included
Differential Revision: D18711313
Pulled By: pdillinger
fbshipit-source-id: 1aa73295f152a995328cb846ef9157ae8a05522a
Summary:
Adds an improved, replacement Bloom filter implementation (FastLocalBloom) for full and partitioned filters in the block-based table. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single filter.
Speed
The improved speed, at least on recent x86_64, comes from
* Using fastrange instead of modulo (%)
* Using our new hash function (XXH3 preview, added in a previous commit), which is much faster for large keys and only *slightly* slower on keys around 12 bytes if hashing the same size many thousands of times in a row.
* Optimizing the Bloom filter queries with AVX2 SIMD operations. (Added AVX2 to the USE_SSE=1 build.) Careful design was required to support (a) SIMD-optimized queries, (b) compatible non-SIMD code that's simple and efficient, (c) flexible choice of number of probes, and (d) essentially maximized accuracy for a cache-local Bloom filter. Probes are made eight at a time, so any number of probes up to 8 is the same speed, then up to 16, etc.
* Prefetching cache lines when building the filter. Although this optimization could be applied to the old structure as well, it seems to balance out the small added cost of accumulating 64 bit hashes for adding to the filter rather than 32 bit hashes.
Here's nominal speed data from filter_bench (200MB in filters, about 10k keys each, 10 bits filter data / key, 6 probes, avg key size 24 bytes, includes hashing time) on Skylake DE (relatively low clock speed):
$ ./filter_bench -quick -impl=2 -net_includes_hashing # New Bloom filter
Build avg ns/key: 47.7135
Mixed inside/outside queries...
Single filter net ns/op: 26.2825
Random filter net ns/op: 150.459
Average FP rate %: 0.954651
$ ./filter_bench -quick -impl=0 -net_includes_hashing # Old Bloom filter
Build avg ns/key: 47.2245
Mixed inside/outside queries...
Single filter net ns/op: 63.2978
Random filter net ns/op: 188.038
Average FP rate %: 1.13823
Similar build time but dramatically faster query times on hot data (63 ns to 26 ns), and somewhat faster on stale data (188 ns to 150 ns). Performance differences on batched and skewed query loads are between these extremes as expected.
The only other interesting thing about speed is "inside" (query key was added to filter) vs. "outside" (query key was not added to filter) query times. The non-SIMD implementations are substantially slower when most queries are "outside" vs. "inside". This goes against what one might expect or would have observed years ago, as "outside" queries only need about two probes on average, due to short-circuiting, while "inside" always have num_probes (say 6). The problem is probably the nastily unpredictable branch. The SIMD implementation has few branches (very predictable) and has pretty consistent running time regardless of query outcome.
Accuracy
The generally improved accuracy (re: Issue https://github.com/facebook/rocksdb/issues/5857) comes from a better design for probing indices
within a cache line (re: Issue https://github.com/facebook/rocksdb/issues/4120) and improved accuracy for millions of keys in a single filter from using a 64-bit hash function (XXH3p). Design details in code comments.
Accuracy data (generalizes, except old impl gets worse with millions of keys):
Memory bits per key: FP rate percent old impl -> FP rate percent new impl
6: 5.70953 -> 5.69888
8: 2.45766 -> 2.29709
10: 1.13977 -> 0.959254
12: 0.662498 -> 0.411593
16: 0.353023 -> 0.0873754
24: 0.261552 -> 0.0060971
50: 0.225453 -> ~0.00003 (less than 1 in a million queries are FP)
Fixes https://github.com/facebook/rocksdb/issues/5857
Fixes https://github.com/facebook/rocksdb/issues/4120
Unlike the old implementation, this implementation has a fixed cache line size (64 bytes). At 10 bits per key, the accuracy of this new implementation is very close to the old implementation with 128-byte cache line size. If there's sufficient demand, this implementation could be generalized.
Compatibility
Although old releases would see the new structure as corrupt filter data and read the table as if there's no filter, we've decided only to enable the new Bloom filter with new format_version=5. This provides a smooth path for automatic adoption over time, with an option for early opt-in.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6007
Test Plan: filter_bench has been used thoroughly to validate speed, accuracy, and correctness. Unit tests have been carefully updated to exercise new and old implementations, as well as the logic to select an implementation based on context (format_version).
Differential Revision: D18294749
Pulled By: pdillinger
fbshipit-source-id: d44c9db3696e4d0a17caaec47075b7755c262c5f
Summary:
This change sets up for alternate implementations underlying
BloomFilterPolicy:
* Refactor BloomFilterPolicy and expose in internal .h file so that it's easy to iterate over / select implementations for testing, regardless of what the best public interface will look like. Most notably updated db_bloom_filter_test to use this.
* Hide FullFilterBitsBuilder from unit tests (alternate derived classes planned); expose the part important for testing (CalculateSpace), as abstract class BuiltinFilterBitsBuilder. (Also cleaned up internally exposed interface to CalculateSpace.)
* Rename BloomTest -> BlockBasedBloomTest for clarity (despite ongoing confusion between block-based table and block-based filter)
* Assert that block-based filter construction interface is only used on BloomFilterPolicy appropriately constructed. (A couple of tests updated to add ", true".)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5967
Test Plan: make check
Differential Revision: D18138704
Pulled By: pdillinger
fbshipit-source-id: 55ef9273423b0696309e251f50b8c1b5e9ec7597
Summary:
Some filtering tests were unfriendly to new implementations of
FilterBitsBuilder because of dynamic_cast to FullFilterBitsBuilder. Most
of those have now been cleaned up, worked around, or at least changed
from crash on dynamic_cast failure to individual test failure.
Also put some clarifying comments on filter-related APIs.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5960
Test Plan: make check
Differential Revision: D18121223
Pulled By: pdillinger
fbshipit-source-id: e83827d9d5d96315d96f8e25a99cd70f497d802c
Summary:
The parts that are used to implement FilterPolicy /
NewBloomFilterPolicy and not used other than for the block-based table
should be consolidated under table/block_based/filter_policy*. I don't
foresee sharing these APIs with e.g. the Plain Table because they don't
expose hashes for reuse in indexing.
This change is step 1 of 2:
(a) mv table/full_filter_bits_builder.h to
table/block_based/filter_policy_internal.h which I expect to expand
soon to internally reveal more implementation details for testing.
(b) consolidate eventual contents of table/block_based/filter_policy.cc
in util/bloom.cc, which has the most elaborate revision history
(see step 2 ...)
Step 2 soon to follow:
mv util/bloom.cc table/block_based/filter_policy.cc
This gets its own PR so that git has the best chance of following the
rename for blame purposes. Note that low-level shared implementation
details of Bloom filters are in util/bloom_impl.h.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5963
Test Plan: make check
Differential Revision: D18121199
Pulled By: pdillinger
fbshipit-source-id: 8f21732c3d8909777e3240e4ac3123d73140326a
Summary:
FullFilterBitsReader, after creating in BloomFilterPolicy, was
responsible for decoding metadata bits. This meant that
FullFilterBitsReader::MayMatch had some metadata checks in order to
implement "always true" or "always false" functionality in the case
of inconsistent or trivial metadata. This made for ugly
mixing-of-concerns code and probably had some runtime cost. It also
didn't really support plugging in alternative filter implementations
with extensions to the existing metadata schema.
BloomFilterPolicy::GetFilterBitsReader is now (exclusively) responsible
for decoding filter metadata bits and constructing appropriate instances
deriving from FilterBitsReader. "Always false" and "always true" derived
classes allow FullFilterBitsReader not to be concerned with handling of
trivial or inconsistent metadata. This also makes for easy expansion
to alternative filter implementations in new, alternative derived
classes. This change makes calls to FilterBitsReader::MayMatch
*necessarily* virtual because there's now more than one built-in
implementation. Compared with the previous implementation's extra
'if' checks in MayMatch, there's no consistent performance difference,
measured by (an older revision of) filter_bench (differences here seem
to be within noise):
Inside queries...
- Dry run (407) ns/op: 35.9996
+ Dry run (407) ns/op: 35.2034
- Single filter ns/op: 47.5483
+ Single filter ns/op: 47.4034
- Batched, prepared ns/op: 43.1559
+ Batched, prepared ns/op: 42.2923
...
- Random filter ns/op: 150.697
+ Random filter ns/op: 149.403
----------------------------
Outside queries...
- Dry run (980) ns/op: 34.6114
+ Dry run (980) ns/op: 34.0405
- Single filter ns/op: 56.8326
+ Single filter ns/op: 55.8414
- Batched, prepared ns/op: 48.2346
+ Batched, prepared ns/op: 47.5667
- Random filter ns/op: 155.377
+ Random filter ns/op: 153.942
Average FP rate %: 1.1386
Also, the FullFilterBitsReader ctor was responsible for a surprising
amount of CPU in production, due in part to inefficient determination of
the CACHE_LINE_SIZE used to construct the filter being read. The
overwhelming common case (same as my CACHE_LINE_SIZE) is now
substantially optimized, as shown with filter_bench with
-new_reader_every=1 (old option - see below) (repeatable result):
Inside queries...
- Dry run (453) ns/op: 118.799
+ Dry run (453) ns/op: 105.869
- Single filter ns/op: 82.5831
+ Single filter ns/op: 74.2509
...
- Random filter ns/op: 224.936
+ Random filter ns/op: 194.833
----------------------------
Outside queries...
- Dry run (aa1) ns/op: 118.503
+ Dry run (aa1) ns/op: 104.925
- Single filter ns/op: 90.3023
+ Single filter ns/op: 83.425
...
- Random filter ns/op: 220.455
+ Random filter ns/op: 175.7
Average FP rate %: 1.13886
However PR#5936 has/will reclaim most of this cost. After that PR, the optimization of this code path is likely negligible, but nonetheless it's clear we aren't making performance any worse.
Also fixed inadequate check of consistency between filter data size and
num_lines. (Unit test updated.)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5941
Test Plan:
previously added unit tests FullBloomTest.CorruptFilters and
FullBloomTest.RawSchema
Differential Revision: D18018353
Pulled By: pdillinger
fbshipit-source-id: 8e04c2b4a7d93223f49a237fd52ef2483929ed9c
Summary:
Broken type for shift in PR#5834. Fixing code means fixing
expected values in test.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5882
Test Plan: thisisthetest
Differential Revision: D17746136
Pulled By: pdillinger
fbshipit-source-id: d3c456ed30b433d55fcab6fc7d836940fe3b46b8
Summary:
There was significant untested logic in FullFilterBitsReader in
the handling of serialized Bloom filter bits that cannot be generated by
FullFilterBitsBuilder in the current compilation. These now test many of
those corner-case behaviors, including bad metadata or filters created
with different cache line size than the current compiled-in value.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5834
Test Plan: thisisthetest
Differential Revision: D17726372
Pulled By: pdillinger
fbshipit-source-id: fb7b8003b5a8e6fb4666fe95206128f3d5835fc7
Summary:
Further apply formatter to more recent commits.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5830
Test Plan: Run all existing tests.
Differential Revision: D17488031
fbshipit-source-id: 137458fd94d56dd271b8b40c522b03036943a2ab
Summary:
Refactoring to consolidate implementation details of legacy
Bloom filters. This helps to organize and document some related,
obscure code.
Also added make/cpp var TEST_CACHE_LINE_SIZE so that it's easy to
compile and run unit tests for non-native cache line size. (Fixed a
related test failure in db_properties_test.)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5784
Test Plan:
make check, including Recently added Bloom schema unit tests
(in ./plain_table_db_test && ./bloom_test), and including with
TEST_CACHE_LINE_SIZE=128U and TEST_CACHE_LINE_SIZE=256U. Tested the
schema tests with temporary fault injection into new implementations.
Some performance testing with modified unit tests suggest a small to moderate
improvement in speed.
Differential Revision: D17381384
Pulled By: pdillinger
fbshipit-source-id: ee42586da996798910fc45ac0b6289147f16d8df
Summary:
This will allow us to fix history by having the code changes for PR#5784 properly attributed to it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5810
Differential Revision: D17400231
Pulled By: pdillinger
fbshipit-source-id: 2da8b1cdf2533cfedb35b5526eadefb38c291f09
Summary:
Check that we don't accidentally change the on-disk format of
existing Bloom filter implementations, including for various
CACHE_LINE_SIZE (by changing temporarily).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5778
Test Plan: thisisthetest
Differential Revision: D17269630
Pulled By: pdillinger
fbshipit-source-id: c77017662f010a77603b7d475892b1f0d5563d8b
Summary:
FullFilterBitsBuilder::CalculateSpace use CACHE_LINE_SIZE which is 64@X86 but 128@ARM64
when it run bloom_test.FullVaryingLengths it failed on ARM64 server,
the assert can be fixed by change 128->CACHE_LINE_SIZE*2 as merged
ASSERT_LE(FilterSize(), (size_t)((length * 10 / 8) + CACHE_LINE_SIZE * 2 + 5)) << length;
run bloom_test
before fix:
/root/rocksdb-master/util/bloom_test.cc:281: Failure
Expected: (FilterSize()) <= ((size_t)((length * 10 / 8) + 128 + 5)), actual: 389 vs 383
200
[ FAILED ] FullBloomTest.FullVaryingLengths (32 ms)
[----------] 4 tests from FullBloomTest (32 ms total)
[----------] Global test environment tear-down
[==========] 7 tests from 2 test cases ran. (116 ms total)
[ PASSED ] 6 tests.
[ FAILED ] 1 test, listed below:
[ FAILED ] FullBloomTest.FullVaryingLengths
after fix:
Filters: 37 good, 0 mediocre
[ OK ] FullBloomTest.FullVaryingLengths (90 ms)
[----------] 4 tests from FullBloomTest (90 ms total)
[----------] Global test environment tear-down
[==========] 7 tests from 2 test cases ran. (174 ms total)
[ PASSED ] 7 tests.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5745
Differential Revision: D17076047
fbshipit-source-id: e7beb5d55d4855fceb2b84bc8119a6b0759de635
Summary:
Many logging related source files are under util/. It will be more structured if they are together.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5387
Differential Revision: D15579036
Pulled By: siying
fbshipit-source-id: 3850134ed50b8c0bb40a0c8ae1f184fa4081303f
Summary:
There are too many types of files under util/. Some test related files don't belong to there or just are just loosely related. Mo
ve them to a new directory test_util/, so that util/ is cleaner.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5377
Differential Revision: D15551366
Pulled By: siying
fbshipit-source-id: 0f5c8653832354ef8caa31749c0143815d719e2c
Summary:
I started adding gflags support for cmake on linux and got frustrated that I'd need to duplicate the build_detect_platform logic, which determines namespace based on attempting compilation. We can do it differently -- use the GFLAGS_NAMESPACE macro if available, and if not, that indicates it's an old gflags version without configurable namespace so we can simply hardcode "google".
Closes https://github.com/facebook/rocksdb/pull/3212
Differential Revision: D6456973
Pulled By: ajkr
fbshipit-source-id: 3e6d5bde3ca00d4496a120a7caf4687399f5d656
Summary:
Currently metadata_block_size controls only index partition size. With this patch a partition is cut after any of index or filter partitions reaches metadata_block_size.
Closes https://github.com/facebook/rocksdb/pull/2452
Differential Revision: D5275651
Pulled By: maysamyabandeh
fbshipit-source-id: 5057e4424b4c8902043782e6bf8c38f0c4f25160
Summary:
Replacement of #2147
The change was squashed due to a lot of conflicts.
Closes https://github.com/facebook/rocksdb/pull/2194
Differential Revision: D4929799
Pulled By: siying
fbshipit-source-id: 5cd49c254737a1d5ac13f3c035f128e86524c581
Summary: If we skip a test, we shouldn't mark `make check` as failure. This fixes travis CI test.
Test Plan: Travis CI
Reviewers: noetzli, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D47031
Summary:
Our existing test notation is very similar to what is used in gtest. It makes it easy to adopt what is different.
In this diff I modify existing [[ https://code.google.com/p/googletest/wiki/Primer#Test_Fixtures:_Using_the_Same_Data_Configuration_for_Multiple_Te | test fixture ]] classes to inherit from `testing::Test`. Also for unit tests that use fixture class, `TEST` is replaced with `TEST_F` as required in gtest.
There are several custom `main` functions in our existing tests. To make this transition easier, I modify all `main` functions to fallow gtest notation. But eventually we can remove them and use implementation of `main` that gtest provides.
```lang=bash
% cat ~/transform
#!/bin/sh
files=$(git ls-files '*test\.cc')
for file in $files
do
if grep -q "rocksdb::test::RunAllTests()" $file
then
if grep -Eq '^class \w+Test {' $file
then
perl -pi -e 's/^(class \w+Test) {/${1}: public testing::Test {/g' $file
perl -pi -e 's/^(TEST)/${1}_F/g' $file
fi
perl -pi -e 's/(int main.*\{)/${1}::testing::InitGoogleTest(&argc, argv);/g' $file
perl -pi -e 's/rocksdb::test::RunAllTests/RUN_ALL_TESTS/g' $file
fi
done
% sh ~/transform
% make format
```
Second iteration of this diff contains only scripted changes.
Third iteration contains manual changes to fix last errors and make it compilable.
Test Plan:
Build and notice no errors.
```lang=bash
% USE_CLANG=1 make check -j55
```
Tests are still testing.
Reviewers: meyering, sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D35157
Summary:
We need to turn on -Wshorten-64-to-32 for mobile. See D1671432 (internal phabricator) for details.
This diff turns on the warning flag and fixes all the errors. There were also some interesting errors that I might call bugs, especially in plain table. Going forward, I think it makes sense to have this flag turned on and be very very careful when converting 64-bit to 32-bit variables.
Test Plan: compiles
Reviewers: ljin, rven, yhchiang, sdong
Reviewed By: yhchiang
Subscribers: bobbaldwin, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D28689
Summary:
1. Make filter_block.h a base class. Derive block_based_filter_block and full_filter_block. The previous one is the traditional filter block. The full_filter_block is newly added. It would generate a filter block that contain all the keys in SST file.
2. When querying a key, table would first check if full_filter is available. If not, it would go to the exact data block and check using block_based filter.
3. User could choose to use full_filter or tradional(block_based_filter). They would be stored in SST file with different meta index name. "filter.filter_policy" or "full_filter.filter_policy". Then, Table reader is able to know the fllter block type.
4. Some optimizations have been done for full_filter_block, thus it requires a different interface compared to the original one in filter_policy.h.
5. Actual implementation of filter bits coding/decoding is placed in util/bloom_impl.cc
Benchmark: base commit 1d23b5c470
Command:
db_bench --db=/dev/shm/rocksdb --num_levels=6 --key_size=20 --prefix_size=20 --keys_per_prefix=0 --value_size=100 --write_buffer_size=134217728 --max_write_buffer_number=2 --target_file_size_base=33554432 --max_bytes_for_level_base=1073741824 --verify_checksum=false --max_background_compactions=4 --use_plain_table=0 --memtablerep=prefix_hash --open_files=-1 --mmap_read=1 --mmap_write=0 --bloom_bits=10 --bloom_locality=1 --memtable_bloom_bits=500000 --compression_type=lz4 --num=393216000 --use_hash_search=1 --block_size=1024 --block_restart_interval=16 --use_existing_db=1 --threads=1 --benchmarks=readrandom —disable_auto_compactions=1
Read QPS increase for about 30% from 2230002 to 2991411.
Test Plan:
make all check
valgrind db_test
db_stress --use_block_based_filter = 0
./auto_sanity_test.sh
Reviewers: igor, yhchiang, ljin, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D20979
Summary: as title
Test Plan: dynamic_bloom_test
Reviewers: dhruba, sdong, kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14385
Summary:
Change namespace from leveldb to rocksdb. This allows a single
application to link in open-source leveldb code as well as
rocksdb code into the same process.
Test Plan: compile rocksdb
Reviewers: emayanke
Reviewed By: emayanke
CC: leveldb
Differential Revision: https://reviews.facebook.net/D13287
Summary: Replace include/leveldb with include/rocksdb.
Test Plan:
make clean; make check
make clean; make release
Differential Revision: https://reviews.facebook.net/D12489
Summary:
The default compilation process now uses "-Wall" to compile.
Fix all compilation error generated by gcc.
Test Plan: make all check
Reviewers: heyongqiang, emayanke, sheki
Reviewed By: heyongqiang
CC: MarkCallaghan
Differential Revision: https://reviews.facebook.net/D6525
In particular, we add a new FilterPolicy class. An instance
of this class can be supplied in Options when opening a
database. If supplied, the instance is used to generate
summaries of keys (e.g., a bloom filter) which are placed in
sstables. These summaries are consulted by DB::Get() so we
can avoid reading sstable blocks that are guaranteed to not
contain the key we are looking for.
This change provides one implementation of FilterPolicy
based on bloom filters.
Other changes:
- Updated version number to 1.4.
- Some build tweaks.
- C binding for CompactRange.
- A few more benchmarks: deleteseq, deleterandom, readmissing, seekrandom.
- Minor .gitignore update.
