New Bloom filter implementation for full and partitioned filters (#6007)

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 #5857) comes from a better design for probing indices
within a cache line (re: Issue #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 #5857
Fixes #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: #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

CMakeLists.txt

@@ -470,7 +470,7 @@ endif()
 include_directories(${PROJECT_SOURCE_DIR})
 include_directories(${PROJECT_SOURCE_DIR}/include)
 include_directories(SYSTEM ${PROJECT_SOURCE_DIR}/third-party/gtest-1.8.1/fused-src)
-if(WITH_FOLLY_DISTRIBUTED_MUTEX)
+if(NOT ROCKSDB_LITE)
   include_directories(${PROJECT_SOURCE_DIR}/third-party/folly)
 endif()
 find_package(Threads REQUIRED)

HISTORY.md

@@ -5,11 +5,14 @@
 * Changed the default value of periodic_compaction_seconds to `UINT64_MAX` which allows RocksDB to auto-tune periodic compaction scheduling. When using the default value, periodic compactions are now auto-enabled if a compaction filter is used. A value of `0` will turn off the feature completely.
 * With FIFO compaction style, options.periodic_compaction_seconds will have the same meaning as options.ttl. Whichever stricter will be used. With the default options.periodic_compaction_seconds value with options.ttl's default of 0, RocksDB will give a default of 30 days.
 * Added an API GetCreationTimeOfOldestFile(uint64_t* creation_time) to get the file_creation_time of the oldest SST file in the DB.
+* An unlikely usage of FilterPolicy is no longer supported. Calling GetFilterBitsBuilder() on the FilterPolicy returned by NewBloomFilterPolicy will now cause an assertion violation in debug builds, because RocksDB has internally migrated to a more elaborate interface that is expected to evolve further. Custom implementations of FilterPolicy should work as before, except those wrapping the return of NewBloomFilterPolicy, which will require a new override of a protected function in FilterPolicy.
 
 ### New Features
 * Universal compaction to support options.periodic_compaction_seconds. A full compaction will be triggered if any file is over the threshold.
 * `GetLiveFilesMetaData` and `GetColumnFamilyMetaData` now expose the file number of SST files as well as the oldest blob file referenced by each SST.
 * A batched MultiGet API (DB::MultiGet()) that supports retrieving keys from multiple column families.
+* Full and partitioned filters in the block-based table use an improved Bloom filter implementation, enabled with format_version 5 (or above) because previous releases cannot read this filter. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single (full) filter. For example, the new Bloom filter has a lower false positive rate at 16 bits per key than the old one at 100 bits per key.
+* Added AVX2 instructions to USE_SSE builds to accelerate the new Bloom filter and XXH3-based hash function on compatible x86_64 platforms (Haswell and later, ~2014).
 
 ### Performance Improvements
 * For 64-bit hashing, RocksDB is standardizing on a slightly modified preview version of XXH3. This function is now used for many non-persisted hashes, along with fastrange64() in place of the modulus operator, and some benchmarks show a slight improvement.

Makefile

@@ -320,7 +320,7 @@ else
 	PLATFORM_CXXFLAGS += -isystem $(GTEST_DIR)
 endif
 
-ifeq ($(USE_FOLLY_DISTRIBUTED_MUTEX),1)
+ifeq ($(filter -DROCKSDB_LITE,$(OPT)),)
 	FOLLY_DIR = ./third-party/folly
 	# AIX: pre-defined system headers are surrounded by an extern "C" block
 	ifeq ($(PLATFORM), OS_AIX)

TARGETS

@@ -77,6 +77,7 @@ ROCKSDB_PREPROCESSOR_FLAGS = [
     # Directories with files for #include
     "-I" + REPO_PATH + "include/",
     "-I" + REPO_PATH,
+    "-I" + REPO_PATH + "third-party/folly/",
 ]
 
 ROCKSDB_ARCH_PREPROCESSOR_FLAGS = {

buckifier/targets_cfg.py

@@ -83,6 +83,7 @@
     # Directories with files for #include
     "-I" + REPO_PATH + "include/",
     "-I" + REPO_PATH,
+    "-I" + REPO_PATH + "third-party/folly/",
 ]
 
 ROCKSDB_ARCH_PREPROCESSOR_FLAGS = {

build_tools/build_detect_platform

@@ -572,6 +572,8 @@ if test "$TRY_SSE_ETC"; then
   TRY_SSE42="-msse4.2"
   # PCLMUL available since westmere, ca. 2010-2011
   TRY_PCLMUL="-mpclmul"
+  # AVX2 available since haswell, ca. 2013-2015
+  TRY_AVX2="-mavx2"
 fi
 
 $CXX $PLATFORM_CXXFLAGS $COMMON_FLAGS $TRY_SSE42 -x c++ - -o /dev/null 2>/dev/null <<EOF
@@ -605,6 +607,21 @@ elif test "$USE_SSE"; then
   echo "warning: USE_SSE specified but compiler could not use PCLMUL intrinsics, disabling" >&2
 fi
 
+$CXX $PLATFORM_CXXFLAGS $COMMON_FLAGS $TRY_AVX2 -x c++ - -o /dev/null 2>/dev/null <<EOF
+  #include <cstdint>
+  #include <immintrin.h>
+  int main() {
+    const auto a = _mm256_setr_epi32(0, 1, 2, 3, 4, 7, 6, 5);
+    const auto b = _mm256_permutevar8x32_epi32(a, a);
+    (void)b;
+  }
+EOF
+if [ "$?" = 0 ]; then
+  COMMON_FLAGS="$COMMON_FLAGS $TRY_AVX2 -DHAVE_AVX2"
+elif test "$USE_SSE"; then
+  echo "warning: USE_SSE specified but compiler could not use AVX2 intrinsics, disabling" >&2
+fi
+
 $CXX $PLATFORM_CXXFLAGS $COMMON_FLAGS -x c++ - -o /dev/null 2>/dev/null <<EOF
   #include <cstdint>
   int main() {

db/db_bloom_filter_test.cc

@@ -7,6 +7,9 @@
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file. See the AUTHORS file for names of contributors.
 
+#ifndef ROCKSDB_LITE
+#include <folly/Optional.h>
+#endif  // ROCKSDB_LITE
 #include "db/db_test_util.h"
 #include "port/stack_trace.h"
 #include "rocksdb/perf_context.h"
@@ -27,10 +30,10 @@ class DBBloomFilterTest : public DBTestBase {
 
 class DBBloomFilterTestWithParam : public DBTestBase,
                                    public testing::WithParamInterface<
-                                       std::tuple<BFP::Impl, bool, uint32_t>> {
+                                       std::tuple<BFP::Mode, bool, uint32_t>> {
   //                             public testing::WithParamInterface<bool> {
  protected:
-  BFP::Impl bfp_impl_;
+  BFP::Mode bfp_impl_;
   bool partition_filters_;
   uint32_t format_version_;
 
@@ -506,23 +509,26 @@ TEST_P(DBBloomFilterTestWithParam, BloomFilter) {
 INSTANTIATE_TEST_CASE_P(
     FormatDef, DBBloomFilterTestDefFormatVersion,
     ::testing::Values(
-        std::make_tuple(BFP::kBlock, false, test::kDefaultFormatVersion),
-        std::make_tuple(BFP::kFull, true, test::kDefaultFormatVersion),
-        std::make_tuple(BFP::kFull, false, test::kDefaultFormatVersion)));
+        std::make_tuple(BFP::kDeprecatedBlock, false,
+                        test::kDefaultFormatVersion),
+        std::make_tuple(BFP::kAuto, true, test::kDefaultFormatVersion),
+        std::make_tuple(BFP::kAuto, false, test::kDefaultFormatVersion)));
 
 INSTANTIATE_TEST_CASE_P(
     FormatDef, DBBloomFilterTestWithParam,
     ::testing::Values(
-        std::make_tuple(BFP::kBlock, false, test::kDefaultFormatVersion),
-        std::make_tuple(BFP::kFull, true, test::kDefaultFormatVersion),
-        std::make_tuple(BFP::kFull, false, test::kDefaultFormatVersion)));
+        std::make_tuple(BFP::kDeprecatedBlock, false,
+                        test::kDefaultFormatVersion),
+        std::make_tuple(BFP::kAuto, true, test::kDefaultFormatVersion),
+        std::make_tuple(BFP::kAuto, false, test::kDefaultFormatVersion)));
 
 INSTANTIATE_TEST_CASE_P(
     FormatLatest, DBBloomFilterTestWithParam,
     ::testing::Values(
-        std::make_tuple(BFP::kBlock, false, test::kLatestFormatVersion),
-        std::make_tuple(BFP::kFull, true, test::kLatestFormatVersion),
-        std::make_tuple(BFP::kFull, false, test::kLatestFormatVersion)));
+        std::make_tuple(BFP::kDeprecatedBlock, false,
+                        test::kLatestFormatVersion),
+        std::make_tuple(BFP::kAuto, true, test::kLatestFormatVersion),
+        std::make_tuple(BFP::kAuto, false, test::kLatestFormatVersion)));
 #endif  // ROCKSDB_VALGRIND_RUN
 
 TEST_F(DBBloomFilterTest, BloomFilterRate) {
@@ -861,27 +867,27 @@ TEST_F(DBBloomFilterTest, MemtablePrefixBloomOutOfDomain) {
 #ifndef ROCKSDB_LITE
 class BloomStatsTestWithParam
     : public DBBloomFilterTest,
-      public testing::WithParamInterface<std::tuple<bool, BFP::Impl, bool>> {
+      public testing::WithParamInterface<
+          std::tuple<folly::Optional<BFP::Mode>, bool>> {
  public:
   BloomStatsTestWithParam() {
-    use_block_table_ = std::get<0>(GetParam());
-    bfp_impl_ = std::get<1>(GetParam());
-    partition_filters_ = std::get<2>(GetParam());
+    bfp_impl_ = std::get<0>(GetParam());
+    partition_filters_ = std::get<1>(GetParam());
 
     options_.create_if_missing = true;
     options_.prefix_extractor.reset(rocksdb::NewFixedPrefixTransform(4));
     options_.memtable_prefix_bloom_size_ratio =
         8.0 * 1024.0 / static_cast<double>(options_.write_buffer_size);
-    if (use_block_table_) {
+    if (bfp_impl_) {
       BlockBasedTableOptions table_options;
       table_options.hash_index_allow_collision = false;
       if (partition_filters_) {
-        assert(bfp_impl_ != BFP::kBlock);
+        assert(*bfp_impl_ != BFP::kDeprecatedBlock);
         table_options.partition_filters = partition_filters_;
         table_options.index_type =
             BlockBasedTableOptions::IndexType::kTwoLevelIndexSearch;
       }
-      table_options.filter_policy.reset(new BFP(10, bfp_impl_));
+      table_options.filter_policy.reset(new BFP(10, *bfp_impl_));
       options_.table_factory.reset(NewBlockBasedTableFactory(table_options));
     } else {
       assert(!partition_filters_);  // not supported in plain table
@@ -903,8 +909,7 @@ class BloomStatsTestWithParam
   static void SetUpTestCase() {}
   static void TearDownTestCase() {}
 
-  bool use_block_table_;
-  BFP::Impl bfp_impl_;
+  folly::Optional<BFP::Mode> bfp_impl_;
   bool partition_filters_;
   Options options_;
 };
@@ -1008,7 +1013,7 @@ TEST_P(BloomStatsTestWithParam, BloomStatsTestWithIter) {
   ASSERT_EQ(value3, iter->value().ToString());
   // The seek doesn't check block-based bloom filter because last index key
   // starts with the same prefix we're seeking to.
-  uint64_t expected_hits = bfp_impl_ == BFP::kBlock ? 1 : 2;
+  uint64_t expected_hits = bfp_impl_ == BFP::kDeprecatedBlock ? 1 : 2;
   ASSERT_EQ(expected_hits, get_perf_context()->bloom_sst_hit_count);
 
   iter->Seek(key2);
@@ -1020,10 +1025,12 @@ TEST_P(BloomStatsTestWithParam, BloomStatsTestWithIter) {
 
 INSTANTIATE_TEST_CASE_P(
     BloomStatsTestWithParam, BloomStatsTestWithParam,
-    ::testing::Values(std::make_tuple(true, BFP::kBlock, false),
-                      std::make_tuple(true, BFP::kFull, false),
-                      std::make_tuple(true, BFP::kFull, true),
-                      std::make_tuple(false, BFP::kFull, false)));
+    ::testing::Values(std::make_tuple(BFP::kDeprecatedBlock, false),
+                      std::make_tuple(BFP::kLegacyBloom, false),
+                      std::make_tuple(BFP::kLegacyBloom, true),
+                      std::make_tuple(BFP::kFastLocalBloom, false),
+                      std::make_tuple(BFP::kFastLocalBloom, true),
+                      std::make_tuple(folly::Optional<BFP::Mode>(), false)));
 
 namespace {
 void PrefixScanInit(DBBloomFilterTest* dbtest) {
@@ -1330,8 +1337,8 @@ int CountIter(std::unique_ptr<Iterator>& iter, const Slice& key) {
 // into the same string, or 2) the transformed seek key is of the same length
 // as the upper bound and two keys are adjacent according to the comparator.
 TEST_F(DBBloomFilterTest, DynamicBloomFilterUpperBound) {
-  for (auto bfp_impl : BFP::kAllImpls) {
-    int using_full_builder = bfp_impl != BFP::kBlock;
+  for (auto bfp_impl : BFP::kAllFixedImpls) {
+    int using_full_builder = bfp_impl != BFP::kDeprecatedBlock;
     Options options;
     options.create_if_missing = true;
     options.prefix_extractor.reset(NewCappedPrefixTransform(4));
@@ -1461,8 +1468,8 @@ TEST_F(DBBloomFilterTest, DynamicBloomFilterUpperBound) {
 // Create multiple SST files each with a different prefix_extractor config,
 // verify iterators can read all SST files using the latest config.
 TEST_F(DBBloomFilterTest, DynamicBloomFilterMultipleSST) {
-  for (auto bfp_impl : BFP::kAllImpls) {
-    int using_full_builder = bfp_impl != BFP::kBlock;
+  for (auto bfp_impl : BFP::kAllFixedImpls) {
+    int using_full_builder = bfp_impl != BFP::kDeprecatedBlock;
     Options options;
     options.create_if_missing = true;
     options.prefix_extractor.reset(NewFixedPrefixTransform(1));
@@ -1596,7 +1603,7 @@ TEST_F(DBBloomFilterTest, DynamicBloomFilterMultipleSST) {
 // as expected
 TEST_F(DBBloomFilterTest, DynamicBloomFilterNewColumnFamily) {
   int iteration = 0;
-  for (auto bfp_impl : BFP::kAllImpls) {
+  for (auto bfp_impl : BFP::kAllFixedImpls) {
     Options options = CurrentOptions();
     options.create_if_missing = true;
     options.prefix_extractor.reset(NewFixedPrefixTransform(1));
@@ -1654,7 +1661,7 @@ TEST_F(DBBloomFilterTest, DynamicBloomFilterNewColumnFamily) {
 // Verify it's possible to change prefix_extractor at runtime and iterators
 // behaves as expected
 TEST_F(DBBloomFilterTest, DynamicBloomFilterOptions) {
-  for (auto bfp_impl : BFP::kAllImpls) {
+  for (auto bfp_impl : BFP::kAllFixedImpls) {
     Options options;
     options.create_if_missing = true;
     options.prefix_extractor.reset(NewFixedPrefixTransform(1));

include/rocksdb/filter_policy.h

@@ -80,6 +80,9 @@ class FilterBitsReader {
   }
 };
 
+// Internal type required for FilterPolicy
+struct FilterBuildingContext;
+
 // We add a new format of filter block called full filter block
 // This new interface gives you more space of customization
 //
@@ -131,6 +134,18 @@ class FilterPolicy {
       const Slice& /*contents*/) const {
     return nullptr;
   }
+
+ protected:
+  // An internal-use-only variant of GetFilterBitsBuilder that allows
+  // a built-in FilterPolicy to customize the builder for contextual
+  // constraints and hints. (Name changed to avoid triggering
+  // -Werror=overloaded-virtual.)
+  virtual FilterBitsBuilder* GetFilterBitsBuilderInternal(
+      const FilterBuildingContext&) const {
+    return GetFilterBitsBuilder();
+  }
+
+  friend FilterBuildingContext;
 };
 
 // Return a new filter policy that uses a bloom filter with approximately

include/rocksdb/table.h

@@ -269,6 +269,9 @@ struct BlockBasedTableOptions {
   // probably use this as it would reduce the index size.
   // This option only affects newly written tables. When reading existing
   // tables, the information about version is read from the footer.
+  // 5 -- Can be read by RocksDB's versions since X.X.X (something after 6.4.6)
+  // Full and partitioned filters use a generally faster and more accurate
+  // Bloom filter implementation, with a different schema.
   uint32_t format_version = 2;
 
   // Store index blocks on disk in compressed format. Changing this option to

table/block_based/block_based_table_builder.cc

@@ -25,7 +25,6 @@
 #include "rocksdb/cache.h"
 #include "rocksdb/comparator.h"
 #include "rocksdb/env.h"
-#include "rocksdb/filter_policy.h"
 #include "rocksdb/flush_block_policy.h"
 #include "rocksdb/merge_operator.h"
 #include "rocksdb/table.h"
@@ -36,6 +35,7 @@
 #include "table/block_based/block_based_table_reader.h"
 #include "table/block_based/block_builder.h"
 #include "table/block_based/filter_block.h"
+#include "table/block_based/filter_policy_internal.h"
 #include "table/block_based/full_filter_block.h"
 #include "table/block_based/partitioned_filter_block.h"
 #include "table/format.h"
@@ -68,7 +68,7 @@ FilterBlockBuilder* CreateFilterBlockBuilder(
   if (table_opt.filter_policy == nullptr) return nullptr;
 
   FilterBitsBuilder* filter_bits_builder =
-      table_opt.filter_policy->GetFilterBitsBuilder();
+      FilterBuildingContext(table_opt).GetBuilder();
   if (filter_bits_builder == nullptr) {
     return new BlockBasedFilterBlockBuilder(mopt.prefix_extractor.get(),
                                             table_opt);