|
|
|
// Copyright (c) Facebook, Inc. and its affiliates. 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).
|
|
|
|
|
|
|
|
#pragma once
|
|
|
|
|
|
|
|
#include <cmath>
|
|
|
|
|
|
|
|
#include "port/port.h" // for PREFETCH
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
#include "util/fastrange.h"
|
|
|
|
#include "util/ribbon_alg.h"
|
|
|
|
|
|
|
|
namespace ROCKSDB_NAMESPACE {
|
|
|
|
|
|
|
|
namespace ribbon {
|
|
|
|
|
|
|
|
// RIBBON PHSF & RIBBON Filter (Rapid Incremental Boolean Banding ON-the-fly)
|
|
|
|
//
|
|
|
|
// ribbon_impl.h: templated (parameterized) standard implementations
|
|
|
|
//
|
|
|
|
// Ribbon is a Perfect Hash Static Function construction useful as a compact
|
|
|
|
// static Bloom filter alternative. See ribbon_alg.h for core algorithms
|
|
|
|
// and core design details.
|
|
|
|
//
|
|
|
|
// TODO: more details on trade-offs and practical issues.
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
//
|
|
|
|
// APIs for configuring Ribbon are in ribbon_config.h
|
|
|
|
|
|
|
|
// Ribbon implementations in this file take these parameters, which must be
|
|
|
|
// provided in a class/struct type with members expressed in this concept:
|
|
|
|
|
|
|
|
// concept TypesAndSettings {
|
|
|
|
// // See RibbonTypes and *Hasher in ribbon_alg.h, except here we have
|
|
|
|
// // the added constraint that Hash be equivalent to either uint32_t or
|
|
|
|
// // uint64_t.
|
|
|
|
// typename Hash;
|
|
|
|
// typename CoeffRow;
|
|
|
|
// typename ResultRow;
|
|
|
|
// typename Index;
|
|
|
|
// typename Key;
|
|
|
|
// static constexpr bool kFirstCoeffAlwaysOne;
|
|
|
|
//
|
|
|
|
// // An unsigned integer type for identifying a hash seed, typically
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
// // uint32_t or uint64_t. Importantly, this is the amount of data
|
|
|
|
// // stored in memory for identifying a raw seed. See StandardHasher.
|
|
|
|
// typename Seed;
|
|
|
|
//
|
|
|
|
// // When true, the PHSF implements a static filter, expecting just
|
|
|
|
// // keys as inputs for construction. When false, implements a general
|
|
|
|
// // PHSF and expects std::pair<Key, ResultRow> as inputs for
|
|
|
|
// // construction.
|
|
|
|
// static constexpr bool kIsFilter;
|
|
|
|
//
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
// // When true, enables a special "homogeneous" filter implementation that
|
|
|
|
// // is slightly faster to construct, and never fails to construct though
|
|
|
|
// // FP rate can quickly explode in cases where corresponding
|
|
|
|
// // non-homogeneous filter would fail (or nearly fail?) to construct.
|
|
|
|
// // For smaller filters, you can configure with ConstructionFailureChance
|
|
|
|
// // smaller than desired FP rate to largely counteract this effect.
|
|
|
|
// // TODO: configuring Homogeneous Ribbon for arbitrarily large filters
|
|
|
|
// // based on data from OptimizeHomogAtScale
|
|
|
|
// static constexpr bool kHomogeneous;
|
|
|
|
//
|
|
|
|
// // When true, adds a tiny bit more hashing logic on queries and
|
|
|
|
// // construction to improve utilization at the beginning and end of
|
|
|
|
// // the structure. Recommended when CoeffRow is only 64 bits (or
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
// // less), so typical num_starts < 10k. Although this is compatible
|
|
|
|
// // with kHomogeneous, the competing space vs. time priorities might
|
|
|
|
// // not be useful.
|
|
|
|
// static constexpr bool kUseSmash;
|
|
|
|
//
|
|
|
|
// // When true, allows number of "starts" to be zero, for best support
|
|
|
|
// // of the "no keys to add" case by always returning false for filter
|
|
|
|
// // queries. (This is distinct from the "keys added but no space for
|
|
|
|
// // any data" case, in which a filter always returns true.) The cost
|
|
|
|
// // supporting this is a conditional branch (probably predictable) in
|
|
|
|
// // queries.
|
|
|
|
// static constexpr bool kAllowZeroStarts;
|
|
|
|
//
|
|
|
|
// // A seedable stock hash function on Keys. All bits of Hash must
|
|
|
|
// // be reasonably high quality. XXH functions recommended, but
|
|
|
|
// // Murmur, City, Farm, etc. also work.
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
// static Hash HashFn(const Key &, Seed raw_seed);
|
|
|
|
// };
|
|
|
|
|
|
|
|
// A bit of a hack to automatically construct the type for
|
|
|
|
// AddInput based on a constexpr bool.
|
|
|
|
template <typename Key, typename ResultRow, bool IsFilter>
|
|
|
|
struct AddInputSelector {
|
|
|
|
// For general PHSF, not filter
|
|
|
|
using T = std::pair<Key, ResultRow>;
|
|
|
|
};
|
|
|
|
|
|
|
|
template <typename Key, typename ResultRow>
|
|
|
|
struct AddInputSelector<Key, ResultRow, true /*IsFilter*/> {
|
|
|
|
// For Filter
|
|
|
|
using T = Key;
|
|
|
|
};
|
|
|
|
|
|
|
|
// To avoid writing 'typename' everywhere that we use types like 'Index'
|
|
|
|
#define IMPORT_RIBBON_TYPES_AND_SETTINGS(TypesAndSettings) \
|
|
|
|
using CoeffRow = typename TypesAndSettings::CoeffRow; \
|
|
|
|
using ResultRow = typename TypesAndSettings::ResultRow; \
|
|
|
|
using Index = typename TypesAndSettings::Index; \
|
|
|
|
using Hash = typename TypesAndSettings::Hash; \
|
|
|
|
using Key = typename TypesAndSettings::Key; \
|
|
|
|
using Seed = typename TypesAndSettings::Seed; \
|
|
|
|
\
|
|
|
|
/* Some more additions */ \
|
|
|
|
using QueryInput = Key; \
|
|
|
|
using AddInput = typename ROCKSDB_NAMESPACE::ribbon::AddInputSelector< \
|
|
|
|
Key, ResultRow, TypesAndSettings::kIsFilter>::T; \
|
|
|
|
static constexpr auto kCoeffBits = \
|
|
|
|
static_cast<Index>(sizeof(CoeffRow) * 8U); \
|
|
|
|
\
|
|
|
|
/* Export to algorithm */ \
|
|
|
|
static constexpr bool kFirstCoeffAlwaysOne = \
|
|
|
|
TypesAndSettings::kFirstCoeffAlwaysOne; \
|
|
|
|
\
|
|
|
|
static_assert(sizeof(CoeffRow) + sizeof(ResultRow) + sizeof(Index) + \
|
|
|
|
sizeof(Hash) + sizeof(Key) + sizeof(Seed) + \
|
|
|
|
sizeof(QueryInput) + sizeof(AddInput) + kCoeffBits + \
|
|
|
|
kFirstCoeffAlwaysOne > \
|
|
|
|
0, \
|
|
|
|
"avoid unused warnings, semicolon expected after macro call")
|
|
|
|
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
#ifdef _MSC_VER
|
|
|
|
#pragma warning(push)
|
|
|
|
#pragma warning(disable : 4309) // cast truncating constant
|
|
|
|
#pragma warning(disable : 4307) // arithmetic constant overflow
|
|
|
|
#endif
|
|
|
|
|
|
|
|
// StandardHasher: A standard implementation of concepts RibbonTypes,
|
|
|
|
// PhsfQueryHasher, FilterQueryHasher, and BandingHasher from ribbon_alg.h.
|
|
|
|
//
|
|
|
|
// This implementation should be suitable for most all practical purposes
|
|
|
|
// as it "behaves" across a wide range of settings, with little room left
|
|
|
|
// for improvement. The key functionality in this hasher is generating
|
|
|
|
// CoeffRows, starts, and (for filters) ResultRows, which could be ~150
|
|
|
|
// bits of data or more, from a modest hash of 64 or even just 32 bits, with
|
|
|
|
// enough uniformity and bitwise independence to be close to "the best you
|
|
|
|
// can do" with available hash information in terms of FP rate and
|
|
|
|
// compactness. (64 bits recommended and sufficient for PHSF practical
|
|
|
|
// purposes.)
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
//
|
|
|
|
// Another feature of this hasher is a minimal "premixing" of seeds before
|
|
|
|
// they are provided to TypesAndSettings::HashFn in case that function does
|
|
|
|
// not provide sufficiently independent hashes when iterating merely
|
|
|
|
// sequentially on seeds. (This for example works around a problem with the
|
|
|
|
// preview version 0.7.2 of XXH3 used in RocksDB, a.k.a. XXH3p or Hash64, and
|
|
|
|
// MurmurHash1 used in RocksDB, a.k.a. Hash.) We say this pre-mixing step
|
|
|
|
// translates "ordinal seeds," which we iterate sequentially to find a
|
|
|
|
// solution, into "raw seeds," with many more bits changing for each
|
|
|
|
// iteration. The translation is an easily reversible lightweight mixing,
|
|
|
|
// not suitable for hashing on its own. An advantage of this approach is that
|
|
|
|
// StandardHasher can store just the raw seed (e.g. 64 bits) for fast query
|
|
|
|
// times, while from the application perspective, we can limit to a small
|
|
|
|
// number of ordinal keys (e.g. 64 in 6 bits) for saving in metadata.
|
|
|
|
//
|
|
|
|
// The default constructor initializes the seed to ordinal seed zero, which
|
|
|
|
// is equal to raw seed zero.
|
|
|
|
//
|
|
|
|
template <class TypesAndSettings>
|
|
|
|
class StandardHasher {
|
|
|
|
public:
|
|
|
|
IMPORT_RIBBON_TYPES_AND_SETTINGS(TypesAndSettings);
|
|
|
|
|
|
|
|
inline Hash GetHash(const Key& key) const {
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
return TypesAndSettings::HashFn(key, raw_seed_);
|
|
|
|
};
|
|
|
|
// For when AddInput == pair<Key, ResultRow> (kIsFilter == false)
|
|
|
|
inline Hash GetHash(const std::pair<Key, ResultRow>& bi) const {
|
|
|
|
return GetHash(bi.first);
|
|
|
|
};
|
|
|
|
inline Index GetStart(Hash h, Index num_starts) const {
|
|
|
|
// This is "critical path" code because it's required before memory
|
|
|
|
// lookup.
|
|
|
|
//
|
|
|
|
// FastRange gives us a fast and effective mapping from h to the
|
|
|
|
// appropriate range. This depends most, sometimes exclusively, on
|
|
|
|
// upper bits of h.
|
|
|
|
//
|
|
|
|
if (TypesAndSettings::kUseSmash) {
|
|
|
|
// Extra logic to "smash" entries at beginning and end, for
|
|
|
|
// better utilization. For example, without smash and with
|
|
|
|
// kFirstCoeffAlwaysOne, there's about a 30% chance that the
|
|
|
|
// first slot in the banding will be unused, and worse without
|
|
|
|
// kFirstCoeffAlwaysOne. The ending slots are even less utilized
|
|
|
|
// without smash.
|
|
|
|
//
|
|
|
|
// But since this only affects roughly kCoeffBits of the slots,
|
|
|
|
// it's usually small enough to be ignorable (less computation in
|
|
|
|
// this function) when number of slots is roughly 10k or larger.
|
|
|
|
//
|
|
|
|
// The best values for these smash weights might depend on how
|
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
|
|
|
// densely you're packing entries, and also kCoeffBits, but this
|
|
|
|
// seems to work well for roughly 95% success probability.
|
|
|
|
//
|
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
|
|
|
constexpr Index kFrontSmash = kCoeffBits / 4;
|
|
|
|
constexpr Index kBackSmash = kCoeffBits / 4;
|
|
|
|
Index start = FastRangeGeneric(h, num_starts + kFrontSmash + kBackSmash);
|
|
|
|
start = std::max(start, kFrontSmash);
|
|
|
|
start -= kFrontSmash;
|
|
|
|
start = std::min(start, num_starts - 1);
|
|
|
|
return start;
|
|
|
|
} else {
|
|
|
|
// For query speed, we allow small number of initial and final
|
|
|
|
// entries to be under-utilized.
|
|
|
|
// NOTE: This call statically enforces that Hash is equivalent to
|
|
|
|
// either uint32_t or uint64_t.
|
|
|
|
return FastRangeGeneric(h, num_starts);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
inline CoeffRow GetCoeffRow(Hash h) const {
|
|
|
|
// This is not so much "critical path" code because it can be done in
|
|
|
|
// parallel (instruction level) with memory lookup.
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
//
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
// When we might have many entries squeezed into a single start,
|
|
|
|
// we need reasonably good remixing for CoeffRow.
|
|
|
|
if (TypesAndSettings::kUseSmash) {
|
|
|
|
// Reasonably good, reasonably fast, reasonably general.
|
|
|
|
// Probably not 1:1 but probably close enough.
|
|
|
|
Unsigned128 a = Multiply64to128(h, kAltCoeffFactor1);
|
|
|
|
Unsigned128 b = Multiply64to128(h, kAltCoeffFactor2);
|
|
|
|
auto cr = static_cast<CoeffRow>(b ^ (a << 64) ^ (a >> 64));
|
|
|
|
|
|
|
|
// Now ensure the value is non-zero
|
|
|
|
if (kFirstCoeffAlwaysOne) {
|
|
|
|
cr |= 1;
|
|
|
|
} else {
|
|
|
|
// Still have to ensure some bit is non-zero
|
|
|
|
cr |= (cr == 0) ? 1 : 0;
|
|
|
|
}
|
|
|
|
return cr;
|
|
|
|
}
|
|
|
|
// If not kUseSmash, we ensure we're not squeezing many entries into a
|
|
|
|
// single start, in part by ensuring num_starts > num_slots / 2. Thus,
|
|
|
|
// here we do not need good remixing for CoeffRow, but just enough that
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
// (a) every bit is reasonably independent from Start.
|
|
|
|
// (b) every Hash-length bit subsequence of the CoeffRow has full or
|
|
|
|
// nearly full entropy from h.
|
|
|
|
// (c) if nontrivial bit subsequences within are correlated, it needs to
|
|
|
|
// be more complicated than exact copy or bitwise not (at least without
|
|
|
|
// kFirstCoeffAlwaysOne), or else there seems to be a kind of
|
|
|
|
// correlated clustering effect.
|
|
|
|
// (d) the CoeffRow is not zero, so that no one input on its own can
|
|
|
|
// doom construction success. (Preferably a mix of 1's and 0's if
|
|
|
|
// satisfying above.)
|
|
|
|
|
|
|
|
// First, establish sufficient bitwise independence from Start, with
|
|
|
|
// multiplication by a large random prime.
|
|
|
|
// Note that we cast to Hash because if we use product bits beyond
|
|
|
|
// original input size, that's going to correlate with Start (FastRange)
|
|
|
|
// even with a (likely) different multiplier here.
|
|
|
|
Hash a = h * kCoeffAndResultFactor;
|
|
|
|
|
|
|
|
static_assert(
|
|
|
|
sizeof(Hash) == sizeof(uint64_t) || sizeof(Hash) == sizeof(uint32_t),
|
|
|
|
"Supported sizes");
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
// If that's big enough, we're done. If not, we have to expand it,
|
|
|
|
// maybe up to 4x size.
|
|
|
|
uint64_t b;
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
if (sizeof(Hash) < sizeof(uint64_t)) {
|
|
|
|
// Almost-trivial hash expansion (OK - see above), favoring roughly
|
|
|
|
// equal number of 1's and 0's in result
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
b = (uint64_t{a} << 32) ^ (a ^ kCoeffXor32);
|
|
|
|
} else {
|
|
|
|
b = a;
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
}
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
static_assert(sizeof(CoeffRow) <= sizeof(Unsigned128), "Supported sizes");
|
|
|
|
Unsigned128 c;
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
if (sizeof(uint64_t) < sizeof(CoeffRow)) {
|
|
|
|
// Almost-trivial hash expansion (OK - see above), favoring roughly
|
|
|
|
// equal number of 1's and 0's in result
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
c = (Unsigned128{b} << 64) ^ (b ^ kCoeffXor64);
|
|
|
|
} else {
|
|
|
|
c = b;
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
}
|
|
|
|
auto cr = static_cast<CoeffRow>(c);
|
|
|
|
|
|
|
|
// Now ensure the value is non-zero
|
|
|
|
if (kFirstCoeffAlwaysOne) {
|
|
|
|
cr |= 1;
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
} else if (sizeof(CoeffRow) == sizeof(Hash)) {
|
|
|
|
// Still have to ensure some bit is non-zero
|
|
|
|
cr |= (cr == 0) ? 1 : 0;
|
|
|
|
} else {
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
// (We did trivial expansion with constant xor, which ensures some
|
|
|
|
// bits are non-zero.)
|
|
|
|
}
|
|
|
|
return cr;
|
|
|
|
}
|
|
|
|
inline ResultRow GetResultRowMask() const {
|
|
|
|
// TODO: will be used with InterleavedSolutionStorage?
|
|
|
|
// For now, all bits set (note: might be a small type so might need to
|
|
|
|
// narrow after promotion)
|
|
|
|
return static_cast<ResultRow>(~ResultRow{0});
|
|
|
|
}
|
|
|
|
inline ResultRow GetResultRowFromHash(Hash h) const {
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
if (TypesAndSettings::kIsFilter && !TypesAndSettings::kHomogeneous) {
|
|
|
|
// This is not so much "critical path" code because it can be done in
|
|
|
|
// parallel (instruction level) with memory lookup.
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
//
|
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
|
|
|
// ResultRow bits only needs to be independent from CoeffRow bits if
|
|
|
|
// many entries might have the same start location, where "many" is
|
|
|
|
// comparable to number of hash bits or kCoeffBits. If !kUseSmash
|
|
|
|
// and num_starts > kCoeffBits, it is safe and efficient to draw from
|
|
|
|
// the same bits computed for CoeffRow, which are reasonably
|
|
|
|
// independent from Start. (Inlining and common subexpression
|
|
|
|
// elimination with GetCoeffRow should make this
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
// a single shared multiplication in generated code when !kUseSmash.)
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
Hash a = h * kCoeffAndResultFactor;
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
// The bits here that are *most* independent of Start are the highest
|
|
|
|
// order bits (as in Knuth multiplicative hash). To make those the
|
|
|
|
// most preferred for use in the result row, we do a bswap here.
|
|
|
|
auto rr = static_cast<ResultRow>(EndianSwapValue(a));
|
|
|
|
return rr & GetResultRowMask();
|
|
|
|
} else {
|
|
|
|
// Must be zero
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// For when AddInput == Key (kIsFilter == true)
|
|
|
|
inline ResultRow GetResultRowFromInput(const Key&) const {
|
|
|
|
// Must be zero
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
// For when AddInput == pair<Key, ResultRow> (kIsFilter == false)
|
|
|
|
inline ResultRow GetResultRowFromInput(
|
|
|
|
const std::pair<Key, ResultRow>& bi) const {
|
|
|
|
// Simple extraction
|
|
|
|
return bi.second;
|
|
|
|
}
|
|
|
|
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
// Seed tracking APIs - see class comment
|
|
|
|
void SetRawSeed(Seed seed) { raw_seed_ = seed; }
|
|
|
|
Seed GetRawSeed() { return raw_seed_; }
|
|
|
|
void SetOrdinalSeed(Seed count) {
|
|
|
|
// A simple, reversible mixing of any size (whole bytes) up to 64 bits.
|
|
|
|
// This allows casting the raw seed to any smaller size we use for
|
|
|
|
// ordinal seeds without risk of duplicate raw seeds for unique ordinal
|
|
|
|
// seeds.
|
|
|
|
|
|
|
|
// Seed type might be smaller than numerical promotion size, but Hash
|
|
|
|
// should be at least that size, so we use Hash as intermediate type.
|
|
|
|
static_assert(sizeof(Seed) <= sizeof(Hash),
|
|
|
|
"Hash must be at least size of Seed");
|
|
|
|
|
|
|
|
// Multiply by a large random prime (one-to-one for any prefix of bits)
|
|
|
|
Hash tmp = count * kToRawSeedFactor;
|
|
|
|
// Within-byte one-to-one mixing
|
|
|
|
static_assert((kSeedMixMask & (kSeedMixMask >> kSeedMixShift)) == 0,
|
|
|
|
"Illegal mask+shift");
|
|
|
|
tmp ^= (tmp & kSeedMixMask) >> kSeedMixShift;
|
|
|
|
raw_seed_ = static_cast<Seed>(tmp);
|
|
|
|
// dynamic verification
|
|
|
|
assert(GetOrdinalSeed() == count);
|
|
|
|
}
|
|
|
|
Seed GetOrdinalSeed() {
|
|
|
|
Hash tmp = raw_seed_;
|
|
|
|
// Within-byte one-to-one mixing (its own inverse)
|
|
|
|
tmp ^= (tmp & kSeedMixMask) >> kSeedMixShift;
|
|
|
|
// Multiply by 64-bit multiplicative inverse
|
|
|
|
static_assert(kToRawSeedFactor * kFromRawSeedFactor == Hash{1},
|
|
|
|
"Must be inverses");
|
|
|
|
return static_cast<Seed>(tmp * kFromRawSeedFactor);
|
|
|
|
}
|
|
|
|
|
|
|
|
protected:
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
// For expanding hash:
|
|
|
|
// large random prime
|
|
|
|
static constexpr Hash kCoeffAndResultFactor =
|
|
|
|
static_cast<Hash>(0xc28f82822b650bedULL);
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
static constexpr uint64_t kAltCoeffFactor1 = 0x876f170be4f1fcb9U;
|
|
|
|
static constexpr uint64_t kAltCoeffFactor2 = 0xf0433a4aecda4c5fU;
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
// random-ish data
|
|
|
|
static constexpr uint32_t kCoeffXor32 = 0xa6293635U;
|
|
|
|
static constexpr uint64_t kCoeffXor64 = 0xc367844a6e52731dU;
|
|
|
|
|
|
|
|
// For pre-mixing seeds
|
|
|
|
static constexpr Hash kSeedMixMask = static_cast<Hash>(0xf0f0f0f0f0f0f0f0ULL);
|
|
|
|
static constexpr unsigned kSeedMixShift = 4U;
|
|
|
|
static constexpr Hash kToRawSeedFactor =
|
|
|
|
static_cast<Hash>(0xc78219a23eeadd03ULL);
|
|
|
|
static constexpr Hash kFromRawSeedFactor =
|
|
|
|
static_cast<Hash>(0xfe1a137d14b475abULL);
|
|
|
|
|
|
|
|
// See class description
|
|
|
|
Seed raw_seed_ = 0;
|
|
|
|
};
|
|
|
|
|
|
|
|
// StandardRehasher (and StandardRehasherAdapter): A variant of
|
|
|
|
// StandardHasher that uses the same type for keys as for hashes.
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
// This is primarily intended for building a Ribbon filter
|
|
|
|
// from existing hashes without going back to original inputs in
|
|
|
|
// order to apply a different seed. This hasher seeds a 1-to-1 mixing
|
|
|
|
// transformation to apply a seed to an existing hash. (Untested for
|
|
|
|
// hash-sized keys that are not already uniformly distributed.) This
|
|
|
|
// transformation builds on the seed pre-mixing done in StandardHasher.
|
|
|
|
//
|
|
|
|
// Testing suggests essentially no degradation of solution success rate
|
|
|
|
// vs. going back to original inputs when changing hash seeds. For example:
|
|
|
|
// Average re-seeds for solution with r=128, 1.02x overhead, and ~100k keys
|
|
|
|
// is about 1.10 for both StandardHasher and StandardRehasher.
|
|
|
|
//
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
// StandardRehasher is not really recommended for general PHSFs (not
|
|
|
|
// filters) because a collision in the original hash could prevent
|
|
|
|
// construction despite re-seeding the Rehasher. (Such collisions
|
|
|
|
// do not interfere with filter construction.)
|
|
|
|
//
|
|
|
|
// concept RehasherTypesAndSettings: like TypesAndSettings but
|
|
|
|
// does not require Key or HashFn.
|
|
|
|
template <class RehasherTypesAndSettings>
|
|
|
|
class StandardRehasherAdapter : public RehasherTypesAndSettings {
|
|
|
|
public:
|
|
|
|
using Hash = typename RehasherTypesAndSettings::Hash;
|
|
|
|
using Key = Hash;
|
|
|
|
using Seed = typename RehasherTypesAndSettings::Seed;
|
|
|
|
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
static Hash HashFn(const Hash& input, Seed raw_seed) {
|
|
|
|
// Note: raw_seed is already lightly pre-mixed, and this multiplication
|
|
|
|
// by a large prime is sufficient mixing (low-to-high bits) on top of
|
|
|
|
// that for good FastRange results, which depends primarily on highest
|
|
|
|
// bits. (The hashed CoeffRow and ResultRow are less sensitive to
|
|
|
|
// mixing than Start.)
|
|
|
|
// Also note: did consider adding ^ (input >> some) before the
|
|
|
|
// multiplication, but doesn't appear to be necessary.
|
|
|
|
return (input ^ raw_seed) * kRehashFactor;
|
|
|
|
}
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
|
|
|
|
private:
|
|
|
|
static constexpr Hash kRehashFactor =
|
|
|
|
static_cast<Hash>(0x6193d459236a3a0dULL);
|
|
|
|
};
|
|
|
|
|
|
|
|
// See comment on StandardRehasherAdapter
|
|
|
|
template <class RehasherTypesAndSettings>
|
|
|
|
using StandardRehasher =
|
|
|
|
StandardHasher<StandardRehasherAdapter<RehasherTypesAndSettings>>;
|
|
|
|
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
#ifdef _MSC_VER
|
|
|
|
#pragma warning(pop)
|
|
|
|
#endif
|
|
|
|
|
|
|
|
// Especially with smaller hashes (e.g. 32 bit), there can be noticeable
|
|
|
|
// false positives due to collisions in the Hash returned by GetHash.
|
|
|
|
// This function returns the expected FP rate due to those collisions,
|
|
|
|
// which can be added to the expected FP rate from the underlying data
|
|
|
|
// structure. (Note: technically, a + b is only a good approximation of
|
|
|
|
// 1-(1-a)(1-b) == a + b - a*b, if a and b are much closer to 0 than to 1.)
|
|
|
|
// The number of entries added can be a double here in case it's an
|
|
|
|
// average.
|
|
|
|
template <class Hasher, typename Numerical>
|
|
|
|
double ExpectedCollisionFpRate(const Hasher& hasher, Numerical added) {
|
|
|
|
// Standardize on the 'double' specialization
|
|
|
|
return ExpectedCollisionFpRate(hasher, 1.0 * added);
|
|
|
|
}
|
|
|
|
template <class Hasher>
|
|
|
|
double ExpectedCollisionFpRate(const Hasher& /*hasher*/, double added) {
|
|
|
|
// Technically, there could be overlap among the added, but ignoring that
|
|
|
|
// is typically close enough.
|
|
|
|
return added / std::pow(256.0, sizeof(typename Hasher::Hash));
|
|
|
|
}
|
|
|
|
|
|
|
|
// StandardBanding: a canonical implementation of BandingStorage and
|
|
|
|
// BacktrackStorage, with convenience API for banding (solving with on-the-fly
|
|
|
|
// Gaussian elimination) with and without backtracking.
|
|
|
|
template <class TypesAndSettings>
|
|
|
|
class StandardBanding : public StandardHasher<TypesAndSettings> {
|
|
|
|
public:
|
|
|
|
IMPORT_RIBBON_TYPES_AND_SETTINGS(TypesAndSettings);
|
|
|
|
|
|
|
|
StandardBanding(Index num_slots = 0, Index backtrack_size = 0) {
|
|
|
|
Reset(num_slots, backtrack_size);
|
|
|
|
}
|
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
|
|
|
|
|
|
|
void Reset(Index num_slots, Index backtrack_size = 0) {
|
|
|
|
if (num_slots == 0) {
|
|
|
|
// Unusual (TypesAndSettings::kAllowZeroStarts) or "uninitialized"
|
|
|
|
num_starts_ = 0;
|
|
|
|
} else {
|
|
|
|
// Normal
|
|
|
|
assert(num_slots >= kCoeffBits);
|
|
|
|
if (num_slots > num_slots_allocated_) {
|
|
|
|
coeff_rows_.reset(new CoeffRow[num_slots]());
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
if (!TypesAndSettings::kHomogeneous) {
|
|
|
|
// Note: don't strictly have to zero-init result_rows,
|
|
|
|
// except possible information leakage, etc ;)
|
|
|
|
result_rows_.reset(new ResultRow[num_slots]());
|
|
|
|
}
|
|
|
|
num_slots_allocated_ = num_slots;
|
|
|
|
} else {
|
|
|
|
for (Index i = 0; i < num_slots; ++i) {
|
|
|
|
coeff_rows_[i] = 0;
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
if (!TypesAndSettings::kHomogeneous) {
|
|
|
|
// Note: don't strictly have to zero-init result_rows,
|
|
|
|
// except possible information leakage, etc ;)
|
|
|
|
result_rows_[i] = 0;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
num_starts_ = num_slots - kCoeffBits + 1;
|
|
|
|
}
|
|
|
|
EnsureBacktrackSize(backtrack_size);
|
|
|
|
}
|
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
|
|
|
|
|
|
|
void EnsureBacktrackSize(Index backtrack_size) {
|
|
|
|
if (backtrack_size > backtrack_size_) {
|
|
|
|
backtrack_.reset(new Index[backtrack_size]);
|
|
|
|
backtrack_size_ = backtrack_size;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// ********************************************************************
|
|
|
|
// From concept BandingStorage
|
|
|
|
|
|
|
|
inline bool UsePrefetch() const {
|
|
|
|
// A rough guesstimate of when prefetching during construction pays off.
|
|
|
|
// TODO: verify/validate
|
|
|
|
return num_starts_ > 1500;
|
|
|
|
}
|
|
|
|
inline void Prefetch(Index i) const {
|
|
|
|
PREFETCH(&coeff_rows_[i], 1 /* rw */, 1 /* locality */);
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
if (!TypesAndSettings::kHomogeneous) {
|
|
|
|
PREFETCH(&result_rows_[i], 1 /* rw */, 1 /* locality */);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
inline void LoadRow(Index i, CoeffRow* cr, ResultRow* rr,
|
|
|
|
bool for_back_subst) const {
|
|
|
|
*cr = coeff_rows_[i];
|
|
|
|
if (TypesAndSettings::kHomogeneous) {
|
|
|
|
if (for_back_subst && *cr == 0) {
|
|
|
|
// Cheap pseudorandom data to fill unconstrained solution rows
|
|
|
|
*rr = static_cast<ResultRow>(i * 0x9E3779B185EBCA87ULL);
|
|
|
|
} else {
|
|
|
|
*rr = 0;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
*rr = result_rows_[i];
|
|
|
|
}
|
|
|
|
}
|
|
|
|
inline void StoreRow(Index i, CoeffRow cr, ResultRow rr) {
|
|
|
|
coeff_rows_[i] = cr;
|
|
|
|
if (TypesAndSettings::kHomogeneous) {
|
|
|
|
assert(rr == 0);
|
|
|
|
} else {
|
|
|
|
result_rows_[i] = rr;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
inline Index GetNumStarts() const { return num_starts_; }
|
|
|
|
|
|
|
|
// from concept BacktrackStorage, for when backtracking is used
|
|
|
|
inline bool UseBacktrack() const { return true; }
|
|
|
|
inline void BacktrackPut(Index i, Index to_save) { backtrack_[i] = to_save; }
|
|
|
|
inline Index BacktrackGet(Index i) const { return backtrack_[i]; }
|
|
|
|
|
|
|
|
// ********************************************************************
|
|
|
|
// Some useful API, still somewhat low level. Here an input is
|
|
|
|
// a Key for filters, or std::pair<Key, ResultRow> for general PHSF.
|
|
|
|
|
|
|
|
// Adds a range of inputs to the banding, returning true if successful.
|
|
|
|
// False means none or some may have been successfully added, so it's
|
|
|
|
// best to Reset this banding before any further use.
|
|
|
|
//
|
|
|
|
// Adding can fail even before all the "slots" are completely "full".
|
|
|
|
//
|
|
|
|
template <typename InputIterator>
|
|
|
|
bool AddRange(InputIterator begin, InputIterator end) {
|
|
|
|
assert(num_starts_ > 0 || TypesAndSettings::kAllowZeroStarts);
|
|
|
|
if (TypesAndSettings::kAllowZeroStarts && num_starts_ == 0) {
|
|
|
|
// Unusual. Can't add any in this case.
|
|
|
|
return begin == end;
|
|
|
|
}
|
|
|
|
// Normal
|
|
|
|
return BandingAddRange(this, *this, begin, end);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Adds a range of inputs to the banding, returning true if successful,
|
|
|
|
// or if unsuccessful, rolls back to state before this call and returns
|
|
|
|
// false. Caller guarantees that the number of inputs in this batch
|
|
|
|
// does not exceed `backtrack_size` provided to Reset.
|
|
|
|
//
|
|
|
|
// Adding can fail even before all the "slots" are completely "full".
|
|
|
|
//
|
|
|
|
template <typename InputIterator>
|
|
|
|
bool AddRangeOrRollBack(InputIterator begin, InputIterator end) {
|
|
|
|
assert(num_starts_ > 0 || TypesAndSettings::kAllowZeroStarts);
|
|
|
|
if (TypesAndSettings::kAllowZeroStarts && num_starts_ == 0) {
|
|
|
|
// Unusual. Can't add any in this case.
|
|
|
|
return begin == end;
|
|
|
|
}
|
|
|
|
// else Normal
|
|
|
|
return BandingAddRange(this, this, *this, begin, end);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Adds a single input to the banding, returning true if successful.
|
|
|
|
// If unsuccessful, returns false and banding state is unchanged.
|
|
|
|
//
|
|
|
|
// Adding can fail even before all the "slots" are completely "full".
|
|
|
|
//
|
|
|
|
bool Add(const AddInput& input) {
|
|
|
|
// Pointer can act as iterator
|
|
|
|
return AddRange(&input, &input + 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Return the number of "occupied" rows (with non-zero coefficients stored).
|
|
|
|
Index GetOccupiedCount() const {
|
|
|
|
Index count = 0;
|
|
|
|
if (num_starts_ > 0) {
|
|
|
|
const Index num_slots = num_starts_ + kCoeffBits - 1;
|
|
|
|
for (Index i = 0; i < num_slots; ++i) {
|
|
|
|
if (coeff_rows_[i] != 0) {
|
|
|
|
++count;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return count;
|
|
|
|
}
|
|
|
|
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
// Returns whether a row is "occupied" in the banding (non-zero
|
|
|
|
// coefficients stored). (Only recommended for debug/test)
|
|
|
|
bool IsOccupied(Index i) { return coeff_rows_[i] != 0; }
|
|
|
|
|
|
|
|
// ********************************************************************
|
|
|
|
// High-level API
|
|
|
|
|
|
|
|
// Iteratively (a) resets the structure for `num_slots`, (b) attempts
|
|
|
|
// to add the range of inputs, and (c) if unsuccessful, chooses next
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
// hash seed, until either successful or unsuccessful with all the
|
|
|
|
// allowed seeds. Returns true if successful. In that case, use
|
|
|
|
// GetOrdinalSeed() or GetRawSeed() to get the successful seed.
|
|
|
|
//
|
|
|
|
// The allowed sequence of hash seeds is determined by
|
|
|
|
// `starting_ordinal_seed,` the first ordinal seed to be attempted
|
|
|
|
// (see StandardHasher), and `ordinal_seed_mask,` a bit mask (power of
|
|
|
|
// two minus one) for the range of ordinal seeds to consider. The
|
|
|
|
// max number of seeds considered will be ordinal_seed_mask + 1.
|
|
|
|
// For filters we suggest `starting_ordinal_seed` be chosen randomly
|
|
|
|
// or round-robin, to minimize false positive correlations between keys.
|
|
|
|
//
|
|
|
|
// If unsuccessful, how best to continue is going to be application
|
|
|
|
// specific. It should be possible to choose parameters such that
|
|
|
|
// failure is extremely unlikely, using max_seed around 32 to 64.
|
|
|
|
// (TODO: APIs to help choose parameters) One option for fallback in
|
|
|
|
// constructing a filter is to construct a Bloom filter instead.
|
|
|
|
// Increasing num_slots is an option, but should not be used often
|
|
|
|
// unless construction maximum latency is a concern (rather than
|
|
|
|
// average running time of construction). Instead, choose parameters
|
|
|
|
// appropriately and trust that seeds are independent. (Also,
|
|
|
|
// increasing num_slots without changing hash seed would have a
|
|
|
|
// significant correlation in success, rather than independence.)
|
|
|
|
template <typename InputIterator>
|
|
|
|
bool ResetAndFindSeedToSolve(Index num_slots, InputIterator begin,
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
InputIterator end,
|
|
|
|
Seed starting_ordinal_seed = 0U,
|
|
|
|
Seed ordinal_seed_mask = 63U) {
|
|
|
|
// power of 2 minus 1
|
|
|
|
assert((ordinal_seed_mask & (ordinal_seed_mask + 1)) == 0);
|
|
|
|
// starting seed is within mask
|
|
|
|
assert((starting_ordinal_seed & ordinal_seed_mask) ==
|
|
|
|
starting_ordinal_seed);
|
|
|
|
starting_ordinal_seed &= ordinal_seed_mask; // if not debug
|
|
|
|
|
|
|
|
Seed cur_ordinal_seed = starting_ordinal_seed;
|
|
|
|
do {
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
StandardHasher<TypesAndSettings>::SetOrdinalSeed(cur_ordinal_seed);
|
|
|
|
Reset(num_slots);
|
|
|
|
bool success = AddRange(begin, end);
|
|
|
|
if (success) {
|
|
|
|
return true;
|
|
|
|
}
|
Ribbon: major re-work of hashing, seeds, and more (#7635)
Summary:
* Fully optimized StandardHasher, in terms of efficiently generating Start, CoeffRow, and ResultRow from a stock hash value, with sufficient independence between them to have no measurably degraded behavior. (Degraded behavior would be an FP rate higher than explainable by 2^-b and, if using a 32-bit stock hash function, expected stock hash collisions.) Details in code comments.
* Our standard 64-bit and 32-bit hash functions do not exhibit sufficient independence on sequential seeds (for one Ribbon construction attempt to have independent probability from the next). I have worked around this in the Ribbon code by "pre-mixing" "ordinal seeds," sequentially tried and appropriate for storage in persisted metadata, into "raw seeds," ready for application and appropriate for in-memory storage. This way the pre-mixing step (though fast) is only applied on loading or configuring the structure, not on each query or banding add.
* Fix a subtle flaw in which backtracking not clearing ResultRow data could lead to elevated FP rate on keys that were backtracked on and should (for generality) exhibit the same FP rate as novel keys.
* Added a basic test for PhsfQuery and construction algorithms (map or "retrieval structure" rather than set or filter), and made a few trivial related fixes.
* Better random configuration generation in unit tests
* Some other minor cleanup / clarification / etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7635
Test Plan: unit tests included
Reviewed By: jay-zhuang
Differential Revision: D24738978
Pulled By: pdillinger
fbshipit-source-id: f9d03599d9e2ca3e30e9d3e7d81cd936b56f76f0
4 years ago
|
|
|
cur_ordinal_seed = (cur_ordinal_seed + 1) & ordinal_seed_mask;
|
|
|
|
} while (cur_ordinal_seed != starting_ordinal_seed);
|
|
|
|
// Reached limit by circling around
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
protected:
|
|
|
|
// TODO: explore combining in a struct
|
|
|
|
std::unique_ptr<CoeffRow[]> coeff_rows_;
|
|
|
|
std::unique_ptr<ResultRow[]> result_rows_;
|
|
|
|
// We generally store "starts" instead of slots for speed of GetStart(),
|
|
|
|
// as in StandardHasher.
|
|
|
|
Index num_starts_ = 0;
|
|
|
|
Index num_slots_allocated_ = 0;
|
|
|
|
std::unique_ptr<Index[]> backtrack_;
|
|
|
|
Index backtrack_size_ = 0;
|
|
|
|
};
|
|
|
|
|
|
|
|
// Implements concept SimpleSolutionStorage, mostly for demonstration
|
|
|
|
// purposes. This is "in memory" only because it does not handle byte
|
|
|
|
// ordering issues for serialization.
|
|
|
|
template <class TypesAndSettings>
|
|
|
|
class InMemSimpleSolution {
|
|
|
|
public:
|
|
|
|
IMPORT_RIBBON_TYPES_AND_SETTINGS(TypesAndSettings);
|
|
|
|
|
|
|
|
void PrepareForNumStarts(Index num_starts) {
|
|
|
|
if (TypesAndSettings::kAllowZeroStarts && num_starts == 0) {
|
|
|
|
// Unusual
|
|
|
|
num_starts_ = 0;
|
|
|
|
} else {
|
|
|
|
// Normal
|
|
|
|
const Index num_slots = num_starts + kCoeffBits - 1;
|
|
|
|
assert(num_slots >= kCoeffBits);
|
|
|
|
if (num_slots > num_slots_allocated_) {
|
|
|
|
// Do not need to init the memory
|
|
|
|
solution_rows_.reset(new ResultRow[num_slots]);
|
|
|
|
num_slots_allocated_ = num_slots;
|
|
|
|
}
|
|
|
|
num_starts_ = num_starts;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
Index GetNumStarts() const { return num_starts_; }
|
|
|
|
|
|
|
|
ResultRow Load(Index slot_num) const { return solution_rows_[slot_num]; }
|
|
|
|
|
|
|
|
void Store(Index slot_num, ResultRow solution_row) {
|
|
|
|
solution_rows_[slot_num] = solution_row;
|
|
|
|
}
|
|
|
|
|
|
|
|
// ********************************************************************
|
|
|
|
// High-level API
|
|
|
|
|
|
|
|
template <typename BandingStorage>
|
|
|
|
void BackSubstFrom(const BandingStorage& bs) {
|
|
|
|
if (TypesAndSettings::kAllowZeroStarts && bs.GetNumStarts() == 0) {
|
|
|
|
// Unusual
|
|
|
|
PrepareForNumStarts(0);
|
|
|
|
} else {
|
|
|
|
// Normal
|
|
|
|
SimpleBackSubst(this, bs);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template <typename PhsfQueryHasher>
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
ResultRow PhsfQuery(const Key& input, const PhsfQueryHasher& hasher) const {
|
|
|
|
// assert(!TypesAndSettings::kIsFilter); Can be useful in testing
|
|
|
|
if (TypesAndSettings::kAllowZeroStarts && num_starts_ == 0) {
|
|
|
|
// Unusual
|
|
|
|
return 0;
|
|
|
|
} else {
|
|
|
|
// Normal
|
|
|
|
return SimplePhsfQuery(input, hasher, *this);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template <typename FilterQueryHasher>
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
bool FilterQuery(const Key& input, const FilterQueryHasher& hasher) const {
|
|
|
|
assert(TypesAndSettings::kIsFilter);
|
|
|
|
if (TypesAndSettings::kAllowZeroStarts && num_starts_ == 0) {
|
|
|
|
// Unusual. Zero starts presumes no keys added -> always false
|
|
|
|
return false;
|
|
|
|
} else {
|
|
|
|
// Normal, or upper_num_columns_ == 0 means "no space for data" and
|
|
|
|
// thus will always return true.
|
|
|
|
return SimpleFilterQuery(input, hasher, *this);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
double ExpectedFpRate() const {
|
|
|
|
assert(TypesAndSettings::kIsFilter);
|
|
|
|
if (TypesAndSettings::kAllowZeroStarts && num_starts_ == 0) {
|
|
|
|
// Unusual, but we don't have FPs if we always return false.
|
|
|
|
return 0.0;
|
|
|
|
}
|
|
|
|
// else Normal
|
|
|
|
|
|
|
|
// Each result (solution) bit (column) cuts FP rate in half
|
|
|
|
return std::pow(0.5, 8U * sizeof(ResultRow));
|
|
|
|
}
|
|
|
|
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
// ********************************************************************
|
|
|
|
// Static high-level API
|
|
|
|
|
|
|
|
// Round up to a number of slots supported by this structure. Note that
|
|
|
|
// this needs to be must be taken into account for the banding if this
|
|
|
|
// solution layout/storage is to be used.
|
|
|
|
static Index RoundUpNumSlots(Index num_slots) {
|
|
|
|
// Must be at least kCoeffBits for at least one start
|
|
|
|
// Or if not smash, even more because hashing not equipped
|
|
|
|
// for stacking up so many entries on a single start location
|
|
|
|
auto min_slots = kCoeffBits * (TypesAndSettings::kUseSmash ? 1 : 2);
|
|
|
|
return std::max(num_slots, static_cast<Index>(min_slots));
|
|
|
|
}
|
|
|
|
|
|
|
|
protected:
|
|
|
|
// We generally store "starts" instead of slots for speed of GetStart(),
|
|
|
|
// as in StandardHasher.
|
|
|
|
Index num_starts_ = 0;
|
|
|
|
Index num_slots_allocated_ = 0;
|
|
|
|
std::unique_ptr<ResultRow[]> solution_rows_;
|
|
|
|
};
|
|
|
|
|
|
|
|
// Implements concept InterleavedSolutionStorage always using little-endian
|
|
|
|
// byte order, so easy for serialization/deserialization. This implementation
|
|
|
|
// fully supports fractional bits per key, where any number of segments
|
|
|
|
// (number of bytes multiple of sizeof(CoeffRow)) can be used with any number
|
|
|
|
// of slots that is a multiple of kCoeffBits.
|
|
|
|
//
|
|
|
|
// The structure is passed an externally allocated/de-allocated byte buffer
|
|
|
|
// that is optionally pre-populated (from storage) for answering queries,
|
|
|
|
// or can be populated by BackSubstFrom.
|
|
|
|
//
|
|
|
|
template <class TypesAndSettings>
|
|
|
|
class SerializableInterleavedSolution {
|
|
|
|
public:
|
|
|
|
IMPORT_RIBBON_TYPES_AND_SETTINGS(TypesAndSettings);
|
|
|
|
|
|
|
|
// Does not take ownership of `data` but uses it (up to `data_len` bytes)
|
|
|
|
// throughout lifetime
|
|
|
|
SerializableInterleavedSolution(char* data, size_t data_len)
|
|
|
|
: data_(data), data_len_(data_len) {}
|
|
|
|
|
|
|
|
void PrepareForNumStarts(Index num_starts) {
|
|
|
|
assert(num_starts == 0 || (num_starts % kCoeffBits == 1));
|
|
|
|
num_starts_ = num_starts;
|
|
|
|
|
|
|
|
InternalConfigure();
|
|
|
|
}
|
|
|
|
|
|
|
|
Index GetNumStarts() const { return num_starts_; }
|
|
|
|
|
|
|
|
Index GetNumBlocks() const {
|
|
|
|
const Index num_slots = num_starts_ + kCoeffBits - 1;
|
|
|
|
return num_slots / kCoeffBits;
|
|
|
|
}
|
|
|
|
|
|
|
|
Index GetUpperNumColumns() const { return upper_num_columns_; }
|
|
|
|
|
|
|
|
Index GetUpperStartBlock() const { return upper_start_block_; }
|
|
|
|
|
|
|
|
Index GetNumSegments() const {
|
|
|
|
return static_cast<Index>(data_len_ / sizeof(CoeffRow));
|
|
|
|
}
|
|
|
|
|
|
|
|
CoeffRow LoadSegment(Index segment_num) const {
|
|
|
|
assert(data_ != nullptr); // suppress clang analyzer report
|
|
|
|
return DecodeFixedGeneric<CoeffRow>(data_ + segment_num * sizeof(CoeffRow));
|
|
|
|
}
|
|
|
|
void StoreSegment(Index segment_num, CoeffRow val) {
|
|
|
|
assert(data_ != nullptr); // suppress clang analyzer report
|
|
|
|
EncodeFixedGeneric(data_ + segment_num * sizeof(CoeffRow), val);
|
|
|
|
}
|
Add prefetching (batched MultiGet) for experimental Ribbon filter (#7889)
Summary:
Adds support for prefetching data in Ribbon queries,
which especially optimizes batched Ribbon queries for MultiGet
(~222ns/key to ~97ns/key) but also single key queries on cold memory
(~333ns to ~226ns) because many queries span more than one cache line.
This required some refactoring of the query algorithm, and there
does not appear to be a noticeable regression in "hot memory" query
times (perhaps from 48ns to 50ns).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7889
Test Plan:
existing unit tests, plus performance validation with
filter_bench:
Each data point is the best of two runs. I saturated the machine
CPUs with other filter_bench runs in the background.
Before:
$ ./filter_bench -impl=3 -m_keys_total_max=200 -average_keys_per_filter=100000 -m_queries=50
WARNING: Assertions are enabled; benchmarks unnecessarily slow
Building...
Build avg ns/key: 125.86
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
Prelim FP rate %: 0.951827
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 48.0111
Batched, prepared net ns/op: 222.384
Batched, unprepared net ns/op: 343.908
Skewed 50% in 1% net ns/op: 252.916
Skewed 80% in 20% net ns/op: 320.579
Random filter net ns/op: 332.957
After:
$ ./filter_bench -impl=3 -m_keys_total_max=200 -average_keys_per_filter=100000 -m_queries=50
WARNING: Assertions are enabled; benchmarks unnecessarily slow
Building...
Build avg ns/key: 128.117
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
Prelim FP rate %: 0.951827
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 49.8812
Batched, prepared net ns/op: 97.1514
Batched, unprepared net ns/op: 222.025
Skewed 50% in 1% net ns/op: 197.48
Skewed 80% in 20% net ns/op: 212.457
Random filter net ns/op: 226.464
Bloom comparison, for reference:
$ ./filter_bench -impl=2 -m_keys_total_max=200 -average_keys_per_filter=100000 -m_queries=50
WARNING: Assertions are enabled; benchmarks unnecessarily slow
Building...
Build avg ns/key: 35.3042
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
Prelim FP rate %: 0.965327
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 9.09931
Batched, prepared net ns/op: 34.21
Batched, unprepared net ns/op: 88.8564
Skewed 50% in 1% net ns/op: 139.75
Skewed 80% in 20% net ns/op: 181.264
Random filter net ns/op: 173.88
Reviewed By: jay-zhuang
Differential Revision: D26378710
Pulled By: pdillinger
fbshipit-source-id: 058428967c55ed763698284cd3b4bbe3351b6e69
4 years ago
|
|
|
void PrefetchSegmentRange(Index begin_segment_num,
|
|
|
|
Index end_segment_num) const {
|
|
|
|
if (end_segment_num == begin_segment_num) {
|
|
|
|
// Nothing to do
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
char* cur = data_ + begin_segment_num * sizeof(CoeffRow);
|
|
|
|
char* last = data_ + (end_segment_num - 1) * sizeof(CoeffRow);
|
|
|
|
while (cur < last) {
|
|
|
|
PREFETCH(cur, 0 /* rw */, 1 /* locality */);
|
|
|
|
cur += CACHE_LINE_SIZE;
|
|
|
|
}
|
|
|
|
PREFETCH(last, 0 /* rw */, 1 /* locality */);
|
|
|
|
}
|
|
|
|
|
|
|
|
// ********************************************************************
|
|
|
|
// High-level API
|
|
|
|
|
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
|
|
|
void ConfigureForNumBlocks(Index num_blocks) {
|
|
|
|
if (num_blocks == 0) {
|
|
|
|
PrepareForNumStarts(0);
|
|
|
|
} else {
|
|
|
|
PrepareForNumStarts(num_blocks * kCoeffBits - kCoeffBits + 1);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void ConfigureForNumSlots(Index num_slots) {
|
|
|
|
assert(num_slots % kCoeffBits == 0);
|
|
|
|
ConfigureForNumBlocks(num_slots / kCoeffBits);
|
|
|
|
}
|
|
|
|
|
|
|
|
template <typename BandingStorage>
|
|
|
|
void BackSubstFrom(const BandingStorage& bs) {
|
|
|
|
if (TypesAndSettings::kAllowZeroStarts && bs.GetNumStarts() == 0) {
|
|
|
|
// Unusual
|
|
|
|
PrepareForNumStarts(0);
|
|
|
|
} else {
|
|
|
|
// Normal
|
|
|
|
InterleavedBackSubst(this, bs);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template <typename PhsfQueryHasher>
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
ResultRow PhsfQuery(const Key& input, const PhsfQueryHasher& hasher) const {
|
|
|
|
// assert(!TypesAndSettings::kIsFilter); Can be useful in testing
|
|
|
|
if (TypesAndSettings::kAllowZeroStarts && num_starts_ == 0) {
|
|
|
|
// Unusual
|
|
|
|
return 0;
|
|
|
|
} else {
|
|
|
|
// Normal
|
Add prefetching (batched MultiGet) for experimental Ribbon filter (#7889)
Summary:
Adds support for prefetching data in Ribbon queries,
which especially optimizes batched Ribbon queries for MultiGet
(~222ns/key to ~97ns/key) but also single key queries on cold memory
(~333ns to ~226ns) because many queries span more than one cache line.
This required some refactoring of the query algorithm, and there
does not appear to be a noticeable regression in "hot memory" query
times (perhaps from 48ns to 50ns).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7889
Test Plan:
existing unit tests, plus performance validation with
filter_bench:
Each data point is the best of two runs. I saturated the machine
CPUs with other filter_bench runs in the background.
Before:
$ ./filter_bench -impl=3 -m_keys_total_max=200 -average_keys_per_filter=100000 -m_queries=50
WARNING: Assertions are enabled; benchmarks unnecessarily slow
Building...
Build avg ns/key: 125.86
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
Prelim FP rate %: 0.951827
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 48.0111
Batched, prepared net ns/op: 222.384
Batched, unprepared net ns/op: 343.908
Skewed 50% in 1% net ns/op: 252.916
Skewed 80% in 20% net ns/op: 320.579
Random filter net ns/op: 332.957
After:
$ ./filter_bench -impl=3 -m_keys_total_max=200 -average_keys_per_filter=100000 -m_queries=50
WARNING: Assertions are enabled; benchmarks unnecessarily slow
Building...
Build avg ns/key: 128.117
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
Prelim FP rate %: 0.951827
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 49.8812
Batched, prepared net ns/op: 97.1514
Batched, unprepared net ns/op: 222.025
Skewed 50% in 1% net ns/op: 197.48
Skewed 80% in 20% net ns/op: 212.457
Random filter net ns/op: 226.464
Bloom comparison, for reference:
$ ./filter_bench -impl=2 -m_keys_total_max=200 -average_keys_per_filter=100000 -m_queries=50
WARNING: Assertions are enabled; benchmarks unnecessarily slow
Building...
Build avg ns/key: 35.3042
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
Prelim FP rate %: 0.965327
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 9.09931
Batched, prepared net ns/op: 34.21
Batched, unprepared net ns/op: 88.8564
Skewed 50% in 1% net ns/op: 139.75
Skewed 80% in 20% net ns/op: 181.264
Random filter net ns/op: 173.88
Reviewed By: jay-zhuang
Differential Revision: D26378710
Pulled By: pdillinger
fbshipit-source-id: 058428967c55ed763698284cd3b4bbe3351b6e69
4 years ago
|
|
|
// NOTE: not using a struct to encourage compiler optimization
|
|
|
|
Hash hash;
|
|
|
|
Index segment_num;
|
|
|
|
Index num_columns;
|
|
|
|
Index start_bit;
|
|
|
|
InterleavedPrepareQuery(input, hasher, *this, &hash, &segment_num,
|
|
|
|
&num_columns, &start_bit);
|
|
|
|
return InterleavedPhsfQuery(hash, segment_num, num_columns, start_bit,
|
|
|
|
hasher, *this);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template <typename FilterQueryHasher>
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
bool FilterQuery(const Key& input, const FilterQueryHasher& hasher) const {
|
|
|
|
assert(TypesAndSettings::kIsFilter);
|
|
|
|
if (TypesAndSettings::kAllowZeroStarts && num_starts_ == 0) {
|
|
|
|
// Unusual. Zero starts presumes no keys added -> always false
|
|
|
|
return false;
|
|
|
|
} else {
|
|
|
|
// Normal, or upper_num_columns_ == 0 means "no space for data" and
|
|
|
|
// thus will always return true.
|
Add prefetching (batched MultiGet) for experimental Ribbon filter (#7889)
Summary:
Adds support for prefetching data in Ribbon queries,
which especially optimizes batched Ribbon queries for MultiGet
(~222ns/key to ~97ns/key) but also single key queries on cold memory
(~333ns to ~226ns) because many queries span more than one cache line.
This required some refactoring of the query algorithm, and there
does not appear to be a noticeable regression in "hot memory" query
times (perhaps from 48ns to 50ns).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7889
Test Plan:
existing unit tests, plus performance validation with
filter_bench:
Each data point is the best of two runs. I saturated the machine
CPUs with other filter_bench runs in the background.
Before:
$ ./filter_bench -impl=3 -m_keys_total_max=200 -average_keys_per_filter=100000 -m_queries=50
WARNING: Assertions are enabled; benchmarks unnecessarily slow
Building...
Build avg ns/key: 125.86
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
Prelim FP rate %: 0.951827
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 48.0111
Batched, prepared net ns/op: 222.384
Batched, unprepared net ns/op: 343.908
Skewed 50% in 1% net ns/op: 252.916
Skewed 80% in 20% net ns/op: 320.579
Random filter net ns/op: 332.957
After:
$ ./filter_bench -impl=3 -m_keys_total_max=200 -average_keys_per_filter=100000 -m_queries=50
WARNING: Assertions are enabled; benchmarks unnecessarily slow
Building...
Build avg ns/key: 128.117
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
Prelim FP rate %: 0.951827
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 49.8812
Batched, prepared net ns/op: 97.1514
Batched, unprepared net ns/op: 222.025
Skewed 50% in 1% net ns/op: 197.48
Skewed 80% in 20% net ns/op: 212.457
Random filter net ns/op: 226.464
Bloom comparison, for reference:
$ ./filter_bench -impl=2 -m_keys_total_max=200 -average_keys_per_filter=100000 -m_queries=50
WARNING: Assertions are enabled; benchmarks unnecessarily slow
Building...
Build avg ns/key: 35.3042
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
Prelim FP rate %: 0.965327
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 9.09931
Batched, prepared net ns/op: 34.21
Batched, unprepared net ns/op: 88.8564
Skewed 50% in 1% net ns/op: 139.75
Skewed 80% in 20% net ns/op: 181.264
Random filter net ns/op: 173.88
Reviewed By: jay-zhuang
Differential Revision: D26378710
Pulled By: pdillinger
fbshipit-source-id: 058428967c55ed763698284cd3b4bbe3351b6e69
4 years ago
|
|
|
// NOTE: not using a struct to encourage compiler optimization
|
|
|
|
Hash hash;
|
|
|
|
Index segment_num;
|
|
|
|
Index num_columns;
|
|
|
|
Index start_bit;
|
|
|
|
InterleavedPrepareQuery(input, hasher, *this, &hash, &segment_num,
|
|
|
|
&num_columns, &start_bit);
|
|
|
|
return InterleavedFilterQuery(hash, segment_num, num_columns, start_bit,
|
|
|
|
hasher, *this);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Refine Ribbon configuration, improve testing, add Homogeneous (#7879)
Summary:
This change only affects non-schema-critical aspects of the production candidate Ribbon filter. Specifically, it refines choice of internal configuration parameters based on inputs. The changes are minor enough that the schema tests in bloom_test, some of which depend on this, are unaffected. There are also some minor optimizations and refactorings.
This would be a schema change for "smash" Ribbon, to fix some known issues with small filters, but "smash" Ribbon is not accessible in public APIs. Unit test CompactnessAndBacktrackAndFpRate updated to test small and medium-large filters. Run with --thoroughness=100 or so for much better detection power (not appropriate for continuous regression testing).
Homogenous Ribbon:
This change adds internally a Ribbon filter variant we call Homogeneous Ribbon, in collaboration with Stefan Walzer. The expected "result" value for every key is zero, instead of computed from a hash. Entropy for queries not to be false positives comes from free variables ("overhead") in the solution structure, which are populated pseudorandomly. Construction is slightly faster for not tracking result values, and never fails. Instead, FP rate can jump up whenever and whereever entries are packed too tightly. For small structures, we can choose overhead to make this FP rate jump unlikely, as seen in updated unit test CompactnessAndBacktrackAndFpRate.
Unlike standard Ribbon, Homogeneous Ribbon seems to scale to arbitrary number of keys when accepting an FP rate penalty for small pockets of high FP rate in the structure. For example, 64-bit ribbon with 8 solution columns and 10% allocated space overhead for slots seems to achieve about 10.5% space overhead vs. information-theoretic minimum based on its observed FP rate with expected pockets of degradation. (FP rate is close to 1/256.) If targeting a higher FP rate with fewer solution columns, Homogeneous Ribbon can be even more space efficient, because the penalty from degradation is relatively smaller. If targeting a lower FP rate, Homogeneous Ribbon is less space efficient, as more allocated overhead is needed to keep the FP rate impact of degradation relatively under control. The new OptimizeHomogAtScale tool in ribbon_test helps to find these optimal allocation overheads for different numbers of solution columns. And Ribbon widths, with 128-bit Ribbon apparently cutting space overheads in half vs. 64-bit.
Other misc item specifics:
* Ribbon APIs in util/ribbon_config.h now provide configuration data for not just 5% construction failure rate (95% success), but also 50% and 0.1%.
* Note that the Ribbon structure does not exhibit "threshold" behavior as standard Xor filter does, so there is a roughly fixed space penalty to cut construction failure rate in half. Thus, there isn't really an "almost sure" setting.
* Although we can extrapolate settings for large filters, we don't have a good formula for configuring smaller filters (< 2^17 slots or so), and efforts to summarize with a formula have failed. Thus, small data is hard-coded from updated FindOccupancy tool.
* Enhances ApproximateNumEntries for public API Ribbon using more precise data (new API GetNumToAdd), thus a more accurate but not perfect reversal of CalculateSpace. (bloom_test updated to expect the greater precision)
* Move EndianSwapValue from coding.h to coding_lean.h to keep Ribbon code easily transferable from RocksDB
* Add some missing 'const' to member functions
* Small optimization to 128-bit BitParity
* Small refactoring of BandingStorage in ribbon_alg.h to support Homogeneous Ribbon
* CompactnessAndBacktrackAndFpRate now has an "expand" test: on construction failure, a possible alternative to re-seeding hash functions is simply to increase the number of slots (allocated space overhead) and try again with essentially the same hash values. (Start locations will be different roundings of the same scaled hash values--because fastrange not mod.) This seems to be as effective or more effective than re-seeding, as long as we increase the number of slots (m) by roughly m += m/w where w is the Ribbon width. This way, there is effectively an expansion by one slot for each ribbon-width window in the banding. (This approach assumes that getting "bad data" from your hash function is as unlikely as it naturally should be, e.g. no adversary.)
* 32-bit and 16-bit Ribbon configurations are added to ribbon_test for understanding their behavior, e.g. with FindOccupancy. They are not considered useful at this time and not tested with CompactnessAndBacktrackAndFpRate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7879
Test Plan: unit test updates included
Reviewed By: jay-zhuang
Differential Revision: D26371245
Pulled By: pdillinger
fbshipit-source-id: da6600d90a3785b99ad17a88b2a3027710b4ea3a
4 years ago
|
|
|
double ExpectedFpRate() const {
|
|
|
|
assert(TypesAndSettings::kIsFilter);
|
|
|
|
if (TypesAndSettings::kAllowZeroStarts && num_starts_ == 0) {
|
|
|
|
// Unusual. Zero starts presumes no keys added -> always false
|
|
|
|
return 0.0;
|
|
|
|
}
|
|
|
|
// else Normal
|
|
|
|
|
|
|
|
// Note: Ignoring smash setting; still close enough in that case
|
|
|
|
double lower_portion =
|
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
|
|
|
(upper_start_block_ * 1.0 * kCoeffBits) / num_starts_;
|
|
|
|
|
|
|
|
// Each result (solution) bit (column) cuts FP rate in half. Weight that
|
|
|
|
// for upper and lower number of bits (columns).
|
|
|
|
return lower_portion * std::pow(0.5, upper_num_columns_ - 1) +
|
|
|
|
(1.0 - lower_portion) * std::pow(0.5, upper_num_columns_);
|
|
|
|
}
|
|
|
|
|
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
|
|
|
// ********************************************************************
|
|
|
|
// Static high-level API
|
|
|
|
|
|
|
|
// Round up to a number of slots supported by this structure. Note that
|
|
|
|
// this needs to be must be taken into account for the banding if this
|
|
|
|
// solution layout/storage is to be used.
|
|
|
|
static Index RoundUpNumSlots(Index num_slots) {
|
|
|
|
// Must be multiple of kCoeffBits
|
|
|
|
Index corrected = (num_slots + kCoeffBits - 1) / kCoeffBits * kCoeffBits;
|
|
|
|
|
|
|
|
// Do not use num_starts==1 unless kUseSmash, because the hashing
|
|
|
|
// might not be equipped for stacking up so many entries on a
|
|
|
|
// single start location.
|
|
|
|
if (!TypesAndSettings::kUseSmash && corrected == kCoeffBits) {
|
|
|
|
corrected += kCoeffBits;
|
|
|
|
}
|
|
|
|
return corrected;
|
|
|
|
}
|
|
|
|
|
Support optimize_filters_for_memory for Ribbon filter (#7774)
Summary:
Primarily this change refactors the optimize_filters_for_memory
code for Bloom filters, based on malloc_usable_size, to also work for
Ribbon filters.
This change also replaces the somewhat slow but general
BuiltinFilterBitsBuilder::ApproximateNumEntries with
implementation-specific versions for Ribbon (new) and Legacy Bloom
(based on a recently deleted version). The reason is to emphasize
speed in ApproximateNumEntries rather than 100% accuracy.
Justification: ApproximateNumEntries (formerly CalculateNumEntry) is
only used by RocksDB for range-partitioned filters, called each time we
start to construct one. (In theory, it should be possible to reuse the
estimate, but the abstractions provided by FilterPolicy don't really
make that workable.) But this is only used as a heuristic estimate for
hitting a desired partitioned filter size because of alignment to data
blocks, which have various numbers of unique keys or prefixes. The two
factors lead us to prioritize reasonable speed over 100% accuracy.
optimize_filters_for_memory adds extra complication, because precisely
calculating num_entries for some allowed number of bytes depends on state
with optimize_filters_for_memory enabled. And the allocator-agnostic
implementation of optimize_filters_for_memory, using malloc_usable_size,
means we would have to actually allocate memory, many times, just to
precisely determine how many entries (keys) could be added and stay below
some size budget, for the current state. (In a draft, I got this
working, and then realized the balance of speed vs. accuracy was all
wrong.)
So related to that, I have made CalculateSpace, an internal-only API
only used for testing, non-authoritative also if
optimize_filters_for_memory is enabled. This simplifies some code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7774
Test Plan:
unit test updated, and for FilterSize test, range of tested
values is greatly expanded (still super fast)
Also tested `db_bench -benchmarks=fillrandom,stats -bloom_bits=10 -num=1000000 -partition_index_and_filters -format_version=5 [-optimize_filters_for_memory] [-use_ribbon_filter]` with temporary debug output of generated filter sizes.
Bloom+optimize_filters_for_memory:
1 Filter size: 197 (224 in memory)
134 Filter size: 3525 (3584 in memory)
107 Filter size: 4037 (4096 in memory)
Total on disk: 904,506
Total in memory: 918,752
Ribbon+optimize_filters_for_memory:
1 Filter size: 3061 (3072 in memory)
110 Filter size: 3573 (3584 in memory)
58 Filter size: 4085 (4096 in memory)
Total on disk: 633,021 (-30.0%)
Total in memory: 634,880 (-30.9%)
Bloom (no offm):
1 Filter size: 261 (320 in memory)
1 Filter size: 3333 (3584 in memory)
240 Filter size: 3717 (4096 in memory)
Total on disk: 895,674 (-1% on disk vs. +offm; known tolerable overhead of offm)
Total in memory: 986,944 (+7.4% vs. +offm)
Ribbon (no offm):
1 Filter size: 2949 (3072 in memory)
1 Filter size: 3381 (3584 in memory)
167 Filter size: 3701 (4096 in memory)
Total on disk: 624,397 (-30.3% vs. Bloom)
Total in memory: 690,688 (-30.0% vs. Bloom)
Note that optimize_filters_for_memory is even more effective for Ribbon filter than for cache-local Bloom, because it can close the unused memory gap even tighter than Bloom filter, because of 16 byte increments for Ribbon vs. 64 byte increments for Bloom.
Reviewed By: jay-zhuang
Differential Revision: D25592970
Pulled By: pdillinger
fbshipit-source-id: 606fdaa025bb790d7e9c21601e8ea86e10541912
4 years ago
|
|
|
// Round down to a number of slots supported by this structure. Note that
|
|
|
|
// this needs to be must be taken into account for the banding if this
|
|
|
|
// solution layout/storage is to be used.
|
|
|
|
static Index RoundDownNumSlots(Index num_slots) {
|
|
|
|
// Must be multiple of kCoeffBits
|
|
|
|
Index corrected = num_slots / kCoeffBits * kCoeffBits;
|
|
|
|
|
|
|
|
// Do not use num_starts==1 unless kUseSmash, because the hashing
|
|
|
|
// might not be equipped for stacking up so many entries on a
|
|
|
|
// single start location.
|
|
|
|
if (!TypesAndSettings::kUseSmash && corrected == kCoeffBits) {
|
|
|
|
corrected = 0;
|
|
|
|
}
|
|
|
|
return corrected;
|
|
|
|
}
|
|
|
|
|
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
|
|
|
// Compute the number of bytes for a given number of slots and desired
|
|
|
|
// FP rate. Since desired FP rate might not be exactly achievable,
|
|
|
|
// rounding_bias32==0 means to always round toward lower FP rate
|
|
|
|
// than desired (more bytes); rounding_bias32==max uint32_t means always
|
|
|
|
// round toward higher FP rate than desired (fewer bytes); other values
|
|
|
|
// act as a proportional threshold or bias between the two.
|
|
|
|
static size_t GetBytesForFpRate(Index num_slots, double desired_fp_rate,
|
|
|
|
uint32_t rounding_bias32) {
|
|
|
|
return InternalGetBytesForFpRate(num_slots, desired_fp_rate,
|
|
|
|
1.0 / desired_fp_rate, rounding_bias32);
|
|
|
|
}
|
|
|
|
|
|
|
|
// The same, but specifying desired accuracy as 1.0 / FP rate, or
|
|
|
|
// one_in_fp_rate. E.g. desired_one_in_fp_rate=100 means 1% FP rate.
|
|
|
|
static size_t GetBytesForOneInFpRate(Index num_slots,
|
|
|
|
double desired_one_in_fp_rate,
|
|
|
|
uint32_t rounding_bias32) {
|
|
|
|
return InternalGetBytesForFpRate(num_slots, 1.0 / desired_one_in_fp_rate,
|
|
|
|
desired_one_in_fp_rate, rounding_bias32);
|
|
|
|
}
|
|
|
|
|
|
|
|
protected:
|
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
|
|
|
static size_t InternalGetBytesForFpRate(Index num_slots,
|
|
|
|
double desired_fp_rate,
|
|
|
|
double desired_one_in_fp_rate,
|
|
|
|
uint32_t rounding_bias32) {
|
|
|
|
assert(TypesAndSettings::kIsFilter);
|
Support optimize_filters_for_memory for Ribbon filter (#7774)
Summary:
Primarily this change refactors the optimize_filters_for_memory
code for Bloom filters, based on malloc_usable_size, to also work for
Ribbon filters.
This change also replaces the somewhat slow but general
BuiltinFilterBitsBuilder::ApproximateNumEntries with
implementation-specific versions for Ribbon (new) and Legacy Bloom
(based on a recently deleted version). The reason is to emphasize
speed in ApproximateNumEntries rather than 100% accuracy.
Justification: ApproximateNumEntries (formerly CalculateNumEntry) is
only used by RocksDB for range-partitioned filters, called each time we
start to construct one. (In theory, it should be possible to reuse the
estimate, but the abstractions provided by FilterPolicy don't really
make that workable.) But this is only used as a heuristic estimate for
hitting a desired partitioned filter size because of alignment to data
blocks, which have various numbers of unique keys or prefixes. The two
factors lead us to prioritize reasonable speed over 100% accuracy.
optimize_filters_for_memory adds extra complication, because precisely
calculating num_entries for some allowed number of bytes depends on state
with optimize_filters_for_memory enabled. And the allocator-agnostic
implementation of optimize_filters_for_memory, using malloc_usable_size,
means we would have to actually allocate memory, many times, just to
precisely determine how many entries (keys) could be added and stay below
some size budget, for the current state. (In a draft, I got this
working, and then realized the balance of speed vs. accuracy was all
wrong.)
So related to that, I have made CalculateSpace, an internal-only API
only used for testing, non-authoritative also if
optimize_filters_for_memory is enabled. This simplifies some code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7774
Test Plan:
unit test updated, and for FilterSize test, range of tested
values is greatly expanded (still super fast)
Also tested `db_bench -benchmarks=fillrandom,stats -bloom_bits=10 -num=1000000 -partition_index_and_filters -format_version=5 [-optimize_filters_for_memory] [-use_ribbon_filter]` with temporary debug output of generated filter sizes.
Bloom+optimize_filters_for_memory:
1 Filter size: 197 (224 in memory)
134 Filter size: 3525 (3584 in memory)
107 Filter size: 4037 (4096 in memory)
Total on disk: 904,506
Total in memory: 918,752
Ribbon+optimize_filters_for_memory:
1 Filter size: 3061 (3072 in memory)
110 Filter size: 3573 (3584 in memory)
58 Filter size: 4085 (4096 in memory)
Total on disk: 633,021 (-30.0%)
Total in memory: 634,880 (-30.9%)
Bloom (no offm):
1 Filter size: 261 (320 in memory)
1 Filter size: 3333 (3584 in memory)
240 Filter size: 3717 (4096 in memory)
Total on disk: 895,674 (-1% on disk vs. +offm; known tolerable overhead of offm)
Total in memory: 986,944 (+7.4% vs. +offm)
Ribbon (no offm):
1 Filter size: 2949 (3072 in memory)
1 Filter size: 3381 (3584 in memory)
167 Filter size: 3701 (4096 in memory)
Total on disk: 624,397 (-30.3% vs. Bloom)
Total in memory: 690,688 (-30.0% vs. Bloom)
Note that optimize_filters_for_memory is even more effective for Ribbon filter than for cache-local Bloom, because it can close the unused memory gap even tighter than Bloom filter, because of 16 byte increments for Ribbon vs. 64 byte increments for Bloom.
Reviewed By: jay-zhuang
Differential Revision: D25592970
Pulled By: pdillinger
fbshipit-source-id: 606fdaa025bb790d7e9c21601e8ea86e10541912
4 years ago
|
|
|
if (TypesAndSettings::kAllowZeroStarts) {
|
|
|
|
if (num_slots == 0) {
|
|
|
|
// Unusual. Zero starts presumes no keys added -> always false (no FPs)
|
|
|
|
return 0U;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
assert(num_slots > 0);
|
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
|
|
|
}
|
|
|
|
// Must be rounded up already.
|
|
|
|
assert(RoundUpNumSlots(num_slots) == num_slots);
|
|
|
|
|
|
|
|
if (desired_one_in_fp_rate > 1.0 && desired_fp_rate < 1.0) {
|
|
|
|
// Typical: less than 100% FP rate
|
|
|
|
if (desired_one_in_fp_rate <= static_cast<ResultRow>(-1)) {
|
|
|
|
// Typical: Less than maximum result row entropy
|
|
|
|
ResultRow rounded = static_cast<ResultRow>(desired_one_in_fp_rate);
|
|
|
|
int lower_columns = FloorLog2(rounded);
|
|
|
|
double lower_columns_fp_rate = std::pow(2.0, -lower_columns);
|
|
|
|
double upper_columns_fp_rate = std::pow(2.0, -(lower_columns + 1));
|
|
|
|
// Floating point don't let me down!
|
|
|
|
assert(lower_columns_fp_rate >= desired_fp_rate);
|
|
|
|
assert(upper_columns_fp_rate <= desired_fp_rate);
|
|
|
|
|
|
|
|
double lower_portion = (desired_fp_rate - upper_columns_fp_rate) /
|
|
|
|
(lower_columns_fp_rate - upper_columns_fp_rate);
|
|
|
|
// Floating point don't let me down!
|
|
|
|
assert(lower_portion >= 0.0);
|
|
|
|
assert(lower_portion <= 1.0);
|
|
|
|
|
|
|
|
double rounding_bias = (rounding_bias32 + 0.5) / double{0x100000000};
|
|
|
|
assert(rounding_bias > 0.0);
|
|
|
|
assert(rounding_bias < 1.0);
|
|
|
|
|
|
|
|
// Note: Ignoring smash setting; still close enough in that case
|
|
|
|
Index num_starts = num_slots - kCoeffBits + 1;
|
|
|
|
// Lower upper_start_block means lower FP rate (higher accuracy)
|
|
|
|
Index upper_start_block = static_cast<Index>(
|
|
|
|
(lower_portion * num_starts + rounding_bias) / kCoeffBits);
|
|
|
|
Index num_blocks = num_slots / kCoeffBits;
|
|
|
|
assert(upper_start_block < num_blocks);
|
|
|
|
|
|
|
|
// Start by assuming all blocks use lower number of columns
|
|
|
|
Index num_segments = num_blocks * static_cast<Index>(lower_columns);
|
|
|
|
// Correct by 1 each for blocks using upper number of columns
|
|
|
|
num_segments += (num_blocks - upper_start_block);
|
|
|
|
// Total bytes
|
|
|
|
return num_segments * sizeof(CoeffRow);
|
|
|
|
} else {
|
|
|
|
// one_in_fp_rate too big, thus requested FP rate is smaller than
|
|
|
|
// supported. Use max number of columns for minimum supported FP rate.
|
|
|
|
return num_slots * sizeof(ResultRow);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// Effectively asking for 100% FP rate, or NaN etc.
|
|
|
|
if (TypesAndSettings::kAllowZeroStarts) {
|
|
|
|
// Zero segments
|
|
|
|
return 0U;
|
|
|
|
} else {
|
|
|
|
// One segment (minimum size, maximizing FP rate)
|
|
|
|
return sizeof(CoeffRow);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void InternalConfigure() {
|
|
|
|
const Index num_blocks = GetNumBlocks();
|
|
|
|
Index num_segments = GetNumSegments();
|
|
|
|
|
|
|
|
if (num_blocks == 0) {
|
|
|
|
// Exceptional
|
|
|
|
upper_num_columns_ = 0;
|
|
|
|
upper_start_block_ = 0;
|
|
|
|
} else {
|
|
|
|
// Normal
|
|
|
|
upper_num_columns_ =
|
|
|
|
(num_segments + /*round up*/ num_blocks - 1) / num_blocks;
|
|
|
|
upper_start_block_ = upper_num_columns_ * num_blocks - num_segments;
|
|
|
|
// Unless that's more columns than supported by ResultRow data type
|
|
|
|
if (upper_num_columns_ > 8U * sizeof(ResultRow)) {
|
|
|
|
// Use maximum columns (there will be space unused)
|
|
|
|
upper_num_columns_ = static_cast<Index>(8U * sizeof(ResultRow));
|
|
|
|
upper_start_block_ = 0;
|
|
|
|
num_segments = num_blocks * upper_num_columns_;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// Update data_len_ for correct rounding and/or unused space
|
|
|
|
// NOTE: unused space stays gone if we PrepareForNumStarts again.
|
|
|
|
// We are prioritizing minimizing the number of fields over making
|
|
|
|
// the "unusued space" feature work well.
|
|
|
|
data_len_ = num_segments * sizeof(CoeffRow);
|
|
|
|
}
|
|
|
|
|
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
|
|
|
char* const data_;
|
|
|
|
size_t data_len_;
|
|
|
|
Index num_starts_ = 0;
|
|
|
|
Index upper_num_columns_ = 0;
|
|
|
|
Index upper_start_block_ = 0;
|
|
|
|
};
|
|
|
|
|
|
|
|
} // namespace ribbon
|
|
|
|
|
|
|
|
} // namespace ROCKSDB_NAMESPACE
|
|
|
|
|
|
|
|
// For convenience working with templates
|
|
|
|
#define IMPORT_RIBBON_IMPL_TYPES(TypesAndSettings) \
|
|
|
|
using Hasher = ROCKSDB_NAMESPACE::ribbon::StandardHasher<TypesAndSettings>; \
|
|
|
|
using Banding = \
|
|
|
|
ROCKSDB_NAMESPACE::ribbon::StandardBanding<TypesAndSettings>; \
|
|
|
|
using SimpleSoln = \
|
|
|
|
ROCKSDB_NAMESPACE::ribbon::InMemSimpleSolution<TypesAndSettings>; \
|
|
|
|
using InterleavedSoln = \
|
|
|
|
ROCKSDB_NAMESPACE::ribbon::SerializableInterleavedSolution< \
|
|
|
|
TypesAndSettings>; \
|
|
|
|
static_assert(sizeof(Hasher) + sizeof(Banding) + sizeof(SimpleSoln) + \
|
|
|
|
sizeof(InterleavedSoln) > \
|
|
|
|
0, \
|
|
|
|
"avoid unused warnings, semicolon expected after macro call")
|