Summary:
Help users that would benefit most from new Bloom filter
implementation by logging a warning that recommends the using
format_version >= 5.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6312
Test Plan:
$ (for BPK in 10 13 14 19 20 50; do ./filter_bench -quick -impl=0 -bits_per_key=$BPK -m_queries=1 2>&1; done) | grep 'its/key'
Bits/key actual: 10.0647
Bits/key actual: 13.0593
[WARN] [/block_based/filter_policy.cc:546] Using legacy Bloom filter with high (14) bits/key. Significant filter space and/or accuracy improvement is available with format_verion>=5.
Bits/key actual: 14.0581
[WARN] [/block_based/filter_policy.cc:546] Using legacy Bloom filter with high (19) bits/key. Significant filter space and/or accuracy improvement is available with format_verion>=5.
Bits/key actual: 19.0542
[WARN] [/block_based/filter_policy.cc:546] Using legacy Bloom filter with high (20) bits/key. Dramatic filter space and/or accuracy improvement is available with format_verion>=5.
Bits/key actual: 20.0584
[WARN] [/block_based/filter_policy.cc:546] Using legacy Bloom filter with high (50) bits/key. Dramatic filter space and/or accuracy improvement is available with format_verion>=5.
Bits/key actual: 50.0577
Differential Revision: D19457191
Pulled By: pdillinger
fbshipit-source-id: 073d94cde5c70e03a160f953e1100c15ea83eda4
Summary:
The filter bits builder collects all the hashes to add in memory before adding them (because the number of keys is not known until we've walked over all the keys). Existing code uses a std::vector for this, which can mean up to 2x than necessary space allocated (and not freed) and up to ~2x write amplification in memory. Using std::deque uses close to minimal space (for large filters, the only time it matters), no write amplification, frees memory while building, and no need for large contiguous memory area. The only cost is more calls to allocator, which does not appear to matter, at least in benchmark test.
For now, this change only applies to the new (format_version=5) Bloom filter implementation, to ease before-and-after comparison downstream.
Temporary memory use during build is about the only way the new Bloom filter could regress vs. the old (because of upgrade to 64-bit hash) and that should only matter for full filters. This change should largely mitigate that potential regression.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6175
Test Plan:
Using filter_bench with -new_builder option and 6M keys per filter is like large full filter (improvement). 10k keys and no -new_builder is like partitioned filters (about the same). (Corresponding configurations run simultaneously on devserver.)
std::vector impl (before)
$ /usr/bin/time -v ./filter_bench -impl=2 -quick -new_builder -working_mem_size_mb=1000 -
average_keys_per_filter=6000000
Build avg ns/key: 52.2027
Maximum resident set size (kbytes): 1105016
$ /usr/bin/time -v ./filter_bench -impl=2 -quick -working_mem_size_mb=1000 -
average_keys_per_filter=10000
Build avg ns/key: 30.5694
Maximum resident set size (kbytes): 1208152
std::deque impl (after)
$ /usr/bin/time -v ./filter_bench -impl=2 -quick -new_builder -working_mem_size_mb=1000 -
average_keys_per_filter=6000000
Build avg ns/key: 39.0697
Maximum resident set size (kbytes): 1087196
$ /usr/bin/time -v ./filter_bench -impl=2 -quick -working_mem_size_mb=1000 -
average_keys_per_filter=10000
Build avg ns/key: 30.9348
Maximum resident set size (kbytes): 1207980
Differential Revision: D19053431
Pulled By: pdillinger
fbshipit-source-id: 2888e748723a19d9ea40403934f13cbb8483430c
Summary:
This change enables custom implementations of FilterPolicy to
wrap a variety of NewBloomFilterPolicy and select among them based on
contextual information such as table level and compaction style.
* Moves FilterBuildingContext to public API and elaborates it with more
useful data. (It would be nice to put more general options-like data,
but at the time this object is constructed, we are using internal APIs
ImmutableCFOptions and MutableCFOptions and don't have easy access to
ColumnFamilyOptions that I can tell.)
* Renames BloomFilterPolicy::GetFilterBitsBuilderInternal to
GetBuilderWithContext, because it's now public.
* Plumbs through the table's "level_at_creation" for filter building
context.
* Simplified some tests by adding GetBuilder() to
MockBlockBasedTableTester.
* Adds test as DBBloomFilterTest.ContextCustomFilterPolicy, including
sample wrapper class LevelAndStyleCustomFilterPolicy.
* Fixes a cross-test bug in DBBloomFilterTest.OptimizeFiltersForHits
where it does not reset perf context.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6088
Test Plan: make check, valgrind on db_bloom_filter_test
Differential Revision: D18697817
Pulled By: pdillinger
fbshipit-source-id: 5f987a2d7b07cc7a33670bc08ca6b4ca698c1cf4
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:
For upcoming new SST filter implementations, we will use a new
64-bit hash function (XXH3 preview, slightly modified). This change
updates hash.{h,cc} for that change, adds unit tests, and out-of-lines
the implementations to keep hash.h as clean/small as possible.
In developing the unit tests, I discovered that the XXH3 preview always
returns zero for the empty string. Zero is problematic for some
algorithms (including an upcoming SST filter implementation) if it
occurs more often than at the "natural" rate, so it should not be
returned from trivial values using trivial seeds. I modified our fork
of XXH3 to return a modest hash of the seed for the empty string.
With hash function details out-of-lines in hash.h, it makes sense to
enable XXH_INLINE_ALL, so that direct calls to XXH64/XXH32/XXH3p
are inlined. To fix array-bounds warnings on some inline calls, I
injected some casts to uintptr_t in xxhash.cc. (Issue reported to Yann.)
Revised: Reverted using XXH_INLINE_ALL for now. Some Facebook
checks are unhappy about #include on xxhash.cc file. I would
fix that by rename to xxhash_cc.h, but to best preserve history I want
to do that in a separate commit (PR) from the uintptr casts.
Also updated filter_bench for this change, improving the performance
predictability of dry run hashing and adding support for 64-bit hash
(for upcoming new SST filter implementations, minor dead code in the
tool for now).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5984
Differential Revision: D18246567
Pulled By: pdillinger
fbshipit-source-id: 6162fbf6381d63c8cc611dd7ec70e1ddc883fbb8
Summary:
filter_bench is a specialized micro-benchmarking tool that
should not be needed with ROCKSDB_LITE. This should fix the LITE build.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5978
Test Plan: make LITE=1 check
Differential Revision: D18177941
Pulled By: pdillinger
fbshipit-source-id: b73a171404661e09e018bc99afcf8d4bf1e2949c
Summary:
* Adds support for plain table filter. This is not critical right now, but does add a -impl flag that will be useful for new filter implementations initially targeted at block-based table (and maybe later ported to plain table)
* Better mixing of inside vs. outside queries, for more realism
* A -best_case option handy for implementation tuning inner loop
* Option for whether to include hashing time in dry run / net timings
No modifications to production code, just filter_bench.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5968
Differential Revision: D18139872
Pulled By: pdillinger
fbshipit-source-id: 5b09eba963111b48f9e0525a706e9921070990e8
Summary:
The first version of filter_bench has selectable key size
but that size does not vary throughout a test run. This artificially
favors "branchy" hash functions like the existing BloomHash,
MurmurHash1, probably because of optimal return for branch prediction.
This change primarily varies those key sizes from -2 to +2 bytes vs.
the average selected size. We also set the default key size at 24 to
better reflect our best guess of typical key size.
But steadily random key sizes may not be realistic either. So this
change introduces a new filter_bench option:
-vary_key_size_log2_interval=n where the same key size is used 2^n
times and then changes to another size. I've set the default at 5
(32 times same size) as a compromise between deployments with
rather consistent vs. rather variable key sizes. On my Skylake
system, the performance boost to MurmurHash1 largely lies between
n=10 and n=15.
Also added -vary_key_alignment (bool, now default=true), though this
doesn't currently seem to matter in hash functions under
consideration.
This change also does a "dry run" for each testing scenario, to improve
the accuracy of those numbers, as there was more difference between
scenarios than expected. Subtracting gross test run times from dry run
times is now also embedded in the output, because these "net" times are
generally the most useful.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5933
Differential Revision: D18121683
Pulled By: pdillinger
fbshipit-source-id: 3c7efee1c5661a5fe43de555e786754ddf80dc1e
Summary:
Amongst other things, PR https://github.com/facebook/rocksdb/issues/5504 refactored the filter block readers so that
only the filter block contents are stored in the block cache (as opposed to the
earlier design where the cache stored the filter block reader itself, leading to
potentially dangling pointers and concurrency bugs). However, this change
introduced a performance hit since with the new code, the metadata fields are
re-parsed upon every access. This patch reunites the block contents with the
filter bits reader to eliminate this overhead; since this is still a self-contained
pure data object, it is safe to store it in the cache. (Note: this is similar to how
the zstd digest is handled.)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5936
Test Plan:
make asan_check
filter_bench results for the old code:
```
$ ./filter_bench -quick
WARNING: Assertions are enabled; benchmarks unnecessarily slow
Building...
Build avg ns/key: 26.7153
Number of filters: 16669
Total memory (MB): 200.009
Bits/key actual: 10.0647
----------------------------
Inside queries...
Dry run (46b) ns/op: 33.4258
Single filter ns/op: 42.5974
Random filter ns/op: 217.861
----------------------------
Outside queries...
Dry run (25d) ns/op: 32.4217
Single filter ns/op: 50.9855
Random filter ns/op: 219.167
Average FP rate %: 1.13993
----------------------------
Done. (For more info, run with -legend or -help.)
$ ./filter_bench -quick -use_full_block_reader
WARNING: Assertions are enabled; benchmarks unnecessarily slow
Building...
Build avg ns/key: 26.5172
Number of filters: 16669
Total memory (MB): 200.009
Bits/key actual: 10.0647
----------------------------
Inside queries...
Dry run (46b) ns/op: 32.3556
Single filter ns/op: 83.2239
Random filter ns/op: 370.676
----------------------------
Outside queries...
Dry run (25d) ns/op: 32.2265
Single filter ns/op: 93.5651
Random filter ns/op: 408.393
Average FP rate %: 1.13993
----------------------------
Done. (For more info, run with -legend or -help.)
```
With the new code:
```
$ ./filter_bench -quick
WARNING: Assertions are enabled; benchmarks unnecessarily slow
Building...
Build avg ns/key: 25.4285
Number of filters: 16669
Total memory (MB): 200.009
Bits/key actual: 10.0647
----------------------------
Inside queries...
Dry run (46b) ns/op: 31.0594
Single filter ns/op: 43.8974
Random filter ns/op: 226.075
----------------------------
Outside queries...
Dry run (25d) ns/op: 31.0295
Single filter ns/op: 50.3824
Random filter ns/op: 226.805
Average FP rate %: 1.13993
----------------------------
Done. (For more info, run with -legend or -help.)
$ ./filter_bench -quick -use_full_block_reader
WARNING: Assertions are enabled; benchmarks unnecessarily slow
Building...
Build avg ns/key: 26.5308
Number of filters: 16669
Total memory (MB): 200.009
Bits/key actual: 10.0647
----------------------------
Inside queries...
Dry run (46b) ns/op: 33.2968
Single filter ns/op: 58.6163
Random filter ns/op: 291.434
----------------------------
Outside queries...
Dry run (25d) ns/op: 32.1839
Single filter ns/op: 66.9039
Random filter ns/op: 292.828
Average FP rate %: 1.13993
----------------------------
Done. (For more info, run with -legend or -help.)
```
Differential Revision: D17991712
Pulled By: ltamasi
fbshipit-source-id: 7ea205550217bfaaa1d5158ebd658e5832e60f29
Summary:
Fixed some spots where converting size_t or uint_fast32_t to
uint32_t. Wrapped mt19937 in a new Random32 class to avoid future
such traps.
NB: I tried using Random32::Uniform (std::uniform_int_distribution) in
filter_bench instead of fastrange, but that more than doubled the dry
run time! So I added fastrange as Random32::Uniformish. ;)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5894
Test Plan: USE_CLANG=1 build, and manual re-run filter_bench
Differential Revision: D17825131
Pulled By: pdillinger
fbshipit-source-id: 68feee333b5f8193c084ded760e3d6679b405ecd
Summary:
Example: using the tool before and after PR https://github.com/facebook/rocksdb/issues/5784 shows that
the refactoring, presumed performance-neutral, actually sped up SST
filters by about 3% to 8% (repeatable result):
Before:
- Dry run ns/op: 22.4725
- Single filter ns/op: 51.1078
- Random filter ns/op: 120.133
After:
+ Dry run ns/op: 22.2301
+ Single filter run ns/op: 47.4313
+ Random filter ns/op: 115.9
Only tests filters for the block-based table (full filters and
partitioned filters - same implementation; not block-based filters),
which seems to be the recommended format/implementation.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5825
Differential Revision: D17804987
Pulled By: pdillinger
fbshipit-source-id: 0f18a9c254c57f7866030d03e7fa4ba503bac3c5