You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
rocksdb/util/filter_bench.cc

808 lines
28 KiB

// Copyright (c) 2011-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).
#if !defined(GFLAGS) || defined(ROCKSDB_LITE)
#include <cstdio>
int main() {
fprintf(stderr, "filter_bench requires gflags and !ROCKSDB_LITE\n");
return 1;
}
#else
#include <cinttypes>
#include <iostream>
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
#include <sstream>
#include <vector>
#include "memory/arena.h"
#include "port/port.h"
#include "port/stack_trace.h"
Account Bloom/Ribbon filter construction memory in global memory limit (#9073) Summary: Note: This PR is the 4th part of a bigger PR stack (https://github.com/facebook/rocksdb/pull/9073) and will rebase/merge only after the first three PRs (https://github.com/facebook/rocksdb/pull/9070, https://github.com/facebook/rocksdb/pull/9071, https://github.com/facebook/rocksdb/pull/9130) merge. **Context:** Similar to https://github.com/facebook/rocksdb/pull/8428, this PR is to track memory usage during (new) Bloom Filter (i.e,FastLocalBloom) and Ribbon Filter (i.e, Ribbon128) construction, moving toward the goal of [single global memory limit using block cache capacity](https://github.com/facebook/rocksdb/wiki/Projects-Being-Developed#improving-memory-efficiency). It also constrains the size of the banding portion of Ribbon Filter during construction by falling back to Bloom Filter if that banding is, at some point, larger than the available space in the cache under `LRUCacheOptions::strict_capacity_limit=true`. The option to turn on this feature is `BlockBasedTableOptions::reserve_table_builder_memory = true` which by default is set to `false`. We [decided](https://github.com/facebook/rocksdb/pull/9073#discussion_r741548409) not to have separate option for separate memory user in table building therefore their memory accounting are all bundled under one general option. **Summary:** - Reserved/released cache for creation/destruction of three main memory users with the passed-in `FilterBuildingContext::cache_res_mgr` during filter construction: - hash entries (i.e`hash_entries`.size(), we bucket-charge hash entries during insertion for performance), - banding (Ribbon Filter only, `bytes_coeff_rows` +`bytes_result_rows` + `bytes_backtrack`), - final filter (i.e, `mutable_buf`'s size). - Implementation details: in order to use `CacheReservationManager::CacheReservationHandle` to account final filter's memory, we have to store the `CacheReservationManager` object and `CacheReservationHandle` for final filter in `XXPH3BitsFilterBuilder` as well as explicitly delete the filter bits builder when done with the final filter in block based table. - Added option fo run `filter_bench` with this memory reservation feature Pull Request resolved: https://github.com/facebook/rocksdb/pull/9073 Test Plan: - Added new tests in `db_bloom_filter_test` to verify filter construction peak cache reservation under combination of `BlockBasedTable::Rep::FilterType` (e.g, `kFullFilter`, `kPartitionedFilter`), `BloomFilterPolicy::Mode`(e.g, `kFastLocalBloom`, `kStandard128Ribbon`, `kDeprecatedBlock`) and `BlockBasedTableOptions::reserve_table_builder_memory` - To address the concern for slow test: tests with memory reservation under `kFullFilter` + `kStandard128Ribbon` and `kPartitionedFilter` take around **3000 - 6000 ms** and others take around **1500 - 2000 ms**, in total adding **20000 - 25000 ms** to the test suit running locally - Added new test in `bloom_test` to verify Ribbon Filter fallback on large banding in FullFilter - Added test in `filter_bench` to verify that this feature does not significantly slow down Bloom/Ribbon Filter construction speed. Local result averaged over **20** run as below: - FastLocalBloom - baseline `./filter_bench -impl=2 -quick -runs 20 | grep 'Build avg'`: - **Build avg ns/key: 29.56295** (DEBUG_LEVEL=1), **29.98153** (DEBUG_LEVEL=0) - new feature (expected to be similar as above)`./filter_bench -impl=2 -quick -runs 20 -reserve_table_builder_memory=true | grep 'Build avg'`: - **Build avg ns/key: 30.99046** (DEBUG_LEVEL=1), **30.48867** (DEBUG_LEVEL=0) - new feature of RibbonFilter with fallback (expected to be similar as above) `./filter_bench -impl=2 -quick -runs 20 -reserve_table_builder_memory=true -strict_capacity_limit=true | grep 'Build avg'` : - **Build avg ns/key: 31.146975** (DEBUG_LEVEL=1), **30.08165** (DEBUG_LEVEL=0) - Ribbon128 - baseline `./filter_bench -impl=3 -quick -runs 20 | grep 'Build avg'`: - **Build avg ns/key: 129.17585** (DEBUG_LEVEL=1), **130.5225** (DEBUG_LEVEL=0) - new feature (expected to be similar as above) `./filter_bench -impl=3 -quick -runs 20 -reserve_table_builder_memory=true | grep 'Build avg' `: - **Build avg ns/key: 131.61645** (DEBUG_LEVEL=1), **132.98075** (DEBUG_LEVEL=0) - new feature of RibbonFilter with fallback (expected to be a lot faster than above due to fallback) `./filter_bench -impl=3 -quick -runs 20 -reserve_table_builder_memory=true -strict_capacity_limit=true | grep 'Build avg'` : - **Build avg ns/key: 52.032965** (DEBUG_LEVEL=1), **52.597825** (DEBUG_LEVEL=0) - And the warning message of `"Cache reservation for Ribbon filter banding failed due to cache full"` is indeed logged to console. Reviewed By: pdillinger Differential Revision: D31991348 Pulled By: hx235 fbshipit-source-id: 9336b2c60f44d530063da518ceaf56dac5f9df8e
3 years ago
#include "rocksdb/cache.h"
#include "rocksdb/system_clock.h"
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
5 years ago
#include "table/block_based/filter_policy_internal.h"
#include "table/block_based/full_filter_block.h"
#include "table/block_based/mock_block_based_table.h"
#include "table/plain/plain_table_bloom.h"
#include "util/cast_util.h"
#include "util/gflags_compat.h"
#include "util/hash.h"
#include "util/random.h"
#include "util/stderr_logger.h"
#include "util/stop_watch.h"
using GFLAGS_NAMESPACE::ParseCommandLineFlags;
using GFLAGS_NAMESPACE::RegisterFlagValidator;
using GFLAGS_NAMESPACE::SetUsageMessage;
DEFINE_uint32(seed, 0, "Seed for random number generators");
DEFINE_double(working_mem_size_mb, 200,
"MB of memory to get up to among all filters, unless "
"m_keys_total_max is specified.");
DEFINE_uint32(average_keys_per_filter, 10000,
"Average number of keys per filter");
DEFINE_double(vary_key_count_ratio, 0.4,
"Vary number of keys by up to +/- vary_key_count_ratio * "
"average_keys_per_filter.");
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
DEFINE_uint32(key_size, 24, "Average number of bytes for each key");
DEFINE_bool(vary_key_alignment, true,
"Whether to vary key alignment (default: at least 32-bit "
"alignment)");
DEFINE_uint32(vary_key_size_log2_interval, 5,
"Use same key size 2^n times, then change. Key size varies from "
"-2 to +2 bytes vs. average, unless n>=30 to fix key size.");
DEFINE_uint32(batch_size, 8, "Number of keys to group in each batch");
Allow fractional bits/key in BloomFilterPolicy (#6092) 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
5 years ago
DEFINE_double(bits_per_key, 10.0, "Bits per key setting for filters");
DEFINE_double(m_queries, 200, "Millions of queries for each test mode");
DEFINE_double(m_keys_total_max, 0,
"Maximum total keys added to filters, in millions. "
"0 (default) disables. Non-zero overrides working_mem_size_mb "
"option.");
DEFINE_bool(use_full_block_reader, false,
"Use FullFilterBlockReader interface rather than FilterBitsReader");
DEFINE_bool(use_plain_table_bloom, false,
"Use PlainTableBloom structure and interface rather than "
"FilterBitsReader/FullFilterBlockReader");
Optimize memory and CPU for building new Bloom filter (#6175) 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
5 years ago
DEFINE_bool(new_builder, false,
"Whether to create a new builder for each new filter");
DEFINE_uint32(impl, 0,
"Select filter implementation. Without -use_plain_table_bloom:"
Experimental (production candidate) SST schema for Ribbon filter (#7658) Summary: Added experimental public API for Ribbon filter: NewExperimentalRibbonFilterPolicy(). This experimental API will take a "Bloom equivalent" bits per key, and configure the Ribbon filter for the same FP rate as Bloom would have but ~30% space savings. (Note: optimize_filters_for_memory is not yet implemented for Ribbon filter. That can be added with no effect on schema.) Internally, the Ribbon filter is configured using a "one_in_fp_rate" value, which is 1 over desired FP rate. For example, use 100 for 1% FP rate. I'm expecting this will be used in the future for configuring Bloom-like filters, as I expect people to more commonly hold constant the filter accuracy and change the space vs. time trade-off, rather than hold constant the space (per key) and change the accuracy vs. time trade-off, though we might make that available. ### Benchmarking ``` $ ./filter_bench -impl=2 -quick -m_keys_total_max=200 -average_keys_per_filter=100000 -net_includes_hashing Building... Build avg ns/key: 34.1341 Number of filters: 1993 Total size (MB): 238.488 Reported total allocated memory (MB): 262.875 Reported internal fragmentation: 10.2255% Bits/key stored: 10.0029 ---------------------------- Mixed inside/outside queries... Single filter net ns/op: 18.7508 Random filter net ns/op: 258.246 Average FP rate %: 0.968672 ---------------------------- Done. (For more info, run with -legend or -help.) $ ./filter_bench -impl=3 -quick -m_keys_total_max=200 -average_keys_per_filter=100000 -net_includes_hashing Building... Build avg ns/key: 130.851 Number of filters: 1993 Total size (MB): 168.166 Reported total allocated memory (MB): 183.211 Reported internal fragmentation: 8.94626% Bits/key stored: 7.05341 ---------------------------- Mixed inside/outside queries... Single filter net ns/op: 58.4523 Random filter net ns/op: 363.717 Average FP rate %: 0.952978 ---------------------------- Done. (For more info, run with -legend or -help.) ``` 168.166 / 238.488 = 0.705 -> 29.5% space reduction 130.851 / 34.1341 = 3.83x construction time for this Ribbon filter vs. lastest Bloom filter (could make that as little as about 2.5x for less space reduction) ### Working around a hashing "flaw" bloom_test discovered a flaw in the simple hashing applied in StandardHasher when num_starts == 1 (num_slots == 128), showing an excessively high FP rate. The problem is that when many entries, on the order of number of hash bits or kCoeffBits, are associated with the same start location, the correlation between the CoeffRow and ResultRow (for efficiency) can lead to a solution that is "universal," or nearly so, for entries mapping to that start location. (Normally, variance in start location breaks the effective association between CoeffRow and ResultRow; the same value for CoeffRow is effectively different if start locations are different.) Without kUseSmash and with num_starts > 1 (thus num_starts ~= num_slots), this flaw should be completely irrelevant. Even with 10M slots, the chances of a single slot having just 16 (or more) entries map to it--not enough to cause an FP problem, which would be local to that slot if it happened--is 1 in millions. This spreadsheet formula shows that: =1/(10000000*(1 - POISSON(15, 1, TRUE))) As kUseSmash==false (the setting for Standard128RibbonBitsBuilder) is intended for CPU efficiency of filters with many more entries/slots than kCoeffBits, a very reasonable work-around is to disallow num_starts==1 when !kUseSmash, by making the minimum non-zero number of slots 2*kCoeffBits. This is the work-around I've applied. This also means that the new Ribbon filter schema (Standard128RibbonBitsBuilder) is not space-efficient for less than a few hundred entries. Because of this, I have made it fall back on constructing a Bloom filter, under existing schema, when that is more space efficient for small filters. (We can change this in the future if we want.) TODO: better unit tests for this case in ribbon_test, and probably update StandardHasher for kUseSmash case so that it can scale nicely to small filters. ### Other related changes * Add Ribbon filter to stress/crash test * Add Ribbon filter to filter_bench as -impl=3 * Add option string support, as in "filter_policy=experimental_ribbon:5.678;" where 5.678 is the Bloom equivalent bits per key. * Rename internal mode BloomFilterPolicy::kAuto to kAutoBloom * Add a general BuiltinFilterBitsBuilder::CalculateNumEntry based on binary searching CalculateSpace (inefficient), so that subclasses (especially experimental ones) don't have to provide an efficient implementation inverting CalculateSpace. * Minor refactor FastLocalBloomBitsBuilder for new base class XXH3pFilterBitsBuilder shared with new Standard128RibbonBitsBuilder, which allows the latter to fall back on Bloom construction in some extreme cases. * Mostly updated bloom_test for Ribbon filter, though a test like FullBloomTest::Schema is a next TODO to ensure schema stability (in case this becomes production-ready schema as it is). * Add some APIs to ribbon_impl.h for configuring Ribbon filters. Although these are reasonably covered by bloom_test, TODO more unit tests in ribbon_test * Added a "tool" FindOccupancyForSuccessRate to ribbon_test to get data for constructing the linear approximations in GetNumSlotsFor95PctSuccess. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7658 Test Plan: Some unit tests updated but other testing is left TODO. This is considered experimental but laying down schema compatibility as early as possible in case it proves production-quality. Also tested in stress/crash test. Reviewed By: jay-zhuang Differential Revision: D24899349 Pulled By: pdillinger fbshipit-source-id: 9715f3e6371c959d923aea8077c9423c7a9f82b8
4 years ago
"0 = legacy full Bloom filter, 1 = block-based Bloom filter, "
"2 = format_version 5 Bloom filter, 3 = Ribbon128 filter. With "
"-use_plain_table_bloom: 0 = no locality, 1 = locality.");
DEFINE_bool(net_includes_hashing, false,
"Whether query net ns/op times should include hashing. "
"(if not, dry run will include hashing) "
"(build times always include hashing)");
Minimize memory internal fragmentation for Bloom filters (#6427) Summary: New experimental option BBTO::optimize_filters_for_memory builds filters that maximize their use of "usable size" from malloc_usable_size, which is also used to compute block cache charges. Rather than always "rounding up," we track state in the BloomFilterPolicy object to mix essentially "rounding down" and "rounding up" so that the average FP rate of all generated filters is the same as without the option. (YMMV as heavily accessed filters might be unluckily lower accuracy.) Thus, the option near-minimizes what the block cache considers as "memory used" for a given target Bloom filter false positive rate and Bloom filter implementation. There are no forward or backward compatibility issues with this change, though it only works on the format_version=5 Bloom filter. With Jemalloc, we see about 10% reduction in memory footprint (and block cache charge) for Bloom filters, but 1-2% increase in storage footprint, due to encoding efficiency losses (FP rate is non-linear with bits/key). Why not weighted random round up/down rather than state tracking? By only requiring malloc_usable_size, we don't actually know what the next larger and next smaller usable sizes for the allocator are. We pick a requested size, accept and use whatever usable size it has, and use the difference to inform our next choice. This allows us to narrow in on the right balance without tracking/predicting usable sizes. Why not weight history of generated filter false positive rates by number of keys? This could lead to excess skew in small filters after generating a large filter. Results from filter_bench with jemalloc (irrelevant details omitted): (normal keys/filter, but high variance) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=30000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.6278 Number of filters: 5516 Total size (MB): 200.046 Reported total allocated memory (MB): 220.597 Reported internal fragmentation: 10.2732% Bits/key stored: 10.0097 Average FP rate %: 0.965228 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=30000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 30.5104 Number of filters: 5464 Total size (MB): 200.015 Reported total allocated memory (MB): 200.322 Reported internal fragmentation: 0.153709% Bits/key stored: 10.1011 Average FP rate %: 0.966313 (very few keys / filter, optimization not as effective due to ~59 byte internal fragmentation in blocked Bloom filter representation) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.5649 Number of filters: 162950 Total size (MB): 200.001 Reported total allocated memory (MB): 224.624 Reported internal fragmentation: 12.3117% Bits/key stored: 10.2951 Average FP rate %: 0.821534 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 31.8057 Number of filters: 159849 Total size (MB): 200 Reported total allocated memory (MB): 208.846 Reported internal fragmentation: 4.42297% Bits/key stored: 10.4948 Average FP rate %: 0.811006 (high keys/filter) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.7017 Number of filters: 164 Total size (MB): 200.352 Reported total allocated memory (MB): 221.5 Reported internal fragmentation: 10.5552% Bits/key stored: 10.0003 Average FP rate %: 0.969358 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 30.7131 Number of filters: 160 Total size (MB): 200.928 Reported total allocated memory (MB): 200.938 Reported internal fragmentation: 0.00448054% Bits/key stored: 10.1852 Average FP rate %: 0.963387 And from db_bench (block cache) with jemalloc: $ ./db_bench -db=/dev/shm/dbbench.no_optimize -benchmarks=fillrandom -format_version=5 -value_size=90 -bloom_bits=10 -num=2000000 -threads=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false $ ./db_bench -db=/dev/shm/dbbench -benchmarks=fillrandom -format_version=5 -value_size=90 -bloom_bits=10 -num=2000000 -threads=8 -optimize_filters_for_memory -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false $ (for FILE in /dev/shm/dbbench.no_optimize/*.sst; do ./sst_dump --file=$FILE --show_properties | grep 'filter block' ; done) | awk '{ t += $4; } END { print t; }' 17063835 $ (for FILE in /dev/shm/dbbench/*.sst; do ./sst_dump --file=$FILE --show_properties | grep 'filter block' ; done) | awk '{ t += $4; } END { print t; }' 17430747 $ #^ 2.1% additional filter storage $ ./db_bench -db=/dev/shm/dbbench.no_optimize -use_existing_db -benchmarks=readrandom,stats -statistics -bloom_bits=10 -num=2000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false -duration=10 -cache_index_and_filter_blocks -cache_size=1000000000 rocksdb.block.cache.index.add COUNT : 33 rocksdb.block.cache.index.bytes.insert COUNT : 8440400 rocksdb.block.cache.filter.add COUNT : 33 rocksdb.block.cache.filter.bytes.insert COUNT : 21087528 rocksdb.bloom.filter.useful COUNT : 4963889 rocksdb.bloom.filter.full.positive COUNT : 1214081 rocksdb.bloom.filter.full.true.positive COUNT : 1161999 $ #^ 1.04 % observed FP rate $ ./db_bench -db=/dev/shm/dbbench -use_existing_db -benchmarks=readrandom,stats -statistics -bloom_bits=10 -num=2000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false -optimize_filters_for_memory -duration=10 -cache_index_and_filter_blocks -cache_size=1000000000 rocksdb.block.cache.index.add COUNT : 33 rocksdb.block.cache.index.bytes.insert COUNT : 8448592 rocksdb.block.cache.filter.add COUNT : 33 rocksdb.block.cache.filter.bytes.insert COUNT : 18220328 rocksdb.bloom.filter.useful COUNT : 5360933 rocksdb.bloom.filter.full.positive COUNT : 1321315 rocksdb.bloom.filter.full.true.positive COUNT : 1262999 $ #^ 1.08 % observed FP rate, 13.6% less memory usage for filters (Due to specific key density, this example tends to generate filters that are "worse than average" for internal fragmentation. "Better than average" cases can show little or no improvement.) Pull Request resolved: https://github.com/facebook/rocksdb/pull/6427 Test Plan: unit test added, 'make check' with gcc, clang and valgrind Reviewed By: siying Differential Revision: D22124374 Pulled By: pdillinger fbshipit-source-id: f3e3aa152f9043ddf4fae25799e76341d0d8714e
5 years ago
DEFINE_bool(optimize_filters_for_memory, false,
"Setting for BlockBasedTableOptions::optimize_filters_for_memory");
Account Bloom/Ribbon filter construction memory in global memory limit (#9073) Summary: Note: This PR is the 4th part of a bigger PR stack (https://github.com/facebook/rocksdb/pull/9073) and will rebase/merge only after the first three PRs (https://github.com/facebook/rocksdb/pull/9070, https://github.com/facebook/rocksdb/pull/9071, https://github.com/facebook/rocksdb/pull/9130) merge. **Context:** Similar to https://github.com/facebook/rocksdb/pull/8428, this PR is to track memory usage during (new) Bloom Filter (i.e,FastLocalBloom) and Ribbon Filter (i.e, Ribbon128) construction, moving toward the goal of [single global memory limit using block cache capacity](https://github.com/facebook/rocksdb/wiki/Projects-Being-Developed#improving-memory-efficiency). It also constrains the size of the banding portion of Ribbon Filter during construction by falling back to Bloom Filter if that banding is, at some point, larger than the available space in the cache under `LRUCacheOptions::strict_capacity_limit=true`. The option to turn on this feature is `BlockBasedTableOptions::reserve_table_builder_memory = true` which by default is set to `false`. We [decided](https://github.com/facebook/rocksdb/pull/9073#discussion_r741548409) not to have separate option for separate memory user in table building therefore their memory accounting are all bundled under one general option. **Summary:** - Reserved/released cache for creation/destruction of three main memory users with the passed-in `FilterBuildingContext::cache_res_mgr` during filter construction: - hash entries (i.e`hash_entries`.size(), we bucket-charge hash entries during insertion for performance), - banding (Ribbon Filter only, `bytes_coeff_rows` +`bytes_result_rows` + `bytes_backtrack`), - final filter (i.e, `mutable_buf`'s size). - Implementation details: in order to use `CacheReservationManager::CacheReservationHandle` to account final filter's memory, we have to store the `CacheReservationManager` object and `CacheReservationHandle` for final filter in `XXPH3BitsFilterBuilder` as well as explicitly delete the filter bits builder when done with the final filter in block based table. - Added option fo run `filter_bench` with this memory reservation feature Pull Request resolved: https://github.com/facebook/rocksdb/pull/9073 Test Plan: - Added new tests in `db_bloom_filter_test` to verify filter construction peak cache reservation under combination of `BlockBasedTable::Rep::FilterType` (e.g, `kFullFilter`, `kPartitionedFilter`), `BloomFilterPolicy::Mode`(e.g, `kFastLocalBloom`, `kStandard128Ribbon`, `kDeprecatedBlock`) and `BlockBasedTableOptions::reserve_table_builder_memory` - To address the concern for slow test: tests with memory reservation under `kFullFilter` + `kStandard128Ribbon` and `kPartitionedFilter` take around **3000 - 6000 ms** and others take around **1500 - 2000 ms**, in total adding **20000 - 25000 ms** to the test suit running locally - Added new test in `bloom_test` to verify Ribbon Filter fallback on large banding in FullFilter - Added test in `filter_bench` to verify that this feature does not significantly slow down Bloom/Ribbon Filter construction speed. Local result averaged over **20** run as below: - FastLocalBloom - baseline `./filter_bench -impl=2 -quick -runs 20 | grep 'Build avg'`: - **Build avg ns/key: 29.56295** (DEBUG_LEVEL=1), **29.98153** (DEBUG_LEVEL=0) - new feature (expected to be similar as above)`./filter_bench -impl=2 -quick -runs 20 -reserve_table_builder_memory=true | grep 'Build avg'`: - **Build avg ns/key: 30.99046** (DEBUG_LEVEL=1), **30.48867** (DEBUG_LEVEL=0) - new feature of RibbonFilter with fallback (expected to be similar as above) `./filter_bench -impl=2 -quick -runs 20 -reserve_table_builder_memory=true -strict_capacity_limit=true | grep 'Build avg'` : - **Build avg ns/key: 31.146975** (DEBUG_LEVEL=1), **30.08165** (DEBUG_LEVEL=0) - Ribbon128 - baseline `./filter_bench -impl=3 -quick -runs 20 | grep 'Build avg'`: - **Build avg ns/key: 129.17585** (DEBUG_LEVEL=1), **130.5225** (DEBUG_LEVEL=0) - new feature (expected to be similar as above) `./filter_bench -impl=3 -quick -runs 20 -reserve_table_builder_memory=true | grep 'Build avg' `: - **Build avg ns/key: 131.61645** (DEBUG_LEVEL=1), **132.98075** (DEBUG_LEVEL=0) - new feature of RibbonFilter with fallback (expected to be a lot faster than above due to fallback) `./filter_bench -impl=3 -quick -runs 20 -reserve_table_builder_memory=true -strict_capacity_limit=true | grep 'Build avg'` : - **Build avg ns/key: 52.032965** (DEBUG_LEVEL=1), **52.597825** (DEBUG_LEVEL=0) - And the warning message of `"Cache reservation for Ribbon filter banding failed due to cache full"` is indeed logged to console. Reviewed By: pdillinger Differential Revision: D31991348 Pulled By: hx235 fbshipit-source-id: 9336b2c60f44d530063da518ceaf56dac5f9df8e
3 years ago
DEFINE_uint32(block_cache_capacity_MB, 8,
"Setting for "
"LRUCacheOptions::capacity");
DEFINE_bool(reserve_table_builder_memory, false,
"Setting for "
"BlockBasedTableOptions::reserve_table_builder_memory");
DEFINE_bool(strict_capacity_limit, false,
"Setting for "
"LRUCacheOptions::strict_capacity_limit");
DEFINE_bool(quick, false, "Run more limited set of tests, fewer queries");
DEFINE_bool(best_case, false, "Run limited tests only for best-case");
DEFINE_bool(allow_bad_fp_rate, false, "Continue even if FP rate is bad");
DEFINE_bool(legend, false,
"Print more information about interpreting results instead of "
"running tests");
DEFINE_uint32(runs, 1, "Number of times to rebuild and run benchmark tests");
void _always_assert_fail(int line, const char *file, const char *expr) {
fprintf(stderr, "%s: %d: Assertion %s failed\n", file, line, expr);
abort();
}
#define ALWAYS_ASSERT(cond) \
((cond) ? (void)0 : ::_always_assert_fail(__LINE__, __FILE__, #cond))
Warn on excessive keys for legacy Bloom filter with 32-bit hash (#6317) Summary: With many millions of keys, the old Bloom filter implementation for the block-based table (format_version <= 4) would have excessive FP rate due to the limitations of feeding the Bloom filter with a 32-bit hash. This change computes an estimated inflated FP rate due to this effect and warns in the log whenever an SST filter is constructed (almost certainly a "full" not "partitioned" filter) that exceeds 1.5x FP rate due to this effect. The detailed condition is only checked if 3 million keys or more have been added to a filter, as this should be a lower bound for common bits/key settings (< 20). Recommended remedies include smaller SST file size, using format_version >= 5 (for new Bloom filter), or using partitioned filters. This does not change behavior other than generating warnings for some constructed filters using the old implementation. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6317 Test Plan: Example with warning, 15M keys @ 15 bits / key: (working_mem_size_mb is just to stop after building one filter if it's large) $ ./filter_bench -quick -impl=0 -working_mem_size_mb=1 -bits_per_key=15 -average_keys_per_filter=15000000 2>&1 | grep 'FP rate' [WARN] [/block_based/filter_policy.cc:292] Using legacy SST/BBT Bloom filter with excessive key count (15.0M @ 15bpk), causing estimated 1.8x higher filter FP rate. Consider using new Bloom with format_version>=5, smaller SST file size, or partitioned filters. Predicted FP rate %: 0.766702 Average FP rate %: 0.66846 Example without warning (150K keys): $ ./filter_bench -quick -impl=0 -working_mem_size_mb=1 -bits_per_key=15 -average_keys_per_filter=150000 2>&1 | grep 'FP rate' Predicted FP rate %: 0.422857 Average FP rate %: 0.379301 $ With more samples at 15 bits/key: 150K keys -> no warning; actual: 0.379% FP rate (baseline) 1M keys -> no warning; actual: 0.396% FP rate, 1.045x 9M keys -> no warning; actual: 0.563% FP rate, 1.485x 10M keys -> warning (1.5x); actual: 0.564% FP rate, 1.488x 15M keys -> warning (1.8x); actual: 0.668% FP rate, 1.76x 25M keys -> warning (2.4x); actual: 0.880% FP rate, 2.32x At 10 bits/key: 150K keys -> no warning; actual: 1.17% FP rate (baseline) 1M keys -> no warning; actual: 1.16% FP rate 10M keys -> no warning; actual: 1.32% FP rate, 1.13x 25M keys -> no warning; actual: 1.63% FP rate, 1.39x 35M keys -> warning (1.6x); actual: 1.81% FP rate, 1.55x At 5 bits/key: 150K keys -> no warning; actual: 9.32% FP rate (baseline) 25M keys -> no warning; actual: 9.62% FP rate, 1.03x 200M keys -> no warning; actual: 12.2% FP rate, 1.31x 250M keys -> warning (1.5x); actual: 12.8% FP rate, 1.37x 300M keys -> warning (1.6x); actual: 13.4% FP rate, 1.43x The reason for the modest inaccuracy at low bits/key is that the assumption of independence between a collision between 32-hash values feeding the filter and an FP in the filter is not quite true for implementations using "simple" logic to compute indices from the stock hash result. There's math on this in my dissertation, but I don't think it's worth the effort just for these extreme cases (> 100 million keys and low-ish bits/key). Differential Revision: D19471715 Pulled By: pdillinger fbshipit-source-id: f80c96893a09bf1152630ff0b964e5cdd7e35c68
5 years ago
#ifndef NDEBUG
// This could affect build times enough that we should not include it for
// accurate speed tests
#define PREDICT_FP_RATE
#endif
using ROCKSDB_NAMESPACE::Arena;
using ROCKSDB_NAMESPACE::BlockContents;
using ROCKSDB_NAMESPACE::BloomFilterPolicy;
using ROCKSDB_NAMESPACE::BloomHash;
using ROCKSDB_NAMESPACE::BuiltinFilterBitsBuilder;
using ROCKSDB_NAMESPACE::CachableEntry;
Account Bloom/Ribbon filter construction memory in global memory limit (#9073) Summary: Note: This PR is the 4th part of a bigger PR stack (https://github.com/facebook/rocksdb/pull/9073) and will rebase/merge only after the first three PRs (https://github.com/facebook/rocksdb/pull/9070, https://github.com/facebook/rocksdb/pull/9071, https://github.com/facebook/rocksdb/pull/9130) merge. **Context:** Similar to https://github.com/facebook/rocksdb/pull/8428, this PR is to track memory usage during (new) Bloom Filter (i.e,FastLocalBloom) and Ribbon Filter (i.e, Ribbon128) construction, moving toward the goal of [single global memory limit using block cache capacity](https://github.com/facebook/rocksdb/wiki/Projects-Being-Developed#improving-memory-efficiency). It also constrains the size of the banding portion of Ribbon Filter during construction by falling back to Bloom Filter if that banding is, at some point, larger than the available space in the cache under `LRUCacheOptions::strict_capacity_limit=true`. The option to turn on this feature is `BlockBasedTableOptions::reserve_table_builder_memory = true` which by default is set to `false`. We [decided](https://github.com/facebook/rocksdb/pull/9073#discussion_r741548409) not to have separate option for separate memory user in table building therefore their memory accounting are all bundled under one general option. **Summary:** - Reserved/released cache for creation/destruction of three main memory users with the passed-in `FilterBuildingContext::cache_res_mgr` during filter construction: - hash entries (i.e`hash_entries`.size(), we bucket-charge hash entries during insertion for performance), - banding (Ribbon Filter only, `bytes_coeff_rows` +`bytes_result_rows` + `bytes_backtrack`), - final filter (i.e, `mutable_buf`'s size). - Implementation details: in order to use `CacheReservationManager::CacheReservationHandle` to account final filter's memory, we have to store the `CacheReservationManager` object and `CacheReservationHandle` for final filter in `XXPH3BitsFilterBuilder` as well as explicitly delete the filter bits builder when done with the final filter in block based table. - Added option fo run `filter_bench` with this memory reservation feature Pull Request resolved: https://github.com/facebook/rocksdb/pull/9073 Test Plan: - Added new tests in `db_bloom_filter_test` to verify filter construction peak cache reservation under combination of `BlockBasedTable::Rep::FilterType` (e.g, `kFullFilter`, `kPartitionedFilter`), `BloomFilterPolicy::Mode`(e.g, `kFastLocalBloom`, `kStandard128Ribbon`, `kDeprecatedBlock`) and `BlockBasedTableOptions::reserve_table_builder_memory` - To address the concern for slow test: tests with memory reservation under `kFullFilter` + `kStandard128Ribbon` and `kPartitionedFilter` take around **3000 - 6000 ms** and others take around **1500 - 2000 ms**, in total adding **20000 - 25000 ms** to the test suit running locally - Added new test in `bloom_test` to verify Ribbon Filter fallback on large banding in FullFilter - Added test in `filter_bench` to verify that this feature does not significantly slow down Bloom/Ribbon Filter construction speed. Local result averaged over **20** run as below: - FastLocalBloom - baseline `./filter_bench -impl=2 -quick -runs 20 | grep 'Build avg'`: - **Build avg ns/key: 29.56295** (DEBUG_LEVEL=1), **29.98153** (DEBUG_LEVEL=0) - new feature (expected to be similar as above)`./filter_bench -impl=2 -quick -runs 20 -reserve_table_builder_memory=true | grep 'Build avg'`: - **Build avg ns/key: 30.99046** (DEBUG_LEVEL=1), **30.48867** (DEBUG_LEVEL=0) - new feature of RibbonFilter with fallback (expected to be similar as above) `./filter_bench -impl=2 -quick -runs 20 -reserve_table_builder_memory=true -strict_capacity_limit=true | grep 'Build avg'` : - **Build avg ns/key: 31.146975** (DEBUG_LEVEL=1), **30.08165** (DEBUG_LEVEL=0) - Ribbon128 - baseline `./filter_bench -impl=3 -quick -runs 20 | grep 'Build avg'`: - **Build avg ns/key: 129.17585** (DEBUG_LEVEL=1), **130.5225** (DEBUG_LEVEL=0) - new feature (expected to be similar as above) `./filter_bench -impl=3 -quick -runs 20 -reserve_table_builder_memory=true | grep 'Build avg' `: - **Build avg ns/key: 131.61645** (DEBUG_LEVEL=1), **132.98075** (DEBUG_LEVEL=0) - new feature of RibbonFilter with fallback (expected to be a lot faster than above due to fallback) `./filter_bench -impl=3 -quick -runs 20 -reserve_table_builder_memory=true -strict_capacity_limit=true | grep 'Build avg'` : - **Build avg ns/key: 52.032965** (DEBUG_LEVEL=1), **52.597825** (DEBUG_LEVEL=0) - And the warning message of `"Cache reservation for Ribbon filter banding failed due to cache full"` is indeed logged to console. Reviewed By: pdillinger Differential Revision: D31991348 Pulled By: hx235 fbshipit-source-id: 9336b2c60f44d530063da518ceaf56dac5f9df8e
3 years ago
using ROCKSDB_NAMESPACE::Cache;
using ROCKSDB_NAMESPACE::EncodeFixed32;
using ROCKSDB_NAMESPACE::FastRange32;
using ROCKSDB_NAMESPACE::FilterBitsReader;
using ROCKSDB_NAMESPACE::FilterBuildingContext;
using ROCKSDB_NAMESPACE::FullFilterBlockReader;
using ROCKSDB_NAMESPACE::GetSliceHash;
using ROCKSDB_NAMESPACE::GetSliceHash64;
using ROCKSDB_NAMESPACE::Lower32of64;
Account Bloom/Ribbon filter construction memory in global memory limit (#9073) Summary: Note: This PR is the 4th part of a bigger PR stack (https://github.com/facebook/rocksdb/pull/9073) and will rebase/merge only after the first three PRs (https://github.com/facebook/rocksdb/pull/9070, https://github.com/facebook/rocksdb/pull/9071, https://github.com/facebook/rocksdb/pull/9130) merge. **Context:** Similar to https://github.com/facebook/rocksdb/pull/8428, this PR is to track memory usage during (new) Bloom Filter (i.e,FastLocalBloom) and Ribbon Filter (i.e, Ribbon128) construction, moving toward the goal of [single global memory limit using block cache capacity](https://github.com/facebook/rocksdb/wiki/Projects-Being-Developed#improving-memory-efficiency). It also constrains the size of the banding portion of Ribbon Filter during construction by falling back to Bloom Filter if that banding is, at some point, larger than the available space in the cache under `LRUCacheOptions::strict_capacity_limit=true`. The option to turn on this feature is `BlockBasedTableOptions::reserve_table_builder_memory = true` which by default is set to `false`. We [decided](https://github.com/facebook/rocksdb/pull/9073#discussion_r741548409) not to have separate option for separate memory user in table building therefore their memory accounting are all bundled under one general option. **Summary:** - Reserved/released cache for creation/destruction of three main memory users with the passed-in `FilterBuildingContext::cache_res_mgr` during filter construction: - hash entries (i.e`hash_entries`.size(), we bucket-charge hash entries during insertion for performance), - banding (Ribbon Filter only, `bytes_coeff_rows` +`bytes_result_rows` + `bytes_backtrack`), - final filter (i.e, `mutable_buf`'s size). - Implementation details: in order to use `CacheReservationManager::CacheReservationHandle` to account final filter's memory, we have to store the `CacheReservationManager` object and `CacheReservationHandle` for final filter in `XXPH3BitsFilterBuilder` as well as explicitly delete the filter bits builder when done with the final filter in block based table. - Added option fo run `filter_bench` with this memory reservation feature Pull Request resolved: https://github.com/facebook/rocksdb/pull/9073 Test Plan: - Added new tests in `db_bloom_filter_test` to verify filter construction peak cache reservation under combination of `BlockBasedTable::Rep::FilterType` (e.g, `kFullFilter`, `kPartitionedFilter`), `BloomFilterPolicy::Mode`(e.g, `kFastLocalBloom`, `kStandard128Ribbon`, `kDeprecatedBlock`) and `BlockBasedTableOptions::reserve_table_builder_memory` - To address the concern for slow test: tests with memory reservation under `kFullFilter` + `kStandard128Ribbon` and `kPartitionedFilter` take around **3000 - 6000 ms** and others take around **1500 - 2000 ms**, in total adding **20000 - 25000 ms** to the test suit running locally - Added new test in `bloom_test` to verify Ribbon Filter fallback on large banding in FullFilter - Added test in `filter_bench` to verify that this feature does not significantly slow down Bloom/Ribbon Filter construction speed. Local result averaged over **20** run as below: - FastLocalBloom - baseline `./filter_bench -impl=2 -quick -runs 20 | grep 'Build avg'`: - **Build avg ns/key: 29.56295** (DEBUG_LEVEL=1), **29.98153** (DEBUG_LEVEL=0) - new feature (expected to be similar as above)`./filter_bench -impl=2 -quick -runs 20 -reserve_table_builder_memory=true | grep 'Build avg'`: - **Build avg ns/key: 30.99046** (DEBUG_LEVEL=1), **30.48867** (DEBUG_LEVEL=0) - new feature of RibbonFilter with fallback (expected to be similar as above) `./filter_bench -impl=2 -quick -runs 20 -reserve_table_builder_memory=true -strict_capacity_limit=true | grep 'Build avg'` : - **Build avg ns/key: 31.146975** (DEBUG_LEVEL=1), **30.08165** (DEBUG_LEVEL=0) - Ribbon128 - baseline `./filter_bench -impl=3 -quick -runs 20 | grep 'Build avg'`: - **Build avg ns/key: 129.17585** (DEBUG_LEVEL=1), **130.5225** (DEBUG_LEVEL=0) - new feature (expected to be similar as above) `./filter_bench -impl=3 -quick -runs 20 -reserve_table_builder_memory=true | grep 'Build avg' `: - **Build avg ns/key: 131.61645** (DEBUG_LEVEL=1), **132.98075** (DEBUG_LEVEL=0) - new feature of RibbonFilter with fallback (expected to be a lot faster than above due to fallback) `./filter_bench -impl=3 -quick -runs 20 -reserve_table_builder_memory=true -strict_capacity_limit=true | grep 'Build avg'` : - **Build avg ns/key: 52.032965** (DEBUG_LEVEL=1), **52.597825** (DEBUG_LEVEL=0) - And the warning message of `"Cache reservation for Ribbon filter banding failed due to cache full"` is indeed logged to console. Reviewed By: pdillinger Differential Revision: D31991348 Pulled By: hx235 fbshipit-source-id: 9336b2c60f44d530063da518ceaf56dac5f9df8e
3 years ago
using ROCKSDB_NAMESPACE::LRUCacheOptions;
using ROCKSDB_NAMESPACE::ParsedFullFilterBlock;
using ROCKSDB_NAMESPACE::PlainTableBloomV1;
using ROCKSDB_NAMESPACE::Random32;
using ROCKSDB_NAMESPACE::Slice;
using ROCKSDB_NAMESPACE::static_cast_with_check;
using ROCKSDB_NAMESPACE::StderrLogger;
using ROCKSDB_NAMESPACE::mock::MockBlockBasedTableTester;
struct KeyMaker {
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
KeyMaker(size_t avg_size)
: smallest_size_(avg_size -
(FLAGS_vary_key_size_log2_interval >= 30 ? 2 : 0)),
buf_size_(avg_size + 11), // pad to vary key size and alignment
buf_(new char[buf_size_]) {
memset(buf_.get(), 0, buf_size_);
assert(smallest_size_ > 8);
}
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
size_t smallest_size_;
size_t buf_size_;
std::unique_ptr<char[]> buf_;
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
// Returns a unique(-ish) key based on the given parameter values. Each
// call returns a Slice from the same buffer so previously returned
// Slices should be considered invalidated.
Slice Get(uint32_t filter_num, uint32_t val_num) {
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
size_t start = FLAGS_vary_key_alignment ? val_num % 4 : 0;
size_t len = smallest_size_;
if (FLAGS_vary_key_size_log2_interval < 30) {
// To get range [avg_size - 2, avg_size + 2]
// use range [smallest_size, smallest_size + 4]
len += FastRange32(
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
(val_num >> FLAGS_vary_key_size_log2_interval) * 1234567891, 5);
}
char * data = buf_.get() + start;
// Populate key data such that all data makes it into a key of at
// least 8 bytes. We also don't want all the within-filter key
// variance confined to a contiguous 32 bits, because then a 32 bit
// hash function can "cheat" the false positive rate by
// approximating a perfect hash.
EncodeFixed32(data, val_num);
EncodeFixed32(data + 4, filter_num + val_num);
// ensure clearing leftovers from different alignment
EncodeFixed32(data + 8, 0);
return Slice(data, len);
}
};
void PrintWarnings() {
#if defined(__GNUC__) && !defined(__OPTIMIZE__)
fprintf(stdout,
"WARNING: Optimization is disabled: benchmarks unnecessarily slow\n");
#endif
#ifndef NDEBUG
fprintf(stdout,
"WARNING: Assertions are enabled; benchmarks unnecessarily slow\n");
#endif
}
struct FilterInfo {
uint32_t filter_id_ = 0;
std::unique_ptr<const char[]> owner_;
Slice filter_;
uint32_t keys_added_ = 0;
std::unique_ptr<FilterBitsReader> reader_;
std::unique_ptr<FullFilterBlockReader> full_block_reader_;
std::unique_ptr<PlainTableBloomV1> plain_table_bloom_;
uint64_t outside_queries_ = 0;
uint64_t false_positives_ = 0;
};
enum TestMode {
kSingleFilter,
kBatchPrepared,
kBatchUnprepared,
kFiftyOneFilter,
kEightyTwentyFilter,
kRandomFilter,
};
static const std::vector<TestMode> allTestModes = {
kSingleFilter, kBatchPrepared, kBatchUnprepared,
kFiftyOneFilter, kEightyTwentyFilter, kRandomFilter,
};
static const std::vector<TestMode> quickTestModes = {
kSingleFilter,
kRandomFilter,
};
static const std::vector<TestMode> bestCaseTestModes = {
kSingleFilter,
};
const char *TestModeToString(TestMode tm) {
switch (tm) {
case kSingleFilter:
return "Single filter";
case kBatchPrepared:
return "Batched, prepared";
case kBatchUnprepared:
return "Batched, unprepared";
case kFiftyOneFilter:
return "Skewed 50% in 1%";
case kEightyTwentyFilter:
return "Skewed 80% in 20%";
case kRandomFilter:
return "Random filter";
}
return "Bad TestMode";
}
// Do just enough to keep some data dependence for the
// compiler / CPU
Add new persistent 64-bit hash (#5984) 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
5 years ago
static uint32_t DryRunNoHash(Slice &s) {
uint32_t sz = static_cast<uint32_t>(s.size());
if (sz >= 4) {
return sz + s.data()[3];
} else {
return sz;
}
}
Add new persistent 64-bit hash (#5984) 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
5 years ago
static uint32_t DryRunHash32(Slice &s) {
// Same perf characteristics as GetSliceHash()
return BloomHash(s);
}
static uint32_t DryRunHash64(Slice &s) {
return Lower32of64(GetSliceHash64(s));
}
struct FilterBench : public MockBlockBasedTableTester {
std::vector<KeyMaker> kms_;
std::vector<FilterInfo> infos_;
Random32 random_;
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
std::ostringstream fp_rate_report_;
Arena arena_;
double m_queries_;
StderrLogger stderr_logger_;
FilterBench()
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
5 years ago
: MockBlockBasedTableTester(new BloomFilterPolicy(
FLAGS_bits_per_key,
static_cast<BloomFilterPolicy::Mode>(FLAGS_impl))),
random_(FLAGS_seed),
m_queries_(0) {
for (uint32_t i = 0; i < FLAGS_batch_size; ++i) {
kms_.emplace_back(FLAGS_key_size < 8 ? 8 : FLAGS_key_size);
}
ioptions_.logger = &stderr_logger_;
Minimize memory internal fragmentation for Bloom filters (#6427) Summary: New experimental option BBTO::optimize_filters_for_memory builds filters that maximize their use of "usable size" from malloc_usable_size, which is also used to compute block cache charges. Rather than always "rounding up," we track state in the BloomFilterPolicy object to mix essentially "rounding down" and "rounding up" so that the average FP rate of all generated filters is the same as without the option. (YMMV as heavily accessed filters might be unluckily lower accuracy.) Thus, the option near-minimizes what the block cache considers as "memory used" for a given target Bloom filter false positive rate and Bloom filter implementation. There are no forward or backward compatibility issues with this change, though it only works on the format_version=5 Bloom filter. With Jemalloc, we see about 10% reduction in memory footprint (and block cache charge) for Bloom filters, but 1-2% increase in storage footprint, due to encoding efficiency losses (FP rate is non-linear with bits/key). Why not weighted random round up/down rather than state tracking? By only requiring malloc_usable_size, we don't actually know what the next larger and next smaller usable sizes for the allocator are. We pick a requested size, accept and use whatever usable size it has, and use the difference to inform our next choice. This allows us to narrow in on the right balance without tracking/predicting usable sizes. Why not weight history of generated filter false positive rates by number of keys? This could lead to excess skew in small filters after generating a large filter. Results from filter_bench with jemalloc (irrelevant details omitted): (normal keys/filter, but high variance) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=30000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.6278 Number of filters: 5516 Total size (MB): 200.046 Reported total allocated memory (MB): 220.597 Reported internal fragmentation: 10.2732% Bits/key stored: 10.0097 Average FP rate %: 0.965228 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=30000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 30.5104 Number of filters: 5464 Total size (MB): 200.015 Reported total allocated memory (MB): 200.322 Reported internal fragmentation: 0.153709% Bits/key stored: 10.1011 Average FP rate %: 0.966313 (very few keys / filter, optimization not as effective due to ~59 byte internal fragmentation in blocked Bloom filter representation) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.5649 Number of filters: 162950 Total size (MB): 200.001 Reported total allocated memory (MB): 224.624 Reported internal fragmentation: 12.3117% Bits/key stored: 10.2951 Average FP rate %: 0.821534 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 31.8057 Number of filters: 159849 Total size (MB): 200 Reported total allocated memory (MB): 208.846 Reported internal fragmentation: 4.42297% Bits/key stored: 10.4948 Average FP rate %: 0.811006 (high keys/filter) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.7017 Number of filters: 164 Total size (MB): 200.352 Reported total allocated memory (MB): 221.5 Reported internal fragmentation: 10.5552% Bits/key stored: 10.0003 Average FP rate %: 0.969358 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 30.7131 Number of filters: 160 Total size (MB): 200.928 Reported total allocated memory (MB): 200.938 Reported internal fragmentation: 0.00448054% Bits/key stored: 10.1852 Average FP rate %: 0.963387 And from db_bench (block cache) with jemalloc: $ ./db_bench -db=/dev/shm/dbbench.no_optimize -benchmarks=fillrandom -format_version=5 -value_size=90 -bloom_bits=10 -num=2000000 -threads=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false $ ./db_bench -db=/dev/shm/dbbench -benchmarks=fillrandom -format_version=5 -value_size=90 -bloom_bits=10 -num=2000000 -threads=8 -optimize_filters_for_memory -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false $ (for FILE in /dev/shm/dbbench.no_optimize/*.sst; do ./sst_dump --file=$FILE --show_properties | grep 'filter block' ; done) | awk '{ t += $4; } END { print t; }' 17063835 $ (for FILE in /dev/shm/dbbench/*.sst; do ./sst_dump --file=$FILE --show_properties | grep 'filter block' ; done) | awk '{ t += $4; } END { print t; }' 17430747 $ #^ 2.1% additional filter storage $ ./db_bench -db=/dev/shm/dbbench.no_optimize -use_existing_db -benchmarks=readrandom,stats -statistics -bloom_bits=10 -num=2000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false -duration=10 -cache_index_and_filter_blocks -cache_size=1000000000 rocksdb.block.cache.index.add COUNT : 33 rocksdb.block.cache.index.bytes.insert COUNT : 8440400 rocksdb.block.cache.filter.add COUNT : 33 rocksdb.block.cache.filter.bytes.insert COUNT : 21087528 rocksdb.bloom.filter.useful COUNT : 4963889 rocksdb.bloom.filter.full.positive COUNT : 1214081 rocksdb.bloom.filter.full.true.positive COUNT : 1161999 $ #^ 1.04 % observed FP rate $ ./db_bench -db=/dev/shm/dbbench -use_existing_db -benchmarks=readrandom,stats -statistics -bloom_bits=10 -num=2000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false -optimize_filters_for_memory -duration=10 -cache_index_and_filter_blocks -cache_size=1000000000 rocksdb.block.cache.index.add COUNT : 33 rocksdb.block.cache.index.bytes.insert COUNT : 8448592 rocksdb.block.cache.filter.add COUNT : 33 rocksdb.block.cache.filter.bytes.insert COUNT : 18220328 rocksdb.bloom.filter.useful COUNT : 5360933 rocksdb.bloom.filter.full.positive COUNT : 1321315 rocksdb.bloom.filter.full.true.positive COUNT : 1262999 $ #^ 1.08 % observed FP rate, 13.6% less memory usage for filters (Due to specific key density, this example tends to generate filters that are "worse than average" for internal fragmentation. "Better than average" cases can show little or no improvement.) Pull Request resolved: https://github.com/facebook/rocksdb/pull/6427 Test Plan: unit test added, 'make check' with gcc, clang and valgrind Reviewed By: siying Differential Revision: D22124374 Pulled By: pdillinger fbshipit-source-id: f3e3aa152f9043ddf4fae25799e76341d0d8714e
5 years ago
table_options_.optimize_filters_for_memory =
FLAGS_optimize_filters_for_memory;
Account Bloom/Ribbon filter construction memory in global memory limit (#9073) Summary: Note: This PR is the 4th part of a bigger PR stack (https://github.com/facebook/rocksdb/pull/9073) and will rebase/merge only after the first three PRs (https://github.com/facebook/rocksdb/pull/9070, https://github.com/facebook/rocksdb/pull/9071, https://github.com/facebook/rocksdb/pull/9130) merge. **Context:** Similar to https://github.com/facebook/rocksdb/pull/8428, this PR is to track memory usage during (new) Bloom Filter (i.e,FastLocalBloom) and Ribbon Filter (i.e, Ribbon128) construction, moving toward the goal of [single global memory limit using block cache capacity](https://github.com/facebook/rocksdb/wiki/Projects-Being-Developed#improving-memory-efficiency). It also constrains the size of the banding portion of Ribbon Filter during construction by falling back to Bloom Filter if that banding is, at some point, larger than the available space in the cache under `LRUCacheOptions::strict_capacity_limit=true`. The option to turn on this feature is `BlockBasedTableOptions::reserve_table_builder_memory = true` which by default is set to `false`. We [decided](https://github.com/facebook/rocksdb/pull/9073#discussion_r741548409) not to have separate option for separate memory user in table building therefore their memory accounting are all bundled under one general option. **Summary:** - Reserved/released cache for creation/destruction of three main memory users with the passed-in `FilterBuildingContext::cache_res_mgr` during filter construction: - hash entries (i.e`hash_entries`.size(), we bucket-charge hash entries during insertion for performance), - banding (Ribbon Filter only, `bytes_coeff_rows` +`bytes_result_rows` + `bytes_backtrack`), - final filter (i.e, `mutable_buf`'s size). - Implementation details: in order to use `CacheReservationManager::CacheReservationHandle` to account final filter's memory, we have to store the `CacheReservationManager` object and `CacheReservationHandle` for final filter in `XXPH3BitsFilterBuilder` as well as explicitly delete the filter bits builder when done with the final filter in block based table. - Added option fo run `filter_bench` with this memory reservation feature Pull Request resolved: https://github.com/facebook/rocksdb/pull/9073 Test Plan: - Added new tests in `db_bloom_filter_test` to verify filter construction peak cache reservation under combination of `BlockBasedTable::Rep::FilterType` (e.g, `kFullFilter`, `kPartitionedFilter`), `BloomFilterPolicy::Mode`(e.g, `kFastLocalBloom`, `kStandard128Ribbon`, `kDeprecatedBlock`) and `BlockBasedTableOptions::reserve_table_builder_memory` - To address the concern for slow test: tests with memory reservation under `kFullFilter` + `kStandard128Ribbon` and `kPartitionedFilter` take around **3000 - 6000 ms** and others take around **1500 - 2000 ms**, in total adding **20000 - 25000 ms** to the test suit running locally - Added new test in `bloom_test` to verify Ribbon Filter fallback on large banding in FullFilter - Added test in `filter_bench` to verify that this feature does not significantly slow down Bloom/Ribbon Filter construction speed. Local result averaged over **20** run as below: - FastLocalBloom - baseline `./filter_bench -impl=2 -quick -runs 20 | grep 'Build avg'`: - **Build avg ns/key: 29.56295** (DEBUG_LEVEL=1), **29.98153** (DEBUG_LEVEL=0) - new feature (expected to be similar as above)`./filter_bench -impl=2 -quick -runs 20 -reserve_table_builder_memory=true | grep 'Build avg'`: - **Build avg ns/key: 30.99046** (DEBUG_LEVEL=1), **30.48867** (DEBUG_LEVEL=0) - new feature of RibbonFilter with fallback (expected to be similar as above) `./filter_bench -impl=2 -quick -runs 20 -reserve_table_builder_memory=true -strict_capacity_limit=true | grep 'Build avg'` : - **Build avg ns/key: 31.146975** (DEBUG_LEVEL=1), **30.08165** (DEBUG_LEVEL=0) - Ribbon128 - baseline `./filter_bench -impl=3 -quick -runs 20 | grep 'Build avg'`: - **Build avg ns/key: 129.17585** (DEBUG_LEVEL=1), **130.5225** (DEBUG_LEVEL=0) - new feature (expected to be similar as above) `./filter_bench -impl=3 -quick -runs 20 -reserve_table_builder_memory=true | grep 'Build avg' `: - **Build avg ns/key: 131.61645** (DEBUG_LEVEL=1), **132.98075** (DEBUG_LEVEL=0) - new feature of RibbonFilter with fallback (expected to be a lot faster than above due to fallback) `./filter_bench -impl=3 -quick -runs 20 -reserve_table_builder_memory=true -strict_capacity_limit=true | grep 'Build avg'` : - **Build avg ns/key: 52.032965** (DEBUG_LEVEL=1), **52.597825** (DEBUG_LEVEL=0) - And the warning message of `"Cache reservation for Ribbon filter banding failed due to cache full"` is indeed logged to console. Reviewed By: pdillinger Differential Revision: D31991348 Pulled By: hx235 fbshipit-source-id: 9336b2c60f44d530063da518ceaf56dac5f9df8e
3 years ago
if (FLAGS_reserve_table_builder_memory) {
table_options_.reserve_table_builder_memory = true;
table_options_.no_block_cache = false;
LRUCacheOptions lo;
lo.capacity = FLAGS_block_cache_capacity_MB * 1024 * 1024;
lo.num_shard_bits = 0; // 2^0 shard
lo.strict_capacity_limit = FLAGS_strict_capacity_limit;
std::shared_ptr<Cache> cache(NewLRUCache(lo));
table_options_.block_cache = cache;
}
}
void Go();
double RandomQueryTest(uint32_t inside_threshold, bool dry_run,
TestMode mode);
};
void FilterBench::Go() {
if (FLAGS_use_plain_table_bloom && FLAGS_use_full_block_reader) {
throw std::runtime_error(
"Can't combine -use_plain_table_bloom and -use_full_block_reader");
}
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
5 years ago
if (FLAGS_use_plain_table_bloom) {
if (FLAGS_impl > 1) {
throw std::runtime_error(
"-impl must currently be >= 0 and <= 1 for Plain table");
}
} else {
if (FLAGS_impl == 1) {
throw std::runtime_error(
"Block-based filter not currently supported by filter_bench");
}
Experimental (production candidate) SST schema for Ribbon filter (#7658) Summary: Added experimental public API for Ribbon filter: NewExperimentalRibbonFilterPolicy(). This experimental API will take a "Bloom equivalent" bits per key, and configure the Ribbon filter for the same FP rate as Bloom would have but ~30% space savings. (Note: optimize_filters_for_memory is not yet implemented for Ribbon filter. That can be added with no effect on schema.) Internally, the Ribbon filter is configured using a "one_in_fp_rate" value, which is 1 over desired FP rate. For example, use 100 for 1% FP rate. I'm expecting this will be used in the future for configuring Bloom-like filters, as I expect people to more commonly hold constant the filter accuracy and change the space vs. time trade-off, rather than hold constant the space (per key) and change the accuracy vs. time trade-off, though we might make that available. ### Benchmarking ``` $ ./filter_bench -impl=2 -quick -m_keys_total_max=200 -average_keys_per_filter=100000 -net_includes_hashing Building... Build avg ns/key: 34.1341 Number of filters: 1993 Total size (MB): 238.488 Reported total allocated memory (MB): 262.875 Reported internal fragmentation: 10.2255% Bits/key stored: 10.0029 ---------------------------- Mixed inside/outside queries... Single filter net ns/op: 18.7508 Random filter net ns/op: 258.246 Average FP rate %: 0.968672 ---------------------------- Done. (For more info, run with -legend or -help.) $ ./filter_bench -impl=3 -quick -m_keys_total_max=200 -average_keys_per_filter=100000 -net_includes_hashing Building... Build avg ns/key: 130.851 Number of filters: 1993 Total size (MB): 168.166 Reported total allocated memory (MB): 183.211 Reported internal fragmentation: 8.94626% Bits/key stored: 7.05341 ---------------------------- Mixed inside/outside queries... Single filter net ns/op: 58.4523 Random filter net ns/op: 363.717 Average FP rate %: 0.952978 ---------------------------- Done. (For more info, run with -legend or -help.) ``` 168.166 / 238.488 = 0.705 -> 29.5% space reduction 130.851 / 34.1341 = 3.83x construction time for this Ribbon filter vs. lastest Bloom filter (could make that as little as about 2.5x for less space reduction) ### Working around a hashing "flaw" bloom_test discovered a flaw in the simple hashing applied in StandardHasher when num_starts == 1 (num_slots == 128), showing an excessively high FP rate. The problem is that when many entries, on the order of number of hash bits or kCoeffBits, are associated with the same start location, the correlation between the CoeffRow and ResultRow (for efficiency) can lead to a solution that is "universal," or nearly so, for entries mapping to that start location. (Normally, variance in start location breaks the effective association between CoeffRow and ResultRow; the same value for CoeffRow is effectively different if start locations are different.) Without kUseSmash and with num_starts > 1 (thus num_starts ~= num_slots), this flaw should be completely irrelevant. Even with 10M slots, the chances of a single slot having just 16 (or more) entries map to it--not enough to cause an FP problem, which would be local to that slot if it happened--is 1 in millions. This spreadsheet formula shows that: =1/(10000000*(1 - POISSON(15, 1, TRUE))) As kUseSmash==false (the setting for Standard128RibbonBitsBuilder) is intended for CPU efficiency of filters with many more entries/slots than kCoeffBits, a very reasonable work-around is to disallow num_starts==1 when !kUseSmash, by making the minimum non-zero number of slots 2*kCoeffBits. This is the work-around I've applied. This also means that the new Ribbon filter schema (Standard128RibbonBitsBuilder) is not space-efficient for less than a few hundred entries. Because of this, I have made it fall back on constructing a Bloom filter, under existing schema, when that is more space efficient for small filters. (We can change this in the future if we want.) TODO: better unit tests for this case in ribbon_test, and probably update StandardHasher for kUseSmash case so that it can scale nicely to small filters. ### Other related changes * Add Ribbon filter to stress/crash test * Add Ribbon filter to filter_bench as -impl=3 * Add option string support, as in "filter_policy=experimental_ribbon:5.678;" where 5.678 is the Bloom equivalent bits per key. * Rename internal mode BloomFilterPolicy::kAuto to kAutoBloom * Add a general BuiltinFilterBitsBuilder::CalculateNumEntry based on binary searching CalculateSpace (inefficient), so that subclasses (especially experimental ones) don't have to provide an efficient implementation inverting CalculateSpace. * Minor refactor FastLocalBloomBitsBuilder for new base class XXH3pFilterBitsBuilder shared with new Standard128RibbonBitsBuilder, which allows the latter to fall back on Bloom construction in some extreme cases. * Mostly updated bloom_test for Ribbon filter, though a test like FullBloomTest::Schema is a next TODO to ensure schema stability (in case this becomes production-ready schema as it is). * Add some APIs to ribbon_impl.h for configuring Ribbon filters. Although these are reasonably covered by bloom_test, TODO more unit tests in ribbon_test * Added a "tool" FindOccupancyForSuccessRate to ribbon_test to get data for constructing the linear approximations in GetNumSlotsFor95PctSuccess. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7658 Test Plan: Some unit tests updated but other testing is left TODO. This is considered experimental but laying down schema compatibility as early as possible in case it proves production-quality. Also tested in stress/crash test. Reviewed By: jay-zhuang Differential Revision: D24899349 Pulled By: pdillinger fbshipit-source-id: 9715f3e6371c959d923aea8077c9423c7a9f82b8
4 years ago
if (FLAGS_impl > 3) {
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
5 years ago
throw std::runtime_error(
Experimental (production candidate) SST schema for Ribbon filter (#7658) Summary: Added experimental public API for Ribbon filter: NewExperimentalRibbonFilterPolicy(). This experimental API will take a "Bloom equivalent" bits per key, and configure the Ribbon filter for the same FP rate as Bloom would have but ~30% space savings. (Note: optimize_filters_for_memory is not yet implemented for Ribbon filter. That can be added with no effect on schema.) Internally, the Ribbon filter is configured using a "one_in_fp_rate" value, which is 1 over desired FP rate. For example, use 100 for 1% FP rate. I'm expecting this will be used in the future for configuring Bloom-like filters, as I expect people to more commonly hold constant the filter accuracy and change the space vs. time trade-off, rather than hold constant the space (per key) and change the accuracy vs. time trade-off, though we might make that available. ### Benchmarking ``` $ ./filter_bench -impl=2 -quick -m_keys_total_max=200 -average_keys_per_filter=100000 -net_includes_hashing Building... Build avg ns/key: 34.1341 Number of filters: 1993 Total size (MB): 238.488 Reported total allocated memory (MB): 262.875 Reported internal fragmentation: 10.2255% Bits/key stored: 10.0029 ---------------------------- Mixed inside/outside queries... Single filter net ns/op: 18.7508 Random filter net ns/op: 258.246 Average FP rate %: 0.968672 ---------------------------- Done. (For more info, run with -legend or -help.) $ ./filter_bench -impl=3 -quick -m_keys_total_max=200 -average_keys_per_filter=100000 -net_includes_hashing Building... Build avg ns/key: 130.851 Number of filters: 1993 Total size (MB): 168.166 Reported total allocated memory (MB): 183.211 Reported internal fragmentation: 8.94626% Bits/key stored: 7.05341 ---------------------------- Mixed inside/outside queries... Single filter net ns/op: 58.4523 Random filter net ns/op: 363.717 Average FP rate %: 0.952978 ---------------------------- Done. (For more info, run with -legend or -help.) ``` 168.166 / 238.488 = 0.705 -> 29.5% space reduction 130.851 / 34.1341 = 3.83x construction time for this Ribbon filter vs. lastest Bloom filter (could make that as little as about 2.5x for less space reduction) ### Working around a hashing "flaw" bloom_test discovered a flaw in the simple hashing applied in StandardHasher when num_starts == 1 (num_slots == 128), showing an excessively high FP rate. The problem is that when many entries, on the order of number of hash bits or kCoeffBits, are associated with the same start location, the correlation between the CoeffRow and ResultRow (for efficiency) can lead to a solution that is "universal," or nearly so, for entries mapping to that start location. (Normally, variance in start location breaks the effective association between CoeffRow and ResultRow; the same value for CoeffRow is effectively different if start locations are different.) Without kUseSmash and with num_starts > 1 (thus num_starts ~= num_slots), this flaw should be completely irrelevant. Even with 10M slots, the chances of a single slot having just 16 (or more) entries map to it--not enough to cause an FP problem, which would be local to that slot if it happened--is 1 in millions. This spreadsheet formula shows that: =1/(10000000*(1 - POISSON(15, 1, TRUE))) As kUseSmash==false (the setting for Standard128RibbonBitsBuilder) is intended for CPU efficiency of filters with many more entries/slots than kCoeffBits, a very reasonable work-around is to disallow num_starts==1 when !kUseSmash, by making the minimum non-zero number of slots 2*kCoeffBits. This is the work-around I've applied. This also means that the new Ribbon filter schema (Standard128RibbonBitsBuilder) is not space-efficient for less than a few hundred entries. Because of this, I have made it fall back on constructing a Bloom filter, under existing schema, when that is more space efficient for small filters. (We can change this in the future if we want.) TODO: better unit tests for this case in ribbon_test, and probably update StandardHasher for kUseSmash case so that it can scale nicely to small filters. ### Other related changes * Add Ribbon filter to stress/crash test * Add Ribbon filter to filter_bench as -impl=3 * Add option string support, as in "filter_policy=experimental_ribbon:5.678;" where 5.678 is the Bloom equivalent bits per key. * Rename internal mode BloomFilterPolicy::kAuto to kAutoBloom * Add a general BuiltinFilterBitsBuilder::CalculateNumEntry based on binary searching CalculateSpace (inefficient), so that subclasses (especially experimental ones) don't have to provide an efficient implementation inverting CalculateSpace. * Minor refactor FastLocalBloomBitsBuilder for new base class XXH3pFilterBitsBuilder shared with new Standard128RibbonBitsBuilder, which allows the latter to fall back on Bloom construction in some extreme cases. * Mostly updated bloom_test for Ribbon filter, though a test like FullBloomTest::Schema is a next TODO to ensure schema stability (in case this becomes production-ready schema as it is). * Add some APIs to ribbon_impl.h for configuring Ribbon filters. Although these are reasonably covered by bloom_test, TODO more unit tests in ribbon_test * Added a "tool" FindOccupancyForSuccessRate to ribbon_test to get data for constructing the linear approximations in GetNumSlotsFor95PctSuccess. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7658 Test Plan: Some unit tests updated but other testing is left TODO. This is considered experimental but laying down schema compatibility as early as possible in case it proves production-quality. Also tested in stress/crash test. Reviewed By: jay-zhuang Differential Revision: D24899349 Pulled By: pdillinger fbshipit-source-id: 9715f3e6371c959d923aea8077c9423c7a9f82b8
4 years ago
"-impl must currently be 0, 2, or 3 for Block-based table");
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
5 years ago
}
}
if (FLAGS_vary_key_count_ratio < 0.0 || FLAGS_vary_key_count_ratio > 1.0) {
throw std::runtime_error("-vary_key_count_ratio must be >= 0.0 and <= 1.0");
}
// For example, average_keys_per_filter = 100, vary_key_count_ratio = 0.1.
// Varys up to +/- 10 keys. variance_range = 21 (generating value 0..20).
// variance_offset = 10, so value - offset average value is always 0.
const uint32_t variance_range =
1 + 2 * static_cast<uint32_t>(FLAGS_vary_key_count_ratio *
FLAGS_average_keys_per_filter);
const uint32_t variance_offset = variance_range / 2;
const std::vector<TestMode> &testModes =
FLAGS_best_case ? bestCaseTestModes
: FLAGS_quick ? quickTestModes : allTestModes;
m_queries_ = FLAGS_m_queries;
double working_mem_size_mb = FLAGS_working_mem_size_mb;
if (FLAGS_quick) {
m_queries_ /= 7.0;
} else if (FLAGS_best_case) {
m_queries_ /= 3.0;
working_mem_size_mb /= 10.0;
}
std::cout << "Building..." << std::endl;
Warn on excessive keys for legacy Bloom filter with 32-bit hash (#6317) Summary: With many millions of keys, the old Bloom filter implementation for the block-based table (format_version <= 4) would have excessive FP rate due to the limitations of feeding the Bloom filter with a 32-bit hash. This change computes an estimated inflated FP rate due to this effect and warns in the log whenever an SST filter is constructed (almost certainly a "full" not "partitioned" filter) that exceeds 1.5x FP rate due to this effect. The detailed condition is only checked if 3 million keys or more have been added to a filter, as this should be a lower bound for common bits/key settings (< 20). Recommended remedies include smaller SST file size, using format_version >= 5 (for new Bloom filter), or using partitioned filters. This does not change behavior other than generating warnings for some constructed filters using the old implementation. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6317 Test Plan: Example with warning, 15M keys @ 15 bits / key: (working_mem_size_mb is just to stop after building one filter if it's large) $ ./filter_bench -quick -impl=0 -working_mem_size_mb=1 -bits_per_key=15 -average_keys_per_filter=15000000 2>&1 | grep 'FP rate' [WARN] [/block_based/filter_policy.cc:292] Using legacy SST/BBT Bloom filter with excessive key count (15.0M @ 15bpk), causing estimated 1.8x higher filter FP rate. Consider using new Bloom with format_version>=5, smaller SST file size, or partitioned filters. Predicted FP rate %: 0.766702 Average FP rate %: 0.66846 Example without warning (150K keys): $ ./filter_bench -quick -impl=0 -working_mem_size_mb=1 -bits_per_key=15 -average_keys_per_filter=150000 2>&1 | grep 'FP rate' Predicted FP rate %: 0.422857 Average FP rate %: 0.379301 $ With more samples at 15 bits/key: 150K keys -> no warning; actual: 0.379% FP rate (baseline) 1M keys -> no warning; actual: 0.396% FP rate, 1.045x 9M keys -> no warning; actual: 0.563% FP rate, 1.485x 10M keys -> warning (1.5x); actual: 0.564% FP rate, 1.488x 15M keys -> warning (1.8x); actual: 0.668% FP rate, 1.76x 25M keys -> warning (2.4x); actual: 0.880% FP rate, 2.32x At 10 bits/key: 150K keys -> no warning; actual: 1.17% FP rate (baseline) 1M keys -> no warning; actual: 1.16% FP rate 10M keys -> no warning; actual: 1.32% FP rate, 1.13x 25M keys -> no warning; actual: 1.63% FP rate, 1.39x 35M keys -> warning (1.6x); actual: 1.81% FP rate, 1.55x At 5 bits/key: 150K keys -> no warning; actual: 9.32% FP rate (baseline) 25M keys -> no warning; actual: 9.62% FP rate, 1.03x 200M keys -> no warning; actual: 12.2% FP rate, 1.31x 250M keys -> warning (1.5x); actual: 12.8% FP rate, 1.37x 300M keys -> warning (1.6x); actual: 13.4% FP rate, 1.43x The reason for the modest inaccuracy at low bits/key is that the assumption of independence between a collision between 32-hash values feeding the filter and an FP in the filter is not quite true for implementations using "simple" logic to compute indices from the stock hash result. There's math on this in my dissertation, but I don't think it's worth the effort just for these extreme cases (> 100 million keys and low-ish bits/key). Differential Revision: D19471715 Pulled By: pdillinger fbshipit-source-id: f80c96893a09bf1152630ff0b964e5cdd7e35c68
5 years ago
std::unique_ptr<BuiltinFilterBitsBuilder> builder;
Optimize memory and CPU for building new Bloom filter (#6175) 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
5 years ago
size_t total_memory_used = 0;
Minimize memory internal fragmentation for Bloom filters (#6427) Summary: New experimental option BBTO::optimize_filters_for_memory builds filters that maximize their use of "usable size" from malloc_usable_size, which is also used to compute block cache charges. Rather than always "rounding up," we track state in the BloomFilterPolicy object to mix essentially "rounding down" and "rounding up" so that the average FP rate of all generated filters is the same as without the option. (YMMV as heavily accessed filters might be unluckily lower accuracy.) Thus, the option near-minimizes what the block cache considers as "memory used" for a given target Bloom filter false positive rate and Bloom filter implementation. There are no forward or backward compatibility issues with this change, though it only works on the format_version=5 Bloom filter. With Jemalloc, we see about 10% reduction in memory footprint (and block cache charge) for Bloom filters, but 1-2% increase in storage footprint, due to encoding efficiency losses (FP rate is non-linear with bits/key). Why not weighted random round up/down rather than state tracking? By only requiring malloc_usable_size, we don't actually know what the next larger and next smaller usable sizes for the allocator are. We pick a requested size, accept and use whatever usable size it has, and use the difference to inform our next choice. This allows us to narrow in on the right balance without tracking/predicting usable sizes. Why not weight history of generated filter false positive rates by number of keys? This could lead to excess skew in small filters after generating a large filter. Results from filter_bench with jemalloc (irrelevant details omitted): (normal keys/filter, but high variance) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=30000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.6278 Number of filters: 5516 Total size (MB): 200.046 Reported total allocated memory (MB): 220.597 Reported internal fragmentation: 10.2732% Bits/key stored: 10.0097 Average FP rate %: 0.965228 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=30000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 30.5104 Number of filters: 5464 Total size (MB): 200.015 Reported total allocated memory (MB): 200.322 Reported internal fragmentation: 0.153709% Bits/key stored: 10.1011 Average FP rate %: 0.966313 (very few keys / filter, optimization not as effective due to ~59 byte internal fragmentation in blocked Bloom filter representation) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.5649 Number of filters: 162950 Total size (MB): 200.001 Reported total allocated memory (MB): 224.624 Reported internal fragmentation: 12.3117% Bits/key stored: 10.2951 Average FP rate %: 0.821534 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 31.8057 Number of filters: 159849 Total size (MB): 200 Reported total allocated memory (MB): 208.846 Reported internal fragmentation: 4.42297% Bits/key stored: 10.4948 Average FP rate %: 0.811006 (high keys/filter) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.7017 Number of filters: 164 Total size (MB): 200.352 Reported total allocated memory (MB): 221.5 Reported internal fragmentation: 10.5552% Bits/key stored: 10.0003 Average FP rate %: 0.969358 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 30.7131 Number of filters: 160 Total size (MB): 200.928 Reported total allocated memory (MB): 200.938 Reported internal fragmentation: 0.00448054% Bits/key stored: 10.1852 Average FP rate %: 0.963387 And from db_bench (block cache) with jemalloc: $ ./db_bench -db=/dev/shm/dbbench.no_optimize -benchmarks=fillrandom -format_version=5 -value_size=90 -bloom_bits=10 -num=2000000 -threads=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false $ ./db_bench -db=/dev/shm/dbbench -benchmarks=fillrandom -format_version=5 -value_size=90 -bloom_bits=10 -num=2000000 -threads=8 -optimize_filters_for_memory -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false $ (for FILE in /dev/shm/dbbench.no_optimize/*.sst; do ./sst_dump --file=$FILE --show_properties | grep 'filter block' ; done) | awk '{ t += $4; } END { print t; }' 17063835 $ (for FILE in /dev/shm/dbbench/*.sst; do ./sst_dump --file=$FILE --show_properties | grep 'filter block' ; done) | awk '{ t += $4; } END { print t; }' 17430747 $ #^ 2.1% additional filter storage $ ./db_bench -db=/dev/shm/dbbench.no_optimize -use_existing_db -benchmarks=readrandom,stats -statistics -bloom_bits=10 -num=2000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false -duration=10 -cache_index_and_filter_blocks -cache_size=1000000000 rocksdb.block.cache.index.add COUNT : 33 rocksdb.block.cache.index.bytes.insert COUNT : 8440400 rocksdb.block.cache.filter.add COUNT : 33 rocksdb.block.cache.filter.bytes.insert COUNT : 21087528 rocksdb.bloom.filter.useful COUNT : 4963889 rocksdb.bloom.filter.full.positive COUNT : 1214081 rocksdb.bloom.filter.full.true.positive COUNT : 1161999 $ #^ 1.04 % observed FP rate $ ./db_bench -db=/dev/shm/dbbench -use_existing_db -benchmarks=readrandom,stats -statistics -bloom_bits=10 -num=2000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false -optimize_filters_for_memory -duration=10 -cache_index_and_filter_blocks -cache_size=1000000000 rocksdb.block.cache.index.add COUNT : 33 rocksdb.block.cache.index.bytes.insert COUNT : 8448592 rocksdb.block.cache.filter.add COUNT : 33 rocksdb.block.cache.filter.bytes.insert COUNT : 18220328 rocksdb.bloom.filter.useful COUNT : 5360933 rocksdb.bloom.filter.full.positive COUNT : 1321315 rocksdb.bloom.filter.full.true.positive COUNT : 1262999 $ #^ 1.08 % observed FP rate, 13.6% less memory usage for filters (Due to specific key density, this example tends to generate filters that are "worse than average" for internal fragmentation. "Better than average" cases can show little or no improvement.) Pull Request resolved: https://github.com/facebook/rocksdb/pull/6427 Test Plan: unit test added, 'make check' with gcc, clang and valgrind Reviewed By: siying Differential Revision: D22124374 Pulled By: pdillinger fbshipit-source-id: f3e3aa152f9043ddf4fae25799e76341d0d8714e
5 years ago
size_t total_size = 0;
size_t total_keys_added = 0;
Warn on excessive keys for legacy Bloom filter with 32-bit hash (#6317) Summary: With many millions of keys, the old Bloom filter implementation for the block-based table (format_version <= 4) would have excessive FP rate due to the limitations of feeding the Bloom filter with a 32-bit hash. This change computes an estimated inflated FP rate due to this effect and warns in the log whenever an SST filter is constructed (almost certainly a "full" not "partitioned" filter) that exceeds 1.5x FP rate due to this effect. The detailed condition is only checked if 3 million keys or more have been added to a filter, as this should be a lower bound for common bits/key settings (< 20). Recommended remedies include smaller SST file size, using format_version >= 5 (for new Bloom filter), or using partitioned filters. This does not change behavior other than generating warnings for some constructed filters using the old implementation. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6317 Test Plan: Example with warning, 15M keys @ 15 bits / key: (working_mem_size_mb is just to stop after building one filter if it's large) $ ./filter_bench -quick -impl=0 -working_mem_size_mb=1 -bits_per_key=15 -average_keys_per_filter=15000000 2>&1 | grep 'FP rate' [WARN] [/block_based/filter_policy.cc:292] Using legacy SST/BBT Bloom filter with excessive key count (15.0M @ 15bpk), causing estimated 1.8x higher filter FP rate. Consider using new Bloom with format_version>=5, smaller SST file size, or partitioned filters. Predicted FP rate %: 0.766702 Average FP rate %: 0.66846 Example without warning (150K keys): $ ./filter_bench -quick -impl=0 -working_mem_size_mb=1 -bits_per_key=15 -average_keys_per_filter=150000 2>&1 | grep 'FP rate' Predicted FP rate %: 0.422857 Average FP rate %: 0.379301 $ With more samples at 15 bits/key: 150K keys -> no warning; actual: 0.379% FP rate (baseline) 1M keys -> no warning; actual: 0.396% FP rate, 1.045x 9M keys -> no warning; actual: 0.563% FP rate, 1.485x 10M keys -> warning (1.5x); actual: 0.564% FP rate, 1.488x 15M keys -> warning (1.8x); actual: 0.668% FP rate, 1.76x 25M keys -> warning (2.4x); actual: 0.880% FP rate, 2.32x At 10 bits/key: 150K keys -> no warning; actual: 1.17% FP rate (baseline) 1M keys -> no warning; actual: 1.16% FP rate 10M keys -> no warning; actual: 1.32% FP rate, 1.13x 25M keys -> no warning; actual: 1.63% FP rate, 1.39x 35M keys -> warning (1.6x); actual: 1.81% FP rate, 1.55x At 5 bits/key: 150K keys -> no warning; actual: 9.32% FP rate (baseline) 25M keys -> no warning; actual: 9.62% FP rate, 1.03x 200M keys -> no warning; actual: 12.2% FP rate, 1.31x 250M keys -> warning (1.5x); actual: 12.8% FP rate, 1.37x 300M keys -> warning (1.6x); actual: 13.4% FP rate, 1.43x The reason for the modest inaccuracy at low bits/key is that the assumption of independence between a collision between 32-hash values feeding the filter and an FP in the filter is not quite true for implementations using "simple" logic to compute indices from the stock hash result. There's math on this in my dissertation, but I don't think it's worth the effort just for these extreme cases (> 100 million keys and low-ish bits/key). Differential Revision: D19471715 Pulled By: pdillinger fbshipit-source-id: f80c96893a09bf1152630ff0b964e5cdd7e35c68
5 years ago
#ifdef PREDICT_FP_RATE
double weighted_predicted_fp_rate = 0.0;
#endif
size_t max_total_keys;
size_t max_mem;
if (FLAGS_m_keys_total_max > 0) {
max_total_keys = static_cast<size_t>(1000000 * FLAGS_m_keys_total_max);
max_mem = SIZE_MAX;
} else {
max_total_keys = SIZE_MAX;
max_mem = static_cast<size_t>(1024 * 1024 * working_mem_size_mb);
}
ROCKSDB_NAMESPACE::StopWatchNano timer(
ROCKSDB_NAMESPACE::SystemClock::Default().get(), true);
infos_.clear();
Minimize memory internal fragmentation for Bloom filters (#6427) Summary: New experimental option BBTO::optimize_filters_for_memory builds filters that maximize their use of "usable size" from malloc_usable_size, which is also used to compute block cache charges. Rather than always "rounding up," we track state in the BloomFilterPolicy object to mix essentially "rounding down" and "rounding up" so that the average FP rate of all generated filters is the same as without the option. (YMMV as heavily accessed filters might be unluckily lower accuracy.) Thus, the option near-minimizes what the block cache considers as "memory used" for a given target Bloom filter false positive rate and Bloom filter implementation. There are no forward or backward compatibility issues with this change, though it only works on the format_version=5 Bloom filter. With Jemalloc, we see about 10% reduction in memory footprint (and block cache charge) for Bloom filters, but 1-2% increase in storage footprint, due to encoding efficiency losses (FP rate is non-linear with bits/key). Why not weighted random round up/down rather than state tracking? By only requiring malloc_usable_size, we don't actually know what the next larger and next smaller usable sizes for the allocator are. We pick a requested size, accept and use whatever usable size it has, and use the difference to inform our next choice. This allows us to narrow in on the right balance without tracking/predicting usable sizes. Why not weight history of generated filter false positive rates by number of keys? This could lead to excess skew in small filters after generating a large filter. Results from filter_bench with jemalloc (irrelevant details omitted): (normal keys/filter, but high variance) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=30000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.6278 Number of filters: 5516 Total size (MB): 200.046 Reported total allocated memory (MB): 220.597 Reported internal fragmentation: 10.2732% Bits/key stored: 10.0097 Average FP rate %: 0.965228 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=30000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 30.5104 Number of filters: 5464 Total size (MB): 200.015 Reported total allocated memory (MB): 200.322 Reported internal fragmentation: 0.153709% Bits/key stored: 10.1011 Average FP rate %: 0.966313 (very few keys / filter, optimization not as effective due to ~59 byte internal fragmentation in blocked Bloom filter representation) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.5649 Number of filters: 162950 Total size (MB): 200.001 Reported total allocated memory (MB): 224.624 Reported internal fragmentation: 12.3117% Bits/key stored: 10.2951 Average FP rate %: 0.821534 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 31.8057 Number of filters: 159849 Total size (MB): 200 Reported total allocated memory (MB): 208.846 Reported internal fragmentation: 4.42297% Bits/key stored: 10.4948 Average FP rate %: 0.811006 (high keys/filter) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.7017 Number of filters: 164 Total size (MB): 200.352 Reported total allocated memory (MB): 221.5 Reported internal fragmentation: 10.5552% Bits/key stored: 10.0003 Average FP rate %: 0.969358 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 30.7131 Number of filters: 160 Total size (MB): 200.928 Reported total allocated memory (MB): 200.938 Reported internal fragmentation: 0.00448054% Bits/key stored: 10.1852 Average FP rate %: 0.963387 And from db_bench (block cache) with jemalloc: $ ./db_bench -db=/dev/shm/dbbench.no_optimize -benchmarks=fillrandom -format_version=5 -value_size=90 -bloom_bits=10 -num=2000000 -threads=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false $ ./db_bench -db=/dev/shm/dbbench -benchmarks=fillrandom -format_version=5 -value_size=90 -bloom_bits=10 -num=2000000 -threads=8 -optimize_filters_for_memory -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false $ (for FILE in /dev/shm/dbbench.no_optimize/*.sst; do ./sst_dump --file=$FILE --show_properties | grep 'filter block' ; done) | awk '{ t += $4; } END { print t; }' 17063835 $ (for FILE in /dev/shm/dbbench/*.sst; do ./sst_dump --file=$FILE --show_properties | grep 'filter block' ; done) | awk '{ t += $4; } END { print t; }' 17430747 $ #^ 2.1% additional filter storage $ ./db_bench -db=/dev/shm/dbbench.no_optimize -use_existing_db -benchmarks=readrandom,stats -statistics -bloom_bits=10 -num=2000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false -duration=10 -cache_index_and_filter_blocks -cache_size=1000000000 rocksdb.block.cache.index.add COUNT : 33 rocksdb.block.cache.index.bytes.insert COUNT : 8440400 rocksdb.block.cache.filter.add COUNT : 33 rocksdb.block.cache.filter.bytes.insert COUNT : 21087528 rocksdb.bloom.filter.useful COUNT : 4963889 rocksdb.bloom.filter.full.positive COUNT : 1214081 rocksdb.bloom.filter.full.true.positive COUNT : 1161999 $ #^ 1.04 % observed FP rate $ ./db_bench -db=/dev/shm/dbbench -use_existing_db -benchmarks=readrandom,stats -statistics -bloom_bits=10 -num=2000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false -optimize_filters_for_memory -duration=10 -cache_index_and_filter_blocks -cache_size=1000000000 rocksdb.block.cache.index.add COUNT : 33 rocksdb.block.cache.index.bytes.insert COUNT : 8448592 rocksdb.block.cache.filter.add COUNT : 33 rocksdb.block.cache.filter.bytes.insert COUNT : 18220328 rocksdb.bloom.filter.useful COUNT : 5360933 rocksdb.bloom.filter.full.positive COUNT : 1321315 rocksdb.bloom.filter.full.true.positive COUNT : 1262999 $ #^ 1.08 % observed FP rate, 13.6% less memory usage for filters (Due to specific key density, this example tends to generate filters that are "worse than average" for internal fragmentation. "Better than average" cases can show little or no improvement.) Pull Request resolved: https://github.com/facebook/rocksdb/pull/6427 Test Plan: unit test added, 'make check' with gcc, clang and valgrind Reviewed By: siying Differential Revision: D22124374 Pulled By: pdillinger fbshipit-source-id: f3e3aa152f9043ddf4fae25799e76341d0d8714e
5 years ago
while ((working_mem_size_mb == 0 || total_size < max_mem) &&
total_keys_added < max_total_keys) {
uint32_t filter_id = random_.Next();
uint32_t keys_to_add = FLAGS_average_keys_per_filter +
FastRange32(random_.Next(), variance_range) -
variance_offset;
if (max_total_keys - total_keys_added < keys_to_add) {
keys_to_add = static_cast<uint32_t>(max_total_keys - total_keys_added);
}
infos_.emplace_back();
FilterInfo &info = infos_.back();
info.filter_id_ = filter_id;
info.keys_added_ = keys_to_add;
if (FLAGS_use_plain_table_bloom) {
info.plain_table_bloom_.reset(new PlainTableBloomV1());
info.plain_table_bloom_->SetTotalBits(
&arena_, static_cast<uint32_t>(keys_to_add * FLAGS_bits_per_key),
FLAGS_impl, 0 /*huge_page*/, nullptr /*logger*/);
for (uint32_t i = 0; i < keys_to_add; ++i) {
uint32_t hash = GetSliceHash(kms_[0].Get(filter_id, i));
info.plain_table_bloom_->AddHash(hash);
}
info.filter_ = info.plain_table_bloom_->GetRawData();
} else {
Optimize memory and CPU for building new Bloom filter (#6175) 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
5 years ago
if (!builder) {
builder.reset(
static_cast_with_check<BuiltinFilterBitsBuilder>(GetBuilder()));
Optimize memory and CPU for building new Bloom filter (#6175) 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
5 years ago
}
for (uint32_t i = 0; i < keys_to_add; ++i) {
builder->AddKey(kms_[0].Get(filter_id, i));
}
info.filter_ = builder->Finish(&info.owner_);
Warn on excessive keys for legacy Bloom filter with 32-bit hash (#6317) Summary: With many millions of keys, the old Bloom filter implementation for the block-based table (format_version <= 4) would have excessive FP rate due to the limitations of feeding the Bloom filter with a 32-bit hash. This change computes an estimated inflated FP rate due to this effect and warns in the log whenever an SST filter is constructed (almost certainly a "full" not "partitioned" filter) that exceeds 1.5x FP rate due to this effect. The detailed condition is only checked if 3 million keys or more have been added to a filter, as this should be a lower bound for common bits/key settings (< 20). Recommended remedies include smaller SST file size, using format_version >= 5 (for new Bloom filter), or using partitioned filters. This does not change behavior other than generating warnings for some constructed filters using the old implementation. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6317 Test Plan: Example with warning, 15M keys @ 15 bits / key: (working_mem_size_mb is just to stop after building one filter if it's large) $ ./filter_bench -quick -impl=0 -working_mem_size_mb=1 -bits_per_key=15 -average_keys_per_filter=15000000 2>&1 | grep 'FP rate' [WARN] [/block_based/filter_policy.cc:292] Using legacy SST/BBT Bloom filter with excessive key count (15.0M @ 15bpk), causing estimated 1.8x higher filter FP rate. Consider using new Bloom with format_version>=5, smaller SST file size, or partitioned filters. Predicted FP rate %: 0.766702 Average FP rate %: 0.66846 Example without warning (150K keys): $ ./filter_bench -quick -impl=0 -working_mem_size_mb=1 -bits_per_key=15 -average_keys_per_filter=150000 2>&1 | grep 'FP rate' Predicted FP rate %: 0.422857 Average FP rate %: 0.379301 $ With more samples at 15 bits/key: 150K keys -> no warning; actual: 0.379% FP rate (baseline) 1M keys -> no warning; actual: 0.396% FP rate, 1.045x 9M keys -> no warning; actual: 0.563% FP rate, 1.485x 10M keys -> warning (1.5x); actual: 0.564% FP rate, 1.488x 15M keys -> warning (1.8x); actual: 0.668% FP rate, 1.76x 25M keys -> warning (2.4x); actual: 0.880% FP rate, 2.32x At 10 bits/key: 150K keys -> no warning; actual: 1.17% FP rate (baseline) 1M keys -> no warning; actual: 1.16% FP rate 10M keys -> no warning; actual: 1.32% FP rate, 1.13x 25M keys -> no warning; actual: 1.63% FP rate, 1.39x 35M keys -> warning (1.6x); actual: 1.81% FP rate, 1.55x At 5 bits/key: 150K keys -> no warning; actual: 9.32% FP rate (baseline) 25M keys -> no warning; actual: 9.62% FP rate, 1.03x 200M keys -> no warning; actual: 12.2% FP rate, 1.31x 250M keys -> warning (1.5x); actual: 12.8% FP rate, 1.37x 300M keys -> warning (1.6x); actual: 13.4% FP rate, 1.43x The reason for the modest inaccuracy at low bits/key is that the assumption of independence between a collision between 32-hash values feeding the filter and an FP in the filter is not quite true for implementations using "simple" logic to compute indices from the stock hash result. There's math on this in my dissertation, but I don't think it's worth the effort just for these extreme cases (> 100 million keys and low-ish bits/key). Differential Revision: D19471715 Pulled By: pdillinger fbshipit-source-id: f80c96893a09bf1152630ff0b964e5cdd7e35c68
5 years ago
#ifdef PREDICT_FP_RATE
weighted_predicted_fp_rate +=
keys_to_add *
builder->EstimatedFpRate(keys_to_add, info.filter_.size());
#endif
Optimize memory and CPU for building new Bloom filter (#6175) 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
5 years ago
if (FLAGS_new_builder) {
builder.reset();
}
info.reader_.reset(
table_options_.filter_policy->GetFilterBitsReader(info.filter_));
CachableEntry<ParsedFullFilterBlock> block(
new ParsedFullFilterBlock(table_options_.filter_policy.get(),
BlockContents(info.filter_)),
nullptr /* cache */, nullptr /* cache_handle */,
true /* own_value */);
info.full_block_reader_.reset(
new FullFilterBlockReader(table_.get(), std::move(block)));
}
Minimize memory internal fragmentation for Bloom filters (#6427) Summary: New experimental option BBTO::optimize_filters_for_memory builds filters that maximize their use of "usable size" from malloc_usable_size, which is also used to compute block cache charges. Rather than always "rounding up," we track state in the BloomFilterPolicy object to mix essentially "rounding down" and "rounding up" so that the average FP rate of all generated filters is the same as without the option. (YMMV as heavily accessed filters might be unluckily lower accuracy.) Thus, the option near-minimizes what the block cache considers as "memory used" for a given target Bloom filter false positive rate and Bloom filter implementation. There are no forward or backward compatibility issues with this change, though it only works on the format_version=5 Bloom filter. With Jemalloc, we see about 10% reduction in memory footprint (and block cache charge) for Bloom filters, but 1-2% increase in storage footprint, due to encoding efficiency losses (FP rate is non-linear with bits/key). Why not weighted random round up/down rather than state tracking? By only requiring malloc_usable_size, we don't actually know what the next larger and next smaller usable sizes for the allocator are. We pick a requested size, accept and use whatever usable size it has, and use the difference to inform our next choice. This allows us to narrow in on the right balance without tracking/predicting usable sizes. Why not weight history of generated filter false positive rates by number of keys? This could lead to excess skew in small filters after generating a large filter. Results from filter_bench with jemalloc (irrelevant details omitted): (normal keys/filter, but high variance) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=30000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.6278 Number of filters: 5516 Total size (MB): 200.046 Reported total allocated memory (MB): 220.597 Reported internal fragmentation: 10.2732% Bits/key stored: 10.0097 Average FP rate %: 0.965228 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=30000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 30.5104 Number of filters: 5464 Total size (MB): 200.015 Reported total allocated memory (MB): 200.322 Reported internal fragmentation: 0.153709% Bits/key stored: 10.1011 Average FP rate %: 0.966313 (very few keys / filter, optimization not as effective due to ~59 byte internal fragmentation in blocked Bloom filter representation) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.5649 Number of filters: 162950 Total size (MB): 200.001 Reported total allocated memory (MB): 224.624 Reported internal fragmentation: 12.3117% Bits/key stored: 10.2951 Average FP rate %: 0.821534 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 31.8057 Number of filters: 159849 Total size (MB): 200 Reported total allocated memory (MB): 208.846 Reported internal fragmentation: 4.42297% Bits/key stored: 10.4948 Average FP rate %: 0.811006 (high keys/filter) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.7017 Number of filters: 164 Total size (MB): 200.352 Reported total allocated memory (MB): 221.5 Reported internal fragmentation: 10.5552% Bits/key stored: 10.0003 Average FP rate %: 0.969358 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 30.7131 Number of filters: 160 Total size (MB): 200.928 Reported total allocated memory (MB): 200.938 Reported internal fragmentation: 0.00448054% Bits/key stored: 10.1852 Average FP rate %: 0.963387 And from db_bench (block cache) with jemalloc: $ ./db_bench -db=/dev/shm/dbbench.no_optimize -benchmarks=fillrandom -format_version=5 -value_size=90 -bloom_bits=10 -num=2000000 -threads=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false $ ./db_bench -db=/dev/shm/dbbench -benchmarks=fillrandom -format_version=5 -value_size=90 -bloom_bits=10 -num=2000000 -threads=8 -optimize_filters_for_memory -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false $ (for FILE in /dev/shm/dbbench.no_optimize/*.sst; do ./sst_dump --file=$FILE --show_properties | grep 'filter block' ; done) | awk '{ t += $4; } END { print t; }' 17063835 $ (for FILE in /dev/shm/dbbench/*.sst; do ./sst_dump --file=$FILE --show_properties | grep 'filter block' ; done) | awk '{ t += $4; } END { print t; }' 17430747 $ #^ 2.1% additional filter storage $ ./db_bench -db=/dev/shm/dbbench.no_optimize -use_existing_db -benchmarks=readrandom,stats -statistics -bloom_bits=10 -num=2000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false -duration=10 -cache_index_and_filter_blocks -cache_size=1000000000 rocksdb.block.cache.index.add COUNT : 33 rocksdb.block.cache.index.bytes.insert COUNT : 8440400 rocksdb.block.cache.filter.add COUNT : 33 rocksdb.block.cache.filter.bytes.insert COUNT : 21087528 rocksdb.bloom.filter.useful COUNT : 4963889 rocksdb.bloom.filter.full.positive COUNT : 1214081 rocksdb.bloom.filter.full.true.positive COUNT : 1161999 $ #^ 1.04 % observed FP rate $ ./db_bench -db=/dev/shm/dbbench -use_existing_db -benchmarks=readrandom,stats -statistics -bloom_bits=10 -num=2000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false -optimize_filters_for_memory -duration=10 -cache_index_and_filter_blocks -cache_size=1000000000 rocksdb.block.cache.index.add COUNT : 33 rocksdb.block.cache.index.bytes.insert COUNT : 8448592 rocksdb.block.cache.filter.add COUNT : 33 rocksdb.block.cache.filter.bytes.insert COUNT : 18220328 rocksdb.bloom.filter.useful COUNT : 5360933 rocksdb.bloom.filter.full.positive COUNT : 1321315 rocksdb.bloom.filter.full.true.positive COUNT : 1262999 $ #^ 1.08 % observed FP rate, 13.6% less memory usage for filters (Due to specific key density, this example tends to generate filters that are "worse than average" for internal fragmentation. "Better than average" cases can show little or no improvement.) Pull Request resolved: https://github.com/facebook/rocksdb/pull/6427 Test Plan: unit test added, 'make check' with gcc, clang and valgrind Reviewed By: siying Differential Revision: D22124374 Pulled By: pdillinger fbshipit-source-id: f3e3aa152f9043ddf4fae25799e76341d0d8714e
5 years ago
total_size += info.filter_.size();
#ifdef ROCKSDB_MALLOC_USABLE_SIZE
total_memory_used +=
malloc_usable_size(const_cast<char *>(info.filter_.data()));
#endif // ROCKSDB_MALLOC_USABLE_SIZE
total_keys_added += keys_to_add;
}
uint64_t elapsed_nanos = timer.ElapsedNanos();
double ns = double(elapsed_nanos) / total_keys_added;
std::cout << "Build avg ns/key: " << ns << std::endl;
std::cout << "Number of filters: " << infos_.size() << std::endl;
Minimize memory internal fragmentation for Bloom filters (#6427) Summary: New experimental option BBTO::optimize_filters_for_memory builds filters that maximize their use of "usable size" from malloc_usable_size, which is also used to compute block cache charges. Rather than always "rounding up," we track state in the BloomFilterPolicy object to mix essentially "rounding down" and "rounding up" so that the average FP rate of all generated filters is the same as without the option. (YMMV as heavily accessed filters might be unluckily lower accuracy.) Thus, the option near-minimizes what the block cache considers as "memory used" for a given target Bloom filter false positive rate and Bloom filter implementation. There are no forward or backward compatibility issues with this change, though it only works on the format_version=5 Bloom filter. With Jemalloc, we see about 10% reduction in memory footprint (and block cache charge) for Bloom filters, but 1-2% increase in storage footprint, due to encoding efficiency losses (FP rate is non-linear with bits/key). Why not weighted random round up/down rather than state tracking? By only requiring malloc_usable_size, we don't actually know what the next larger and next smaller usable sizes for the allocator are. We pick a requested size, accept and use whatever usable size it has, and use the difference to inform our next choice. This allows us to narrow in on the right balance without tracking/predicting usable sizes. Why not weight history of generated filter false positive rates by number of keys? This could lead to excess skew in small filters after generating a large filter. Results from filter_bench with jemalloc (irrelevant details omitted): (normal keys/filter, but high variance) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=30000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.6278 Number of filters: 5516 Total size (MB): 200.046 Reported total allocated memory (MB): 220.597 Reported internal fragmentation: 10.2732% Bits/key stored: 10.0097 Average FP rate %: 0.965228 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=30000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 30.5104 Number of filters: 5464 Total size (MB): 200.015 Reported total allocated memory (MB): 200.322 Reported internal fragmentation: 0.153709% Bits/key stored: 10.1011 Average FP rate %: 0.966313 (very few keys / filter, optimization not as effective due to ~59 byte internal fragmentation in blocked Bloom filter representation) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.5649 Number of filters: 162950 Total size (MB): 200.001 Reported total allocated memory (MB): 224.624 Reported internal fragmentation: 12.3117% Bits/key stored: 10.2951 Average FP rate %: 0.821534 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 31.8057 Number of filters: 159849 Total size (MB): 200 Reported total allocated memory (MB): 208.846 Reported internal fragmentation: 4.42297% Bits/key stored: 10.4948 Average FP rate %: 0.811006 (high keys/filter) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.7017 Number of filters: 164 Total size (MB): 200.352 Reported total allocated memory (MB): 221.5 Reported internal fragmentation: 10.5552% Bits/key stored: 10.0003 Average FP rate %: 0.969358 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 30.7131 Number of filters: 160 Total size (MB): 200.928 Reported total allocated memory (MB): 200.938 Reported internal fragmentation: 0.00448054% Bits/key stored: 10.1852 Average FP rate %: 0.963387 And from db_bench (block cache) with jemalloc: $ ./db_bench -db=/dev/shm/dbbench.no_optimize -benchmarks=fillrandom -format_version=5 -value_size=90 -bloom_bits=10 -num=2000000 -threads=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false $ ./db_bench -db=/dev/shm/dbbench -benchmarks=fillrandom -format_version=5 -value_size=90 -bloom_bits=10 -num=2000000 -threads=8 -optimize_filters_for_memory -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false $ (for FILE in /dev/shm/dbbench.no_optimize/*.sst; do ./sst_dump --file=$FILE --show_properties | grep 'filter block' ; done) | awk '{ t += $4; } END { print t; }' 17063835 $ (for FILE in /dev/shm/dbbench/*.sst; do ./sst_dump --file=$FILE --show_properties | grep 'filter block' ; done) | awk '{ t += $4; } END { print t; }' 17430747 $ #^ 2.1% additional filter storage $ ./db_bench -db=/dev/shm/dbbench.no_optimize -use_existing_db -benchmarks=readrandom,stats -statistics -bloom_bits=10 -num=2000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false -duration=10 -cache_index_and_filter_blocks -cache_size=1000000000 rocksdb.block.cache.index.add COUNT : 33 rocksdb.block.cache.index.bytes.insert COUNT : 8440400 rocksdb.block.cache.filter.add COUNT : 33 rocksdb.block.cache.filter.bytes.insert COUNT : 21087528 rocksdb.bloom.filter.useful COUNT : 4963889 rocksdb.bloom.filter.full.positive COUNT : 1214081 rocksdb.bloom.filter.full.true.positive COUNT : 1161999 $ #^ 1.04 % observed FP rate $ ./db_bench -db=/dev/shm/dbbench -use_existing_db -benchmarks=readrandom,stats -statistics -bloom_bits=10 -num=2000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false -optimize_filters_for_memory -duration=10 -cache_index_and_filter_blocks -cache_size=1000000000 rocksdb.block.cache.index.add COUNT : 33 rocksdb.block.cache.index.bytes.insert COUNT : 8448592 rocksdb.block.cache.filter.add COUNT : 33 rocksdb.block.cache.filter.bytes.insert COUNT : 18220328 rocksdb.bloom.filter.useful COUNT : 5360933 rocksdb.bloom.filter.full.positive COUNT : 1321315 rocksdb.bloom.filter.full.true.positive COUNT : 1262999 $ #^ 1.08 % observed FP rate, 13.6% less memory usage for filters (Due to specific key density, this example tends to generate filters that are "worse than average" for internal fragmentation. "Better than average" cases can show little or no improvement.) Pull Request resolved: https://github.com/facebook/rocksdb/pull/6427 Test Plan: unit test added, 'make check' with gcc, clang and valgrind Reviewed By: siying Differential Revision: D22124374 Pulled By: pdillinger fbshipit-source-id: f3e3aa152f9043ddf4fae25799e76341d0d8714e
5 years ago
std::cout << "Total size (MB): " << total_size / 1024.0 / 1024.0 << std::endl;
if (total_memory_used > 0) {
std::cout << "Reported total allocated memory (MB): "
<< total_memory_used / 1024.0 / 1024.0 << std::endl;
std::cout << "Reported internal fragmentation: "
<< (total_memory_used - total_size) * 100.0 / total_size << "%"
<< std::endl;
}
Minimize memory internal fragmentation for Bloom filters (#6427) Summary: New experimental option BBTO::optimize_filters_for_memory builds filters that maximize their use of "usable size" from malloc_usable_size, which is also used to compute block cache charges. Rather than always "rounding up," we track state in the BloomFilterPolicy object to mix essentially "rounding down" and "rounding up" so that the average FP rate of all generated filters is the same as without the option. (YMMV as heavily accessed filters might be unluckily lower accuracy.) Thus, the option near-minimizes what the block cache considers as "memory used" for a given target Bloom filter false positive rate and Bloom filter implementation. There are no forward or backward compatibility issues with this change, though it only works on the format_version=5 Bloom filter. With Jemalloc, we see about 10% reduction in memory footprint (and block cache charge) for Bloom filters, but 1-2% increase in storage footprint, due to encoding efficiency losses (FP rate is non-linear with bits/key). Why not weighted random round up/down rather than state tracking? By only requiring malloc_usable_size, we don't actually know what the next larger and next smaller usable sizes for the allocator are. We pick a requested size, accept and use whatever usable size it has, and use the difference to inform our next choice. This allows us to narrow in on the right balance without tracking/predicting usable sizes. Why not weight history of generated filter false positive rates by number of keys? This could lead to excess skew in small filters after generating a large filter. Results from filter_bench with jemalloc (irrelevant details omitted): (normal keys/filter, but high variance) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=30000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.6278 Number of filters: 5516 Total size (MB): 200.046 Reported total allocated memory (MB): 220.597 Reported internal fragmentation: 10.2732% Bits/key stored: 10.0097 Average FP rate %: 0.965228 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=30000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 30.5104 Number of filters: 5464 Total size (MB): 200.015 Reported total allocated memory (MB): 200.322 Reported internal fragmentation: 0.153709% Bits/key stored: 10.1011 Average FP rate %: 0.966313 (very few keys / filter, optimization not as effective due to ~59 byte internal fragmentation in blocked Bloom filter representation) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.5649 Number of filters: 162950 Total size (MB): 200.001 Reported total allocated memory (MB): 224.624 Reported internal fragmentation: 12.3117% Bits/key stored: 10.2951 Average FP rate %: 0.821534 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 31.8057 Number of filters: 159849 Total size (MB): 200 Reported total allocated memory (MB): 208.846 Reported internal fragmentation: 4.42297% Bits/key stored: 10.4948 Average FP rate %: 0.811006 (high keys/filter) $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000000 -vary_key_count_ratio=0.9 Build avg ns/key: 29.7017 Number of filters: 164 Total size (MB): 200.352 Reported total allocated memory (MB): 221.5 Reported internal fragmentation: 10.5552% Bits/key stored: 10.0003 Average FP rate %: 0.969358 $ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory Build avg ns/key: 30.7131 Number of filters: 160 Total size (MB): 200.928 Reported total allocated memory (MB): 200.938 Reported internal fragmentation: 0.00448054% Bits/key stored: 10.1852 Average FP rate %: 0.963387 And from db_bench (block cache) with jemalloc: $ ./db_bench -db=/dev/shm/dbbench.no_optimize -benchmarks=fillrandom -format_version=5 -value_size=90 -bloom_bits=10 -num=2000000 -threads=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false $ ./db_bench -db=/dev/shm/dbbench -benchmarks=fillrandom -format_version=5 -value_size=90 -bloom_bits=10 -num=2000000 -threads=8 -optimize_filters_for_memory -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false $ (for FILE in /dev/shm/dbbench.no_optimize/*.sst; do ./sst_dump --file=$FILE --show_properties | grep 'filter block' ; done) | awk '{ t += $4; } END { print t; }' 17063835 $ (for FILE in /dev/shm/dbbench/*.sst; do ./sst_dump --file=$FILE --show_properties | grep 'filter block' ; done) | awk '{ t += $4; } END { print t; }' 17430747 $ #^ 2.1% additional filter storage $ ./db_bench -db=/dev/shm/dbbench.no_optimize -use_existing_db -benchmarks=readrandom,stats -statistics -bloom_bits=10 -num=2000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false -duration=10 -cache_index_and_filter_blocks -cache_size=1000000000 rocksdb.block.cache.index.add COUNT : 33 rocksdb.block.cache.index.bytes.insert COUNT : 8440400 rocksdb.block.cache.filter.add COUNT : 33 rocksdb.block.cache.filter.bytes.insert COUNT : 21087528 rocksdb.bloom.filter.useful COUNT : 4963889 rocksdb.bloom.filter.full.positive COUNT : 1214081 rocksdb.bloom.filter.full.true.positive COUNT : 1161999 $ #^ 1.04 % observed FP rate $ ./db_bench -db=/dev/shm/dbbench -use_existing_db -benchmarks=readrandom,stats -statistics -bloom_bits=10 -num=2000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false -optimize_filters_for_memory -duration=10 -cache_index_and_filter_blocks -cache_size=1000000000 rocksdb.block.cache.index.add COUNT : 33 rocksdb.block.cache.index.bytes.insert COUNT : 8448592 rocksdb.block.cache.filter.add COUNT : 33 rocksdb.block.cache.filter.bytes.insert COUNT : 18220328 rocksdb.bloom.filter.useful COUNT : 5360933 rocksdb.bloom.filter.full.positive COUNT : 1321315 rocksdb.bloom.filter.full.true.positive COUNT : 1262999 $ #^ 1.08 % observed FP rate, 13.6% less memory usage for filters (Due to specific key density, this example tends to generate filters that are "worse than average" for internal fragmentation. "Better than average" cases can show little or no improvement.) Pull Request resolved: https://github.com/facebook/rocksdb/pull/6427 Test Plan: unit test added, 'make check' with gcc, clang and valgrind Reviewed By: siying Differential Revision: D22124374 Pulled By: pdillinger fbshipit-source-id: f3e3aa152f9043ddf4fae25799e76341d0d8714e
5 years ago
double bpk = total_size * 8.0 / total_keys_added;
std::cout << "Bits/key stored: " << bpk << std::endl;
Warn on excessive keys for legacy Bloom filter with 32-bit hash (#6317) Summary: With many millions of keys, the old Bloom filter implementation for the block-based table (format_version <= 4) would have excessive FP rate due to the limitations of feeding the Bloom filter with a 32-bit hash. This change computes an estimated inflated FP rate due to this effect and warns in the log whenever an SST filter is constructed (almost certainly a "full" not "partitioned" filter) that exceeds 1.5x FP rate due to this effect. The detailed condition is only checked if 3 million keys or more have been added to a filter, as this should be a lower bound for common bits/key settings (< 20). Recommended remedies include smaller SST file size, using format_version >= 5 (for new Bloom filter), or using partitioned filters. This does not change behavior other than generating warnings for some constructed filters using the old implementation. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6317 Test Plan: Example with warning, 15M keys @ 15 bits / key: (working_mem_size_mb is just to stop after building one filter if it's large) $ ./filter_bench -quick -impl=0 -working_mem_size_mb=1 -bits_per_key=15 -average_keys_per_filter=15000000 2>&1 | grep 'FP rate' [WARN] [/block_based/filter_policy.cc:292] Using legacy SST/BBT Bloom filter with excessive key count (15.0M @ 15bpk), causing estimated 1.8x higher filter FP rate. Consider using new Bloom with format_version>=5, smaller SST file size, or partitioned filters. Predicted FP rate %: 0.766702 Average FP rate %: 0.66846 Example without warning (150K keys): $ ./filter_bench -quick -impl=0 -working_mem_size_mb=1 -bits_per_key=15 -average_keys_per_filter=150000 2>&1 | grep 'FP rate' Predicted FP rate %: 0.422857 Average FP rate %: 0.379301 $ With more samples at 15 bits/key: 150K keys -> no warning; actual: 0.379% FP rate (baseline) 1M keys -> no warning; actual: 0.396% FP rate, 1.045x 9M keys -> no warning; actual: 0.563% FP rate, 1.485x 10M keys -> warning (1.5x); actual: 0.564% FP rate, 1.488x 15M keys -> warning (1.8x); actual: 0.668% FP rate, 1.76x 25M keys -> warning (2.4x); actual: 0.880% FP rate, 2.32x At 10 bits/key: 150K keys -> no warning; actual: 1.17% FP rate (baseline) 1M keys -> no warning; actual: 1.16% FP rate 10M keys -> no warning; actual: 1.32% FP rate, 1.13x 25M keys -> no warning; actual: 1.63% FP rate, 1.39x 35M keys -> warning (1.6x); actual: 1.81% FP rate, 1.55x At 5 bits/key: 150K keys -> no warning; actual: 9.32% FP rate (baseline) 25M keys -> no warning; actual: 9.62% FP rate, 1.03x 200M keys -> no warning; actual: 12.2% FP rate, 1.31x 250M keys -> warning (1.5x); actual: 12.8% FP rate, 1.37x 300M keys -> warning (1.6x); actual: 13.4% FP rate, 1.43x The reason for the modest inaccuracy at low bits/key is that the assumption of independence between a collision between 32-hash values feeding the filter and an FP in the filter is not quite true for implementations using "simple" logic to compute indices from the stock hash result. There's math on this in my dissertation, but I don't think it's worth the effort just for these extreme cases (> 100 million keys and low-ish bits/key). Differential Revision: D19471715 Pulled By: pdillinger fbshipit-source-id: f80c96893a09bf1152630ff0b964e5cdd7e35c68
5 years ago
#ifdef PREDICT_FP_RATE
std::cout << "Predicted FP rate %: "
<< 100.0 * (weighted_predicted_fp_rate / total_keys_added)
<< std::endl;
#endif
if (!FLAGS_quick && !FLAGS_best_case) {
double tolerable_rate = std::pow(2.0, -(bpk - 1.0) / (1.4 + bpk / 50.0));
std::cout << "Best possible FP rate %: " << 100.0 * std::pow(2.0, -bpk)
<< std::endl;
std::cout << "Tolerable FP rate %: " << 100.0 * tolerable_rate << std::endl;
std::cout << "----------------------------" << std::endl;
std::cout << "Verifying..." << std::endl;
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
5 years ago
uint32_t outside_q_per_f =
static_cast<uint32_t>(m_queries_ * 1000000 / infos_.size());
uint64_t fps = 0;
for (uint32_t i = 0; i < infos_.size(); ++i) {
FilterInfo &info = infos_[i];
for (uint32_t j = 0; j < info.keys_added_; ++j) {
if (FLAGS_use_plain_table_bloom) {
uint32_t hash = GetSliceHash(kms_[0].Get(info.filter_id_, j));
ALWAYS_ASSERT(info.plain_table_bloom_->MayContainHash(hash));
} else {
ALWAYS_ASSERT(
info.reader_->MayMatch(kms_[0].Get(info.filter_id_, j)));
}
}
for (uint32_t j = 0; j < outside_q_per_f; ++j) {
if (FLAGS_use_plain_table_bloom) {
uint32_t hash =
GetSliceHash(kms_[0].Get(info.filter_id_, j | 0x80000000));
fps += info.plain_table_bloom_->MayContainHash(hash);
} else {
fps += info.reader_->MayMatch(
kms_[0].Get(info.filter_id_, j | 0x80000000));
}
}
}
std::cout << " No FNs :)" << std::endl;
double prelim_rate = double(fps) / outside_q_per_f / infos_.size();
std::cout << " Prelim FP rate %: " << (100.0 * prelim_rate) << std::endl;
if (!FLAGS_allow_bad_fp_rate) {
ALWAYS_ASSERT(prelim_rate < tolerable_rate);
}
}
std::cout << "----------------------------" << std::endl;
std::cout << "Mixed inside/outside queries..." << std::endl;
// 50% each inside and outside
uint32_t inside_threshold = UINT32_MAX / 2;
for (TestMode tm : testModes) {
random_.Seed(FLAGS_seed + 1);
double f = RandomQueryTest(inside_threshold, /*dry_run*/ false, tm);
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
random_.Seed(FLAGS_seed + 1);
double d = RandomQueryTest(inside_threshold, /*dry_run*/ true, tm);
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
std::cout << " " << TestModeToString(tm) << " net ns/op: " << (f - d)
<< std::endl;
}
if (!FLAGS_quick) {
std::cout << "----------------------------" << std::endl;
std::cout << "Inside queries (mostly)..." << std::endl;
// Do about 95% inside queries rather than 100% so that branch predictor
// can't give itself an artifically crazy advantage.
inside_threshold = UINT32_MAX / 20 * 19;
for (TestMode tm : testModes) {
random_.Seed(FLAGS_seed + 1);
double f = RandomQueryTest(inside_threshold, /*dry_run*/ false, tm);
random_.Seed(FLAGS_seed + 1);
double d = RandomQueryTest(inside_threshold, /*dry_run*/ true, tm);
std::cout << " " << TestModeToString(tm) << " net ns/op: " << (f - d)
<< std::endl;
}
std::cout << "----------------------------" << std::endl;
std::cout << "Outside queries (mostly)..." << std::endl;
// Do about 95% outside queries rather than 100% so that branch predictor
// can't give itself an artifically crazy advantage.
inside_threshold = UINT32_MAX / 20;
for (TestMode tm : testModes) {
random_.Seed(FLAGS_seed + 2);
double f = RandomQueryTest(inside_threshold, /*dry_run*/ false, tm);
random_.Seed(FLAGS_seed + 2);
double d = RandomQueryTest(inside_threshold, /*dry_run*/ true, tm);
std::cout << " " << TestModeToString(tm) << " net ns/op: " << (f - d)
<< std::endl;
}
}
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
std::cout << fp_rate_report_.str();
std::cout << "----------------------------" << std::endl;
std::cout << "Done. (For more info, run with -legend or -help.)" << std::endl;
}
double FilterBench::RandomQueryTest(uint32_t inside_threshold, bool dry_run,
TestMode mode) {
for (auto &info : infos_) {
info.outside_queries_ = 0;
info.false_positives_ = 0;
}
Add new persistent 64-bit hash (#5984) 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
5 years ago
auto dry_run_hash_fn = DryRunNoHash;
if (!FLAGS_net_includes_hashing) {
if (FLAGS_impl < 2 || FLAGS_use_plain_table_bloom) {
dry_run_hash_fn = DryRunHash32;
} else {
dry_run_hash_fn = DryRunHash64;
}
}
uint32_t num_infos = static_cast<uint32_t>(infos_.size());
uint32_t dry_run_hash = 0;
uint64_t max_queries = static_cast<uint64_t>(m_queries_ * 1000000 + 0.50);
// Some filters may be considered secondary in order to implement skewed
// queries. num_primary_filters is the number that are to be treated as
// equal, and any remainder will be treated as secondary.
uint32_t num_primary_filters = num_infos;
// The proportion (when divided by 2^32 - 1) of filter queries going to
// the primary filters (default = all). The remainder of queries are
// against secondary filters.
uint32_t primary_filter_threshold = 0xffffffff;
if (mode == kSingleFilter) {
// 100% of queries to 1 filter
num_primary_filters = 1;
} else if (mode == kFiftyOneFilter) {
Use size_t for filter APIs, protect against overflow (#7726) Summary: Deprecate CalculateNumEntry and replace with ApproximateNumEntries (better name) using size_t instead of int and uint32_t, to minimize confusing casts and bad overflow behavior (possible though probably not realistic). Bloom sizes are now explicitly capped at max size supported by implementations: just under 4GiB for fv=5 Bloom, and just under 512MiB for fv<5 Legacy Bloom. This hardening could help to set up for fuzzing. Also, since RocksDB only uses this information as an approximation for trying to hit certain sizes for partitioned filters, it's more important that the function be reasonably fast than for it to be completely accurate. It's hard enough to be 100% accurate for Ribbon (currently reversing CalculateSpace) that adding optimize_filters_for_memory into the mix is just not worth trying to be 100% accurate for num entries for bytes. Also: - Cleaned up filter_policy.h to remove MSVC warning handling and potentially unsafe use of exception for "not implemented" - Correct the number of entries limit beyond which current Ribbon implementation falls back on Bloom instead. - Consistently use "num_entries" rather than "num_entry" - Remove LegacyBloomBitsBuilder::CalculateNumEntry as it's essentially obsolete from general implementation BuiltinFilterBitsBuilder::CalculateNumEntries. - Fix filter_bench to skip some tests that don't make sense when only one or a small number of filters has been generated. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7726 Test Plan: expanded existing unit tests for CalculateSpace / ApproximateNumEntries. Also manually used filter_bench to verify Legacy and fv=5 Bloom size caps work (much too expensive for unit test). Note that the actual bits per key is below requested due to space cap. $ ./filter_bench -impl=0 -bits_per_key=20 -average_keys_per_filter=256000000 -vary_key_count_ratio=0 -m_keys_total_max=256 -allow_bad_fp_rate ... Total size (MB): 511.992 Bits/key stored: 16.777 ... $ ./filter_bench -impl=2 -bits_per_key=20 -average_keys_per_filter=2000000000 -vary_key_count_ratio=0 -m_keys_total_max=2000 ... Total size (MB): 4096 Bits/key stored: 17.1799 ... $ Reviewed By: jay-zhuang Differential Revision: D25239800 Pulled By: pdillinger fbshipit-source-id: f94e6d065efd31e05ec630ae1a82e6400d8390c4
4 years ago
if (num_infos < 50) {
return 0.0; // skip
}
// 50% of queries
primary_filter_threshold /= 2;
// to 1% of filters
num_primary_filters = (num_primary_filters + 99) / 100;
} else if (mode == kEightyTwentyFilter) {
Use size_t for filter APIs, protect against overflow (#7726) Summary: Deprecate CalculateNumEntry and replace with ApproximateNumEntries (better name) using size_t instead of int and uint32_t, to minimize confusing casts and bad overflow behavior (possible though probably not realistic). Bloom sizes are now explicitly capped at max size supported by implementations: just under 4GiB for fv=5 Bloom, and just under 512MiB for fv<5 Legacy Bloom. This hardening could help to set up for fuzzing. Also, since RocksDB only uses this information as an approximation for trying to hit certain sizes for partitioned filters, it's more important that the function be reasonably fast than for it to be completely accurate. It's hard enough to be 100% accurate for Ribbon (currently reversing CalculateSpace) that adding optimize_filters_for_memory into the mix is just not worth trying to be 100% accurate for num entries for bytes. Also: - Cleaned up filter_policy.h to remove MSVC warning handling and potentially unsafe use of exception for "not implemented" - Correct the number of entries limit beyond which current Ribbon implementation falls back on Bloom instead. - Consistently use "num_entries" rather than "num_entry" - Remove LegacyBloomBitsBuilder::CalculateNumEntry as it's essentially obsolete from general implementation BuiltinFilterBitsBuilder::CalculateNumEntries. - Fix filter_bench to skip some tests that don't make sense when only one or a small number of filters has been generated. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7726 Test Plan: expanded existing unit tests for CalculateSpace / ApproximateNumEntries. Also manually used filter_bench to verify Legacy and fv=5 Bloom size caps work (much too expensive for unit test). Note that the actual bits per key is below requested due to space cap. $ ./filter_bench -impl=0 -bits_per_key=20 -average_keys_per_filter=256000000 -vary_key_count_ratio=0 -m_keys_total_max=256 -allow_bad_fp_rate ... Total size (MB): 511.992 Bits/key stored: 16.777 ... $ ./filter_bench -impl=2 -bits_per_key=20 -average_keys_per_filter=2000000000 -vary_key_count_ratio=0 -m_keys_total_max=2000 ... Total size (MB): 4096 Bits/key stored: 17.1799 ... $ Reviewed By: jay-zhuang Differential Revision: D25239800 Pulled By: pdillinger fbshipit-source-id: f94e6d065efd31e05ec630ae1a82e6400d8390c4
4 years ago
if (num_infos < 5) {
return 0.0; // skip
}
// 80% of queries
primary_filter_threshold = primary_filter_threshold / 5 * 4;
// to 20% of filters
num_primary_filters = (num_primary_filters + 4) / 5;
Use size_t for filter APIs, protect against overflow (#7726) Summary: Deprecate CalculateNumEntry and replace with ApproximateNumEntries (better name) using size_t instead of int and uint32_t, to minimize confusing casts and bad overflow behavior (possible though probably not realistic). Bloom sizes are now explicitly capped at max size supported by implementations: just under 4GiB for fv=5 Bloom, and just under 512MiB for fv<5 Legacy Bloom. This hardening could help to set up for fuzzing. Also, since RocksDB only uses this information as an approximation for trying to hit certain sizes for partitioned filters, it's more important that the function be reasonably fast than for it to be completely accurate. It's hard enough to be 100% accurate for Ribbon (currently reversing CalculateSpace) that adding optimize_filters_for_memory into the mix is just not worth trying to be 100% accurate for num entries for bytes. Also: - Cleaned up filter_policy.h to remove MSVC warning handling and potentially unsafe use of exception for "not implemented" - Correct the number of entries limit beyond which current Ribbon implementation falls back on Bloom instead. - Consistently use "num_entries" rather than "num_entry" - Remove LegacyBloomBitsBuilder::CalculateNumEntry as it's essentially obsolete from general implementation BuiltinFilterBitsBuilder::CalculateNumEntries. - Fix filter_bench to skip some tests that don't make sense when only one or a small number of filters has been generated. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7726 Test Plan: expanded existing unit tests for CalculateSpace / ApproximateNumEntries. Also manually used filter_bench to verify Legacy and fv=5 Bloom size caps work (much too expensive for unit test). Note that the actual bits per key is below requested due to space cap. $ ./filter_bench -impl=0 -bits_per_key=20 -average_keys_per_filter=256000000 -vary_key_count_ratio=0 -m_keys_total_max=256 -allow_bad_fp_rate ... Total size (MB): 511.992 Bits/key stored: 16.777 ... $ ./filter_bench -impl=2 -bits_per_key=20 -average_keys_per_filter=2000000000 -vary_key_count_ratio=0 -m_keys_total_max=2000 ... Total size (MB): 4096 Bits/key stored: 17.1799 ... $ Reviewed By: jay-zhuang Differential Revision: D25239800 Pulled By: pdillinger fbshipit-source-id: f94e6d065efd31e05ec630ae1a82e6400d8390c4
4 years ago
} else if (mode == kRandomFilter) {
if (num_infos == 1) {
return 0.0; // skip
}
}
uint32_t batch_size = 1;
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
std::unique_ptr<Slice[]> batch_slices;
std::unique_ptr<Slice *[]> batch_slice_ptrs;
std::unique_ptr<bool[]> batch_results;
if (mode == kBatchPrepared || mode == kBatchUnprepared) {
batch_size = static_cast<uint32_t>(kms_.size());
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
}
batch_slices.reset(new Slice[batch_size]);
batch_slice_ptrs.reset(new Slice *[batch_size]);
batch_results.reset(new bool[batch_size]);
for (uint32_t i = 0; i < batch_size; ++i) {
batch_results[i] = false;
batch_slice_ptrs[i] = &batch_slices[i];
}
ROCKSDB_NAMESPACE::StopWatchNano timer(
ROCKSDB_NAMESPACE::SystemClock::Default().get(), true);
for (uint64_t q = 0; q < max_queries; q += batch_size) {
bool inside_this_time = random_.Next() <= inside_threshold;
uint32_t filter_index;
if (random_.Next() <= primary_filter_threshold) {
filter_index = random_.Uniformish(num_primary_filters);
} else {
// secondary
filter_index = num_primary_filters +
random_.Uniformish(num_infos - num_primary_filters);
}
FilterInfo &info = infos_[filter_index];
for (uint32_t i = 0; i < batch_size; ++i) {
if (inside_this_time) {
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
batch_slices[i] =
kms_[i].Get(info.filter_id_, random_.Uniformish(info.keys_added_));
} else {
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
batch_slices[i] =
kms_[i].Get(info.filter_id_, random_.Uniformish(info.keys_added_) |
uint32_t{0x80000000});
info.outside_queries_++;
}
}
// TODO: implement batched interface to full block reader
// TODO: implement batched interface to plain table bloom
if (mode == kBatchPrepared && !FLAGS_use_full_block_reader &&
!FLAGS_use_plain_table_bloom) {
for (uint32_t i = 0; i < batch_size; ++i) {
batch_results[i] = false;
}
if (dry_run) {
for (uint32_t i = 0; i < batch_size; ++i) {
batch_results[i] = true;
Add new persistent 64-bit hash (#5984) 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
5 years ago
dry_run_hash += dry_run_hash_fn(batch_slices[i]);
}
} else {
info.reader_->MayMatch(batch_size, batch_slice_ptrs.get(),
batch_results.get());
}
for (uint32_t i = 0; i < batch_size; ++i) {
if (inside_this_time) {
ALWAYS_ASSERT(batch_results[i]);
} else {
info.false_positives_ += batch_results[i];
}
}
} else {
for (uint32_t i = 0; i < batch_size; ++i) {
bool may_match;
if (FLAGS_use_plain_table_bloom) {
if (dry_run) {
Add new persistent 64-bit hash (#5984) 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
5 years ago
dry_run_hash += dry_run_hash_fn(batch_slices[i]);
may_match = true;
} else {
uint32_t hash = GetSliceHash(batch_slices[i]);
may_match = info.plain_table_bloom_->MayContainHash(hash);
}
} else if (FLAGS_use_full_block_reader) {
if (dry_run) {
Add new persistent 64-bit hash (#5984) 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
5 years ago
dry_run_hash += dry_run_hash_fn(batch_slices[i]);
may_match = true;
} else {
may_match = info.full_block_reader_->KeyMayMatch(
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
batch_slices[i],
/*prefix_extractor=*/nullptr,
/*block_offset=*/ROCKSDB_NAMESPACE::kNotValid,
/*no_io=*/false, /*const_ikey_ptr=*/nullptr,
/*get_context=*/nullptr,
/*lookup_context=*/nullptr);
}
} else {
if (dry_run) {
Add new persistent 64-bit hash (#5984) 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
5 years ago
dry_run_hash += dry_run_hash_fn(batch_slices[i]);
may_match = true;
} else {
may_match = info.reader_->MayMatch(batch_slices[i]);
}
}
if (inside_this_time) {
ALWAYS_ASSERT(may_match);
} else {
info.false_positives_ += may_match;
}
}
}
}
uint64_t elapsed_nanos = timer.ElapsedNanos();
double ns = double(elapsed_nanos) / max_queries;
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
if (!FLAGS_quick) {
if (dry_run) {
// Printing part of hash prevents dry run components from being optimized
// away by compiler
std::cout << " Dry run (" << std::hex << (dry_run_hash & 0xfffff)
<< std::dec << ") ";
} else {
std::cout << " Gross filter ";
}
std::cout << "ns/op: " << ns << std::endl;
}
if (!dry_run) {
fp_rate_report_.str("");
uint64_t q = 0;
uint64_t fp = 0;
double worst_fp_rate = 0.0;
double best_fp_rate = 1.0;
for (auto &info : infos_) {
q += info.outside_queries_;
fp += info.false_positives_;
if (info.outside_queries_ > 0) {
double fp_rate = double(info.false_positives_) / info.outside_queries_;
worst_fp_rate = std::max(worst_fp_rate, fp_rate);
best_fp_rate = std::min(best_fp_rate, fp_rate);
}
}
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
fp_rate_report_ << " Average FP rate %: " << 100.0 * fp / q << std::endl;
if (!FLAGS_quick && !FLAGS_best_case) {
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
fp_rate_report_ << " Worst FP rate %: " << 100.0 * worst_fp_rate
<< std::endl;
fp_rate_report_ << " Best FP rate %: " << 100.0 * best_fp_rate
<< std::endl;
fp_rate_report_ << " Best possible bits/key: "
<< -std::log(double(fp) / q) / std::log(2.0) << std::endl;
}
}
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
return ns;
}
int main(int argc, char **argv) {
ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
SetUsageMessage(std::string("\nUSAGE:\n") + std::string(argv[0]) +
" [-quick] [OTHER OPTIONS]...");
ParseCommandLineFlags(&argc, &argv, true);
PrintWarnings();
if (FLAGS_legend) {
std::cout
<< "Legend:" << std::endl
<< " \"Inside\" - key that was added to filter" << std::endl
<< " \"Outside\" - key that was not added to filter" << std::endl
<< " \"FN\" - false negative query (must not happen)" << std::endl
<< " \"FP\" - false positive query (OK at low rate)" << std::endl
Vary key size and alignment in filter_bench (#5933) 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
5 years ago
<< " \"Dry run\" - cost of testing and hashing overhead." << std::endl
<< " \"Gross filter\" - cost of filter queries including testing "
<< "\n and hashing overhead." << std::endl
<< " \"net\" - best estimate of time in filter operation, without "
<< "\n testing and hashing overhead (gross filter - dry run)"
<< std::endl
<< " \"ns/op\" - nanoseconds per operation (key query or add)"
<< std::endl
<< " \"Single filter\" - essentially minimum cost, assuming filter"
<< "\n fits easily in L1 CPU cache." << std::endl
<< " \"Batched, prepared\" - several queries at once against a"
<< "\n randomly chosen filter, using multi-query interface."
<< std::endl
<< " \"Batched, unprepared\" - similar, but using serial calls"
<< "\n to single query interface." << std::endl
<< " \"Random filter\" - a filter is chosen at random as target"
<< "\n of each query." << std::endl
<< " \"Skewed X% in Y%\" - like \"Random filter\" except Y% of"
<< "\n the filters are designated as \"hot\" and receive X%"
<< "\n of queries." << std::endl;
} else {
FilterBench b;
for (uint32_t i = 0; i < FLAGS_runs; ++i) {
b.Go();
FLAGS_seed += 100;
b.random_.Seed(FLAGS_seed);
}
}
return 0;
}
#endif // !defined(GFLAGS) || defined(ROCKSDB_LITE)