rocksdb/util/bloom_impl.h
Peter Dillinger f059c7d9b9 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 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
2019-11-13 16:44:01 -08:00

337 lines
14 KiB
C++

// Copyright (c) 2019-present, Facebook, Inc. All rights reserved.
// This source code is licensed under both the GPLv2 (found in the
// COPYING file in the root directory) and Apache 2.0 License
// (found in the LICENSE.Apache file in the root directory).
//
// Implementation details of various Bloom filter implementations used in
// RocksDB. (DynamicBloom is in a separate file for now because it
// supports concurrent write.)
#pragma once
#include <stddef.h>
#include <stdint.h>
#include "rocksdb/slice.h"
#include "util/hash.h"
#ifdef HAVE_AVX2
#include <immintrin.h>
#endif
namespace rocksdb {
// A fast, flexible, and accurate cache-local Bloom implementation with
// SIMD-optimized query performance (currently using AVX2 on Intel). Write
// performance and non-SIMD read are very good, benefiting from fastrange32
// used in place of % and single-cycle multiplication on recent processors.
//
// Most other SIMD Bloom implementations sacrifice flexibility and/or
// accuracy by requiring num_probes to be a power of two and restricting
// where each probe can occur in a cache line. This implementation sacrifices
// SIMD-optimization for add (might still be possible, especially with AVX512)
// in favor of allowing any num_probes, not crossing cache line boundary,
// and accuracy close to theoretical best accuracy for a cache-local Bloom.
// E.g. theoretical best for 10 bits/key, num_probes=6, and 512-bit bucket
// (Intel cache line size) is 0.9535% FP rate. This implementation yields
// about 0.957%. (Compare to LegacyLocalityBloomImpl<false> at 1.138%, or
// about 0.951% for 1024-bit buckets, cache line size for some ARM CPUs.)
//
// This implementation can use a 32-bit hash (let h2 be h1 * 0x9e3779b9) or
// a 64-bit hash (split into two uint32s). With many millions of keys, the
// false positive rate associated with using a 32-bit hash can dominate the
// false positive rate of the underlying filter. At 10 bits/key setting, the
// inflection point is about 40 million keys, so 32-bit hash is a bad idea
// with 10s of millions of keys or more.
//
// Despite accepting a 64-bit hash, this implementation uses 32-bit fastrange
// to pick a cache line, which can be faster than 64-bit in some cases.
// This only hurts accuracy as you get into 10s of GB for a single filter,
// and accuracy abruptly breaks down at 256GB (2^32 cache lines). Switch to
// 64-bit fastrange if you need filters so big. ;)
//
// Using only a 32-bit input hash within each cache line has negligible
// impact for any reasonable cache line / bucket size, for arbitrary filter
// size, and potentially saves intermediate data size in some cases vs.
// tracking full 64 bits. (Even in an implementation using 64-bit arithmetic
// to generate indices, I might do the same, as a single multiplication
// suffices to generate a sufficiently mixed 64 bits from 32 bits.)
//
// This implementation is currently tied to Intel cache line size, 64 bytes ==
// 512 bits. If there's sufficient demand for other cache line sizes, this is
// a pretty good implementation to extend, but slight performance enhancements
// are possible with an alternate implementation (probably not very compatible
// with SIMD):
// (1) Use rotation in addition to multiplication for remixing
// (like murmur hash). (Using multiplication alone *slightly* hurts accuracy
// because lower bits never depend on original upper bits.)
// (2) Extract more than one bit index from each re-mix. (Only if rotation
// or similar is part of remix, because otherwise you're making the
// multiplication-only problem worse.)
// (3) Re-mix full 64 bit hash, to get maximum number of bit indices per
// re-mix.
//
class FastLocalBloomImpl {
public:
static inline void AddHash(uint32_t h1, uint32_t h2, uint32_t len_bytes,
int num_probes, char *data) {
uint32_t bytes_to_cache_line = fastrange32(len_bytes >> 6, h1) << 6;
AddHashPrepared(h2, num_probes, data + bytes_to_cache_line);
}
static inline void AddHashPrepared(uint32_t h2, int num_probes,
char *data_at_cache_line) {
uint32_t h = h2;
for (int i = 0; i < num_probes; ++i, h *= uint32_t{0x9e3779b9}) {
// 9-bit address within 512 bit cache line
int bitpos = h >> (32 - 9);
data_at_cache_line[bitpos >> 3] |= (uint8_t{1} << (bitpos & 7));
}
}
static inline void PrepareHash(uint32_t h1, uint32_t len_bytes,
const char *data,
uint32_t /*out*/ *byte_offset) {
uint32_t bytes_to_cache_line = fastrange32(len_bytes >> 6, h1) << 6;
PREFETCH(data + bytes_to_cache_line, 0 /* rw */, 1 /* locality */);
PREFETCH(data + bytes_to_cache_line + 63, 0 /* rw */, 1 /* locality */);
*byte_offset = bytes_to_cache_line;
}
static inline bool HashMayMatch(uint32_t h1, uint32_t h2, uint32_t len_bytes,
int num_probes, const char *data) {
uint32_t bytes_to_cache_line = fastrange32(len_bytes >> 6, h1) << 6;
return HashMayMatchPrepared(h2, num_probes, data + bytes_to_cache_line);
}
static inline bool HashMayMatchPrepared(uint32_t h2, int num_probes,
const char *data_at_cache_line) {
uint32_t h = h2;
#ifdef HAVE_AVX2
int rem_probes = num_probes;
// NOTE: For better performance for num_probes in {1, 2, 9, 10, 17, 18,
// etc.} one can insert specialized code for rem_probes <= 2, bypassing
// the SIMD code in those cases. There is a detectable but minor overhead
// applied to other values of num_probes (when not statically determined),
// but smoother performance curve vs. num_probes. But for now, when
// in doubt, don't add unnecessary code.
// Powers of 32-bit golden ratio, mod 2**32.
const __m256i multipliers =
_mm256_setr_epi32(0x00000001, 0x9e3779b9, 0xe35e67b1, 0x734297e9,
0x35fbe861, 0xdeb7c719, 0x448b211, 0x3459b749);
for (;;) {
// Eight copies of hash
__m256i hash_vector = _mm256_set1_epi32(h);
// Same effect as repeated multiplication by 0x9e3779b9 thanks to
// associativity of multiplication.
hash_vector = _mm256_mullo_epi32(hash_vector, multipliers);
// Now the top 9 bits of each of the eight 32-bit values in
// hash_vector are bit addresses for probes within the cache line.
// While the platform-independent code uses byte addressing (6 bits
// to pick a byte + 3 bits to pick a bit within a byte), here we work
// with 32-bit words (4 bits to pick a word + 5 bits to pick a bit
// within a word) because that works well with AVX2 and is equivalent
// under little-endian.
// Shift each right by 28 bits to get 4-bit word addresses.
const __m256i word_addresses = _mm256_srli_epi32(hash_vector, 28);
// Gather 32-bit values spread over 512 bits by 4-bit address. In
// essence, we are dereferencing eight pointers within the cache
// line.
//
// Option 1: AVX2 gather (seems to be a little slow - understandable)
// const __m256i value_vector =
// _mm256_i32gather_epi32(static_cast<const int
// *>(data_at_cache_line),
// word_addresses,
// /*bytes / i32*/ 4);
// END Option 1
// Potentially unaligned as we're not *always* cache-aligned -> loadu
const __m256i *mm_data =
reinterpret_cast<const __m256i *>(data_at_cache_line);
__m256i lower = _mm256_loadu_si256(mm_data);
__m256i upper = _mm256_loadu_si256(mm_data + 1);
// Option 2: AVX512VL permute hack
// Only negligibly faster than Option 3, so not yet worth supporting
// const __m256i value_vector =
// _mm256_permutex2var_epi32(lower, word_addresses, upper);
// END Option 2
// Option 3: AVX2 permute+blend hack
// Use lowest three bits to order probing values, as if all from same
// 256 bit piece.
lower = _mm256_permutevar8x32_epi32(lower, word_addresses);
upper = _mm256_permutevar8x32_epi32(upper, word_addresses);
// Just top 1 bit of address, to select between lower and upper.
const __m256i upper_lower_selector = _mm256_srai_epi32(hash_vector, 31);
// Finally: the next 8 probed 32-bit values, in probing sequence order.
const __m256i value_vector =
_mm256_blendv_epi8(lower, upper, upper_lower_selector);
// END Option 3
// We might not need to probe all 8, so build a mask for selecting only
// what we need. (The k_selector(s) could be pre-computed but that
// doesn't seem to make a noticeable performance difference.)
const __m256i zero_to_seven = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
// Subtract rem_probes from each of those constants
__m256i k_selector =
_mm256_sub_epi32(zero_to_seven, _mm256_set1_epi32(rem_probes));
// Negative after subtract -> use/select
// Keep only high bit (logical shift right each by 31).
k_selector = _mm256_srli_epi32(k_selector, 31);
// Strip off the 4 bit word address (shift left)
__m256i bit_addresses = _mm256_slli_epi32(hash_vector, 4);
// And keep only 5-bit (32 - 27) bit-within-32-bit-word addresses.
bit_addresses = _mm256_srli_epi32(bit_addresses, 27);
// Build a bit mask
const __m256i bit_mask = _mm256_sllv_epi32(k_selector, bit_addresses);
// Like ((~value_vector) & bit_mask) == 0)
bool match = _mm256_testc_si256(value_vector, bit_mask) != 0;
// This check first so that it's easy for branch predictor to optimize
// num_probes <= 8 case, making it free of unpredictable branches.
if (rem_probes <= 8) {
return match;
} else if (!match) {
return false;
}
// otherwise
// Need another iteration. 0xab25f4c1 == golden ratio to the 8th power
h *= 0xab25f4c1;
rem_probes -= 8;
}
#else
for (int i = 0; i < num_probes; ++i, h *= uint32_t{0x9e3779b9}) {
// 9-bit address within 512 bit cache line
int bitpos = h >> (32 - 9);
if ((data_at_cache_line[bitpos >> 3] & (char(1) << (bitpos & 7))) == 0) {
return false;
}
}
return true;
#endif
}
};
// A legacy Bloom filter implementation with no locality of probes (slow).
// It uses double hashing to generate a sequence of hash values.
// Asymptotic analysis is in [Kirsch,Mitzenmacher 2006], but known to have
// subtle accuracy flaws for practical sizes [Dillinger,Manolios 2004].
//
// DO NOT REUSE
//
class LegacyNoLocalityBloomImpl {
public:
static inline void AddHash(uint32_t h, uint32_t total_bits, int num_probes,
char *data) {
const uint32_t delta = (h >> 17) | (h << 15); // Rotate right 17 bits
for (int i = 0; i < num_probes; i++) {
const uint32_t bitpos = h % total_bits;
data[bitpos / 8] |= (1 << (bitpos % 8));
h += delta;
}
}
static inline bool HashMayMatch(uint32_t h, uint32_t total_bits,
int num_probes, const char *data) {
const uint32_t delta = (h >> 17) | (h << 15); // Rotate right 17 bits
for (int i = 0; i < num_probes; i++) {
const uint32_t bitpos = h % total_bits;
if ((data[bitpos / 8] & (1 << (bitpos % 8))) == 0) {
return false;
}
h += delta;
}
return true;
}
};
// A legacy Bloom filter implementation with probes local to a single
// cache line (fast). Because SST files might be transported between
// platforms, the cache line size is a parameter rather than hard coded.
// (But if specified as a constant parameter, an optimizing compiler
// should take advantage of that.)
//
// When ExtraRotates is false, this implementation is notably deficient in
// accuracy. Specifically, it uses double hashing with a 1/512 chance of the
// increment being zero (when cache line size is 512 bits). Thus, there's a
// 1/512 chance of probing only one index, which we'd expect to incur about
// a 1/2 * 1/512 or absolute 0.1% FP rate penalty. More detail at
// https://github.com/facebook/rocksdb/issues/4120
//
// DO NOT REUSE
//
template <bool ExtraRotates>
class LegacyLocalityBloomImpl {
private:
static inline uint32_t GetLine(uint32_t h, uint32_t num_lines) {
uint32_t offset_h = ExtraRotates ? (h >> 11) | (h << 21) : h;
return offset_h % num_lines;
}
public:
static inline void AddHash(uint32_t h, uint32_t num_lines, int num_probes,
char *data, int log2_cache_line_bytes) {
const int log2_cache_line_bits = log2_cache_line_bytes + 3;
char *data_at_offset =
data + (GetLine(h, num_lines) << log2_cache_line_bytes);
const uint32_t delta = (h >> 17) | (h << 15);
for (int i = 0; i < num_probes; ++i) {
// Mask to bit-within-cache-line address
const uint32_t bitpos = h & ((1 << log2_cache_line_bits) - 1);
data_at_offset[bitpos / 8] |= (1 << (bitpos % 8));
if (ExtraRotates) {
h = (h >> log2_cache_line_bits) | (h << (32 - log2_cache_line_bits));
}
h += delta;
}
}
static inline void PrepareHashMayMatch(uint32_t h, uint32_t num_lines,
const char *data,
uint32_t /*out*/ *byte_offset,
int log2_cache_line_bytes) {
uint32_t b = GetLine(h, num_lines) << log2_cache_line_bytes;
PREFETCH(data + b, 0 /* rw */, 1 /* locality */);
PREFETCH(data + b + ((1 << log2_cache_line_bytes) - 1), 0 /* rw */,
1 /* locality */);
*byte_offset = b;
}
static inline bool HashMayMatch(uint32_t h, uint32_t num_lines,
int num_probes, const char *data,
int log2_cache_line_bytes) {
uint32_t b = GetLine(h, num_lines) << log2_cache_line_bytes;
return HashMayMatchPrepared(h, num_probes, data + b, log2_cache_line_bytes);
}
static inline bool HashMayMatchPrepared(uint32_t h, int num_probes,
const char *data_at_offset,
int log2_cache_line_bytes) {
const int log2_cache_line_bits = log2_cache_line_bytes + 3;
const uint32_t delta = (h >> 17) | (h << 15);
for (int i = 0; i < num_probes; ++i) {
// Mask to bit-within-cache-line address
const uint32_t bitpos = h & ((1 << log2_cache_line_bits) - 1);
if (((data_at_offset[bitpos / 8]) & (1 << (bitpos % 8))) == 0) {
return false;
}
if (ExtraRotates) {
h = (h >> log2_cache_line_bits) | (h << (32 - log2_cache_line_bits));
}
h += delta;
}
return true;
}
};
} // namespace rocksdb