|
|
|
// 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).
|
|
|
|
|
|
|
|
#include "rocksdb/system_clock.h"
|
|
|
|
#include "test_util/testharness.h"
|
|
|
|
#include "util/bloom_impl.h"
|
|
|
|
#include "util/coding.h"
|
|
|
|
#include "util/hash.h"
|
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/ribbon_config.h"
|
|
|
|
#include "util/ribbon_impl.h"
|
|
|
|
#include "util/stop_watch.h"
|
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/string_util.h"
|
|
|
|
|
|
|
|
#ifndef GFLAGS
|
|
|
|
uint32_t FLAGS_thoroughness = 5;
|
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
|
|
|
uint32_t FLAGS_max_add = 0;
|
|
|
|
uint32_t FLAGS_min_check = 4000;
|
|
|
|
uint32_t FLAGS_max_check = 100000;
|
|
|
|
bool FLAGS_verbose = false;
|
|
|
|
|
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
|
|
|
bool FLAGS_find_occ = false;
|
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 FLAGS_find_slot_occ = false;
|
|
|
|
double FLAGS_find_next_factor = 1.618;
|
|
|
|
uint32_t FLAGS_find_iters = 10000;
|
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
|
|
|
uint32_t FLAGS_find_min_slots = 128;
|
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
|
|
|
uint32_t FLAGS_find_max_slots = 1000000;
|
|
|
|
|
|
|
|
bool FLAGS_optimize_homog = false;
|
|
|
|
uint32_t FLAGS_optimize_homog_slots = 30000000;
|
|
|
|
uint32_t FLAGS_optimize_homog_check = 200000;
|
|
|
|
double FLAGS_optimize_homog_granularity = 0.002;
|
|
|
|
#else
|
|
|
|
#include "util/gflags_compat.h"
|
|
|
|
using GFLAGS_NAMESPACE::ParseCommandLineFlags;
|
|
|
|
// Using 500 is a good test when you have time to be thorough.
|
|
|
|
// Default is for general RocksDB regression test runs.
|
|
|
|
DEFINE_uint32(thoroughness, 5, "iterations per configuration");
|
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
|
|
|
DEFINE_uint32(max_add, 0,
|
|
|
|
"Add up to this number of entries to a single filter in "
|
|
|
|
"CompactnessAndBacktrackAndFpRate; 0 == reasonable default");
|
|
|
|
DEFINE_uint32(min_check, 4000,
|
|
|
|
"Minimum number of novel entries for testing FP rate");
|
|
|
|
DEFINE_uint32(max_check, 10000,
|
|
|
|
"Maximum number of novel entries for testing FP rate");
|
|
|
|
DEFINE_bool(verbose, false, "Print extra details");
|
|
|
|
|
|
|
|
// Options for FindOccupancy, which is more of a tool than a test.
|
|
|
|
DEFINE_bool(find_occ, false, "whether to run the FindOccupancy tool");
|
|
|
|
DEFINE_bool(find_slot_occ, false,
|
|
|
|
"whether to show individual slot occupancies with "
|
|
|
|
"FindOccupancy tool");
|
|
|
|
DEFINE_double(find_next_factor, 1.618,
|
|
|
|
"factor to next num_slots for FindOccupancy");
|
|
|
|
DEFINE_uint32(find_iters, 10000, "number of samples for FindOccupancy");
|
|
|
|
DEFINE_uint32(find_min_slots, 128, "number of slots for FindOccupancy");
|
|
|
|
DEFINE_uint32(find_max_slots, 1000000, "number of slots for FindOccupancy");
|
|
|
|
|
|
|
|
// Options for OptimizeHomogAtScale, which is more of a tool than a test.
|
|
|
|
DEFINE_bool(optimize_homog, false,
|
|
|
|
"whether to run the OptimizeHomogAtScale tool");
|
|
|
|
DEFINE_uint32(optimize_homog_slots, 30000000,
|
|
|
|
"number of slots for OptimizeHomogAtScale");
|
|
|
|
DEFINE_uint32(optimize_homog_check, 200000,
|
|
|
|
"number of queries for checking FP rate in OptimizeHomogAtScale");
|
|
|
|
DEFINE_double(
|
|
|
|
optimize_homog_granularity, 0.002,
|
|
|
|
"overhead change between FP rate checking in OptimizeHomogAtScale");
|
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
|
|
|
|
|
|
|
#endif // GFLAGS
|
|
|
|
|
|
|
|
template <typename TypesAndSettings>
|
|
|
|
class RibbonTypeParamTest : public ::testing::Test {};
|
|
|
|
|
|
|
|
class RibbonTest : public ::testing::Test {};
|
|
|
|
|
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
|
|
|
namespace {
|
|
|
|
|
|
|
|
// Different ways of generating keys for testing
|
|
|
|
|
|
|
|
// Generate semi-sequential keys
|
|
|
|
struct StandardKeyGen {
|
|
|
|
StandardKeyGen(const std::string& prefix, uint64_t id)
|
|
|
|
: id_(id), str_(prefix) {
|
|
|
|
ROCKSDB_NAMESPACE::PutFixed64(&str_, /*placeholder*/ 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Prefix (only one required)
|
|
|
|
StandardKeyGen& operator++() {
|
|
|
|
++id_;
|
|
|
|
return *this;
|
|
|
|
}
|
|
|
|
|
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
|
|
|
StandardKeyGen& operator+=(uint64_t i) {
|
|
|
|
id_ += i;
|
|
|
|
return *this;
|
|
|
|
}
|
|
|
|
|
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
|
|
|
const std::string& operator*() {
|
|
|
|
// Use multiplication to mix things up a little in the key
|
|
|
|
ROCKSDB_NAMESPACE::EncodeFixed64(&str_[str_.size() - 8],
|
|
|
|
id_ * uint64_t{0x1500000001});
|
|
|
|
return str_;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool operator==(const StandardKeyGen& other) {
|
|
|
|
// Same prefix is assumed
|
|
|
|
return id_ == other.id_;
|
|
|
|
}
|
|
|
|
bool operator!=(const StandardKeyGen& other) {
|
|
|
|
// Same prefix is assumed
|
|
|
|
return id_ != other.id_;
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t id_;
|
|
|
|
std::string str_;
|
|
|
|
};
|
|
|
|
|
|
|
|
// Generate small sequential keys, that can misbehave with sequential seeds
|
|
|
|
// as in https://github.com/Cyan4973/xxHash/issues/469.
|
|
|
|
// These keys are only heuristically unique, but that's OK with 64 bits,
|
|
|
|
// for testing purposes.
|
|
|
|
struct SmallKeyGen {
|
|
|
|
SmallKeyGen(const std::string& prefix, uint64_t id) : id_(id) {
|
|
|
|
// Hash the prefix for a heuristically unique offset
|
|
|
|
id_ += ROCKSDB_NAMESPACE::GetSliceHash64(prefix);
|
|
|
|
ROCKSDB_NAMESPACE::PutFixed64(&str_, id_);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Prefix (only one required)
|
|
|
|
SmallKeyGen& operator++() {
|
|
|
|
++id_;
|
|
|
|
return *this;
|
|
|
|
}
|
|
|
|
|
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
|
|
|
SmallKeyGen& operator+=(uint64_t i) {
|
|
|
|
id_ += i;
|
|
|
|
return *this;
|
|
|
|
}
|
|
|
|
|
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
|
|
|
const std::string& operator*() {
|
|
|
|
ROCKSDB_NAMESPACE::EncodeFixed64(&str_[str_.size() - 8], id_);
|
|
|
|
return str_;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool operator==(const SmallKeyGen& other) { return id_ == other.id_; }
|
|
|
|
bool operator!=(const SmallKeyGen& other) { return id_ != other.id_; }
|
|
|
|
|
|
|
|
uint64_t id_;
|
|
|
|
std::string str_;
|
|
|
|
};
|
|
|
|
|
|
|
|
template <typename KeyGen>
|
|
|
|
struct Hash32KeyGenWrapper : public KeyGen {
|
|
|
|
Hash32KeyGenWrapper(const std::string& prefix, uint64_t id)
|
|
|
|
: KeyGen(prefix, id) {}
|
|
|
|
uint32_t operator*() {
|
|
|
|
auto& key = *static_cast<KeyGen&>(*this);
|
|
|
|
// unseeded
|
|
|
|
return ROCKSDB_NAMESPACE::GetSliceHash(key);
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
template <typename KeyGen>
|
|
|
|
struct Hash64KeyGenWrapper : public KeyGen {
|
|
|
|
Hash64KeyGenWrapper(const std::string& prefix, uint64_t id)
|
|
|
|
: KeyGen(prefix, id) {}
|
|
|
|
uint64_t operator*() {
|
|
|
|
auto& key = *static_cast<KeyGen&>(*this);
|
|
|
|
// unseeded
|
|
|
|
return ROCKSDB_NAMESPACE::GetSliceHash64(key);
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
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
|
|
|
using ROCKSDB_NAMESPACE::ribbon::ConstructionFailureChance;
|
|
|
|
|
|
|
|
const std::vector<ConstructionFailureChance> kFailureOnly50Pct = {
|
|
|
|
ROCKSDB_NAMESPACE::ribbon::kOneIn2};
|
|
|
|
|
|
|
|
const std::vector<ConstructionFailureChance> kFailureOnlyRare = {
|
|
|
|
ROCKSDB_NAMESPACE::ribbon::kOneIn1000};
|
|
|
|
|
|
|
|
const std::vector<ConstructionFailureChance> kFailureAll = {
|
|
|
|
ROCKSDB_NAMESPACE::ribbon::kOneIn2, ROCKSDB_NAMESPACE::ribbon::kOneIn20,
|
|
|
|
ROCKSDB_NAMESPACE::ribbon::kOneIn1000};
|
|
|
|
|
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
|
|
|
} // namespace
|
|
|
|
|
|
|
|
using ROCKSDB_NAMESPACE::ribbon::ExpectedCollisionFpRate;
|
|
|
|
using ROCKSDB_NAMESPACE::ribbon::StandardHasher;
|
|
|
|
using ROCKSDB_NAMESPACE::ribbon::StandardRehasherAdapter;
|
|
|
|
|
|
|
|
struct DefaultTypesAndSettings {
|
|
|
|
using CoeffRow = ROCKSDB_NAMESPACE::Unsigned128;
|
|
|
|
using ResultRow = uint8_t;
|
|
|
|
using Index = uint32_t;
|
|
|
|
using Hash = uint64_t;
|
|
|
|
using Seed = uint32_t;
|
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
|
|
|
using Key = ROCKSDB_NAMESPACE::Slice;
|
|
|
|
static constexpr bool kIsFilter = true;
|
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 bool kHomogeneous = false;
|
|
|
|
static constexpr bool kFirstCoeffAlwaysOne = true;
|
|
|
|
static constexpr bool kUseSmash = false;
|
|
|
|
static constexpr bool kAllowZeroStarts = false;
|
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& key, uint64_t raw_seed) {
|
|
|
|
// This version 0.7.2 preview of XXH3 (a.k.a. XXH3p) function does
|
|
|
|
// not pass SmallKeyGen tests below without some seed premixing from
|
|
|
|
// StandardHasher. See https://github.com/Cyan4973/xxHash/issues/469
|
|
|
|
return ROCKSDB_NAMESPACE::Hash64(key.data(), key.size(), raw_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
|
|
|
// For testing
|
|
|
|
using KeyGen = StandardKeyGen;
|
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 const std::vector<ConstructionFailureChance>& FailureChanceToTest() {
|
|
|
|
return kFailureAll;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
using TypesAndSettings_Coeff128 = DefaultTypesAndSettings;
|
|
|
|
struct TypesAndSettings_Coeff128Smash : public DefaultTypesAndSettings {
|
|
|
|
static constexpr bool kUseSmash = true;
|
|
|
|
};
|
|
|
|
struct TypesAndSettings_Coeff64 : public DefaultTypesAndSettings {
|
|
|
|
using CoeffRow = uint64_t;
|
|
|
|
};
|
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
|
|
|
struct TypesAndSettings_Coeff64Smash : public TypesAndSettings_Coeff64 {
|
|
|
|
static constexpr bool kUseSmash = true;
|
|
|
|
};
|
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
|
|
|
struct TypesAndSettings_Coeff64Smash0 : public TypesAndSettings_Coeff64Smash {
|
|
|
|
static constexpr bool kFirstCoeffAlwaysOne = false;
|
|
|
|
};
|
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
|
|
|
|
|
|
|
// Homogeneous Ribbon configurations
|
|
|
|
struct TypesAndSettings_Coeff128_Homog : public DefaultTypesAndSettings {
|
|
|
|
static constexpr bool kHomogeneous = true;
|
|
|
|
// Since our best construction success setting still has 1/1000 failure
|
|
|
|
// rate, the best FP rate we test is 1/256
|
|
|
|
using ResultRow = uint8_t;
|
|
|
|
// Homogeneous only makes sense with sufficient slots for equivalent of
|
|
|
|
// almost sure construction success
|
|
|
|
static const std::vector<ConstructionFailureChance>& FailureChanceToTest() {
|
|
|
|
return kFailureOnlyRare;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
struct TypesAndSettings_Coeff128Smash_Homog
|
|
|
|
: public TypesAndSettings_Coeff128_Homog {
|
|
|
|
// Smash (extra time to save space) + Homog (extra space to save time)
|
|
|
|
// doesn't make much sense in practice, but we minimally test it
|
|
|
|
static constexpr bool kUseSmash = true;
|
|
|
|
};
|
|
|
|
struct TypesAndSettings_Coeff64_Homog : public TypesAndSettings_Coeff128_Homog {
|
|
|
|
using CoeffRow = uint64_t;
|
|
|
|
};
|
|
|
|
struct TypesAndSettings_Coeff64Smash_Homog
|
|
|
|
: public TypesAndSettings_Coeff64_Homog {
|
|
|
|
// Smash (extra time to save space) + Homog (extra space to save time)
|
|
|
|
// doesn't make much sense in practice, but we minimally test it
|
|
|
|
static constexpr bool kUseSmash = true;
|
|
|
|
};
|
|
|
|
|
|
|
|
// Less exhaustive mix of coverage, but still covering the most stressful case
|
|
|
|
// (only 50% construction success)
|
|
|
|
struct AbridgedTypesAndSettings : public DefaultTypesAndSettings {
|
|
|
|
static const std::vector<ConstructionFailureChance>& FailureChanceToTest() {
|
|
|
|
return kFailureOnly50Pct;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
struct TypesAndSettings_Result16 : public AbridgedTypesAndSettings {
|
|
|
|
using ResultRow = uint16_t;
|
|
|
|
};
|
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
|
|
|
struct TypesAndSettings_Result32 : public AbridgedTypesAndSettings {
|
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
|
|
|
using ResultRow = uint32_t;
|
|
|
|
};
|
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
|
|
|
struct TypesAndSettings_IndexSizeT : public AbridgedTypesAndSettings {
|
|
|
|
using Index = size_t;
|
|
|
|
};
|
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
|
|
|
struct TypesAndSettings_Hash32 : public AbridgedTypesAndSettings {
|
|
|
|
using Hash = uint32_t;
|
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& key, Hash raw_seed) {
|
|
|
|
// This MurmurHash1 function does not pass tests below without the
|
|
|
|
// seed premixing from StandardHasher. In fact, it needs more than
|
|
|
|
// just a multiplication mixer on the ordinal seed.
|
|
|
|
return ROCKSDB_NAMESPACE::Hash(key.data(), key.size(), raw_seed);
|
|
|
|
}
|
|
|
|
};
|
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
|
|
|
struct TypesAndSettings_Hash32_Result16 : public AbridgedTypesAndSettings {
|
|
|
|
using ResultRow = uint16_t;
|
|
|
|
};
|
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
|
|
|
struct TypesAndSettings_KeyString : public AbridgedTypesAndSettings {
|
|
|
|
using Key = std::string;
|
|
|
|
};
|
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
|
|
|
struct TypesAndSettings_Seed8 : public AbridgedTypesAndSettings {
|
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 not a generally recommended configuration. With the configured
|
|
|
|
// hash function, it would fail with SmallKeyGen due to insufficient
|
|
|
|
// independence among the seeds.
|
|
|
|
using Seed = uint8_t;
|
|
|
|
};
|
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
|
|
|
struct TypesAndSettings_NoAlwaysOne : public AbridgedTypesAndSettings {
|
|
|
|
static constexpr bool kFirstCoeffAlwaysOne = false;
|
|
|
|
};
|
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
|
|
|
struct TypesAndSettings_AllowZeroStarts : public AbridgedTypesAndSettings {
|
|
|
|
static constexpr bool kAllowZeroStarts = true;
|
|
|
|
};
|
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
|
|
|
struct TypesAndSettings_Seed64 : public AbridgedTypesAndSettings {
|
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
|
|
|
using Seed = uint64_t;
|
|
|
|
};
|
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
|
|
|
struct TypesAndSettings_Rehasher
|
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
|
|
|
: public StandardRehasherAdapter<AbridgedTypesAndSettings> {
|
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
|
|
|
using KeyGen = Hash64KeyGenWrapper<StandardKeyGen>;
|
|
|
|
};
|
|
|
|
struct TypesAndSettings_Rehasher_Result16 : public TypesAndSettings_Rehasher {
|
|
|
|
using ResultRow = uint16_t;
|
|
|
|
};
|
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
|
|
|
struct TypesAndSettings_Rehasher_Result32 : public TypesAndSettings_Rehasher {
|
|
|
|
using ResultRow = uint32_t;
|
|
|
|
};
|
|
|
|
struct TypesAndSettings_Rehasher_Seed64
|
|
|
|
: public StandardRehasherAdapter<TypesAndSettings_Seed64> {
|
|
|
|
using KeyGen = Hash64KeyGenWrapper<StandardKeyGen>;
|
|
|
|
// Note: 64-bit seed with Rehasher gives slightly better average reseeds
|
|
|
|
};
|
|
|
|
struct TypesAndSettings_Rehasher32
|
|
|
|
: public StandardRehasherAdapter<TypesAndSettings_Hash32> {
|
|
|
|
using KeyGen = Hash32KeyGenWrapper<StandardKeyGen>;
|
|
|
|
};
|
|
|
|
struct TypesAndSettings_Rehasher32_Coeff64
|
|
|
|
: public TypesAndSettings_Rehasher32 {
|
|
|
|
using CoeffRow = uint64_t;
|
|
|
|
};
|
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
|
|
|
struct TypesAndSettings_SmallKeyGen : public AbridgedTypesAndSettings {
|
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
|
|
|
// SmallKeyGen stresses the independence of different hash seeds
|
|
|
|
using KeyGen = SmallKeyGen;
|
|
|
|
};
|
|
|
|
struct TypesAndSettings_Hash32_SmallKeyGen : public TypesAndSettings_Hash32 {
|
|
|
|
// SmallKeyGen stresses the independence of different hash seeds
|
|
|
|
using KeyGen = SmallKeyGen;
|
|
|
|
};
|
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
|
|
|
struct TypesAndSettings_Coeff32 : public DefaultTypesAndSettings {
|
|
|
|
using CoeffRow = uint32_t;
|
|
|
|
};
|
|
|
|
struct TypesAndSettings_Coeff32Smash : public TypesAndSettings_Coeff32 {
|
|
|
|
static constexpr bool kUseSmash = true;
|
|
|
|
};
|
|
|
|
struct TypesAndSettings_Coeff16 : public DefaultTypesAndSettings {
|
|
|
|
using CoeffRow = uint16_t;
|
|
|
|
};
|
|
|
|
struct TypesAndSettings_Coeff16Smash : public TypesAndSettings_Coeff16 {
|
|
|
|
static constexpr bool kUseSmash = true;
|
|
|
|
};
|
|
|
|
|
|
|
|
using TestTypesAndSettings = ::testing::Types<
|
|
|
|
TypesAndSettings_Coeff128, TypesAndSettings_Coeff128Smash,
|
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
|
|
|
TypesAndSettings_Coeff64, TypesAndSettings_Coeff64Smash,
|
|
|
|
TypesAndSettings_Coeff64Smash0, TypesAndSettings_Coeff128_Homog,
|
|
|
|
TypesAndSettings_Coeff128Smash_Homog, TypesAndSettings_Coeff64_Homog,
|
|
|
|
TypesAndSettings_Coeff64Smash_Homog, TypesAndSettings_Result16,
|
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
|
|
|
TypesAndSettings_Result32, TypesAndSettings_IndexSizeT,
|
|
|
|
TypesAndSettings_Hash32, TypesAndSettings_Hash32_Result16,
|
|
|
|
TypesAndSettings_KeyString, TypesAndSettings_Seed8,
|
|
|
|
TypesAndSettings_NoAlwaysOne, TypesAndSettings_AllowZeroStarts,
|
|
|
|
TypesAndSettings_Seed64, TypesAndSettings_Rehasher,
|
|
|
|
TypesAndSettings_Rehasher_Result16, TypesAndSettings_Rehasher_Result32,
|
|
|
|
TypesAndSettings_Rehasher_Seed64, TypesAndSettings_Rehasher32,
|
|
|
|
TypesAndSettings_Rehasher32_Coeff64, TypesAndSettings_SmallKeyGen,
|
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
|
|
|
TypesAndSettings_Hash32_SmallKeyGen, TypesAndSettings_Coeff32,
|
|
|
|
TypesAndSettings_Coeff32Smash, TypesAndSettings_Coeff16,
|
|
|
|
TypesAndSettings_Coeff16Smash>;
|
|
|
|
TYPED_TEST_CASE(RibbonTypeParamTest, TestTypesAndSettings);
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
|
|
|
|
// For testing Poisson-distributed (or similar) statistics, get value for
|
|
|
|
// `stddevs_allowed` standard deviations above expected mean
|
|
|
|
// `expected_count`.
|
|
|
|
// (Poisson approximates Binomial only if probability of a trial being
|
|
|
|
// in the count is low.)
|
|
|
|
uint64_t PoissonUpperBound(double expected_count, double stddevs_allowed) {
|
|
|
|
return static_cast<uint64_t>(
|
|
|
|
expected_count + stddevs_allowed * std::sqrt(expected_count) + 1.0);
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t PoissonLowerBound(double expected_count, double stddevs_allowed) {
|
|
|
|
return static_cast<uint64_t>(std::max(
|
|
|
|
0.0, expected_count - stddevs_allowed * std::sqrt(expected_count)));
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t FrequentPoissonUpperBound(double expected_count) {
|
|
|
|
// Allow up to 5.0 standard deviations for frequently checked statistics
|
|
|
|
return PoissonUpperBound(expected_count, 5.0);
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t FrequentPoissonLowerBound(double expected_count) {
|
|
|
|
return PoissonLowerBound(expected_count, 5.0);
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t InfrequentPoissonUpperBound(double expected_count) {
|
|
|
|
// Allow up to 3 standard deviations for infrequently checked statistics
|
|
|
|
return PoissonUpperBound(expected_count, 3.0);
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t InfrequentPoissonLowerBound(double expected_count) {
|
|
|
|
return PoissonLowerBound(expected_count, 3.0);
|
|
|
|
}
|
|
|
|
|
|
|
|
} // namespace
|
|
|
|
|
|
|
|
TYPED_TEST(RibbonTypeParamTest, CompactnessAndBacktrackAndFpRate) {
|
|
|
|
IMPORT_RIBBON_TYPES_AND_SETTINGS(TypeParam);
|
|
|
|
IMPORT_RIBBON_IMPL_TYPES(TypeParam);
|
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
|
|
|
using KeyGen = typename TypeParam::KeyGen;
|
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
|
|
|
using ConfigHelper =
|
|
|
|
ROCKSDB_NAMESPACE::ribbon::BandingConfigHelper<TypeParam>;
|
|
|
|
|
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 (sizeof(CoeffRow) < 8) {
|
|
|
|
ROCKSDB_GTEST_SKIP("Not fully supported");
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
const auto log2_thoroughness =
|
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_cast<uint32_t>(ROCKSDB_NAMESPACE::FloorLog2(FLAGS_thoroughness));
|
|
|
|
|
|
|
|
// We are going to choose num_to_add using an exponential distribution,
|
|
|
|
// so that we have good representation of small-to-medium filters.
|
|
|
|
// Here we just pick some reasonable, practical upper bound based on
|
|
|
|
// kCoeffBits or option.
|
|
|
|
const double log_max_add = std::log(
|
|
|
|
FLAGS_max_add > 0 ? FLAGS_max_add
|
|
|
|
: static_cast<uint32_t>(kCoeffBits * kCoeffBits) *
|
|
|
|
std::max(FLAGS_thoroughness, uint32_t{32}));
|
|
|
|
|
|
|
|
// This needs to be enough below the minimum number of slots to get a
|
|
|
|
// reasonable number of samples with the minimum number of slots.
|
|
|
|
const double log_min_add = std::log(0.66 * SimpleSoln::RoundUpNumSlots(1));
|
|
|
|
|
|
|
|
ASSERT_GT(log_max_add, log_min_add);
|
|
|
|
|
|
|
|
const double diff_log_add = log_max_add - log_min_add;
|
|
|
|
|
|
|
|
for (ConstructionFailureChance cs : TypeParam::FailureChanceToTest()) {
|
|
|
|
double expected_reseeds;
|
|
|
|
switch (cs) {
|
|
|
|
default:
|
|
|
|
assert(false);
|
|
|
|
FALLTHROUGH_INTENDED;
|
|
|
|
case ROCKSDB_NAMESPACE::ribbon::kOneIn2:
|
|
|
|
fprintf(stderr, "== Failure: 50 percent\n");
|
|
|
|
expected_reseeds = 1.0;
|
|
|
|
break;
|
|
|
|
case ROCKSDB_NAMESPACE::ribbon::kOneIn20:
|
|
|
|
fprintf(stderr, "== Failure: 95 percent\n");
|
|
|
|
expected_reseeds = 0.053;
|
|
|
|
break;
|
|
|
|
case ROCKSDB_NAMESPACE::ribbon::kOneIn1000:
|
|
|
|
fprintf(stderr, "== Failure: 1/1000\n");
|
|
|
|
expected_reseeds = 0.001;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
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
|
|
|
uint64_t total_reseeds = 0;
|
|
|
|
uint64_t total_singles = 0;
|
|
|
|
uint64_t total_single_failures = 0;
|
|
|
|
uint64_t total_batch = 0;
|
|
|
|
uint64_t total_batch_successes = 0;
|
|
|
|
uint64_t total_fp_count = 0;
|
|
|
|
uint64_t total_added = 0;
|
|
|
|
uint64_t total_expand_trials = 0;
|
|
|
|
uint64_t total_expand_failures = 0;
|
|
|
|
double total_expand_overhead = 0.0;
|
|
|
|
|
|
|
|
uint64_t soln_query_nanos = 0;
|
|
|
|
uint64_t soln_query_count = 0;
|
|
|
|
uint64_t bloom_query_nanos = 0;
|
|
|
|
uint64_t isoln_query_nanos = 0;
|
|
|
|
uint64_t isoln_query_count = 0;
|
|
|
|
|
|
|
|
// Take different samples if you change thoroughness
|
|
|
|
ROCKSDB_NAMESPACE::Random32 rnd(FLAGS_thoroughness);
|
|
|
|
|
|
|
|
for (uint32_t i = 0; i < FLAGS_thoroughness; ++i) {
|
|
|
|
// We are going to choose num_to_add using an exponential distribution
|
|
|
|
// as noted above, but instead of randomly choosing them, we generate
|
|
|
|
// samples linearly using the golden ratio, which ensures a nice spread
|
|
|
|
// even for a small number of samples, and starting with the minimum
|
|
|
|
// number of slots to ensure it is tested.
|
|
|
|
double log_add =
|
|
|
|
std::fmod(0.6180339887498948482 * diff_log_add * i, diff_log_add) +
|
|
|
|
log_min_add;
|
|
|
|
uint32_t num_to_add = static_cast<uint32_t>(std::exp(log_add));
|
|
|
|
|
|
|
|
// Most of the time, test the Interleaved solution storage, but when
|
|
|
|
// we do we have to make num_slots a multiple of kCoeffBits. So
|
|
|
|
// sometimes we want to test without that limitation.
|
|
|
|
bool test_interleaved = (i % 7) != 6;
|
|
|
|
|
|
|
|
// Compute num_slots, and re-adjust num_to_add to get as close as possible
|
|
|
|
// to next num_slots, to stress that num_slots in terms of construction
|
|
|
|
// success. Ensure at least one iteration:
|
|
|
|
Index num_slots = Index{0} - 1;
|
|
|
|
--num_to_add;
|
|
|
|
for (;;) {
|
|
|
|
Index next_num_slots = SimpleSoln::RoundUpNumSlots(
|
|
|
|
ConfigHelper::GetNumSlots(num_to_add + 1, cs));
|
|
|
|
if (test_interleaved) {
|
|
|
|
next_num_slots = InterleavedSoln::RoundUpNumSlots(next_num_slots);
|
|
|
|
// assert idempotent
|
|
|
|
EXPECT_EQ(next_num_slots,
|
|
|
|
InterleavedSoln::RoundUpNumSlots(next_num_slots));
|
|
|
|
}
|
|
|
|
// assert idempotent with InterleavedSoln::RoundUpNumSlots
|
|
|
|
EXPECT_EQ(next_num_slots, SimpleSoln::RoundUpNumSlots(next_num_slots));
|
|
|
|
|
|
|
|
if (next_num_slots > num_slots) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
num_slots = next_num_slots;
|
|
|
|
++num_to_add;
|
|
|
|
}
|
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
|
|
|
assert(num_slots < Index{0} - 1);
|
|
|
|
|
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
|
|
|
total_added += num_to_add;
|
|
|
|
|
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
|
|
|
std::string prefix;
|
|
|
|
ROCKSDB_NAMESPACE::PutFixed32(&prefix, rnd.Next());
|
|
|
|
|
|
|
|
// Batch that must be added
|
|
|
|
std::string added_str = prefix + "added";
|
|
|
|
KeyGen keys_begin(added_str, 0);
|
|
|
|
KeyGen keys_end(added_str, num_to_add);
|
|
|
|
|
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 couple more that will probably be added
|
|
|
|
KeyGen one_more(prefix + "more", 1);
|
|
|
|
KeyGen two_more(prefix + "more", 2);
|
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
|
|
|
// Batch that may or may not be added
|
|
|
|
uint32_t batch_size =
|
|
|
|
static_cast<uint32_t>(2.0 * std::sqrt(num_slots - num_to_add));
|
|
|
|
if (batch_size < 10U) {
|
|
|
|
batch_size = 0;
|
|
|
|
}
|
|
|
|
std::string batch_str = prefix + "batch";
|
|
|
|
KeyGen batch_begin(batch_str, 0);
|
|
|
|
KeyGen batch_end(batch_str, batch_size);
|
|
|
|
|
|
|
|
// Batch never (successfully) added, but used for querying FP rate
|
|
|
|
std::string not_str = prefix + "not";
|
|
|
|
KeyGen other_keys_begin(not_str, 0);
|
|
|
|
KeyGen other_keys_end(not_str, FLAGS_max_check);
|
|
|
|
|
|
|
|
double overhead_ratio = 1.0 * num_slots / num_to_add;
|
|
|
|
if (FLAGS_verbose) {
|
|
|
|
fprintf(stderr, "Adding(%s) %u / %u Overhead: %g Batch size: %u\n",
|
|
|
|
test_interleaved ? "i" : "s", (unsigned)num_to_add,
|
|
|
|
(unsigned)num_slots, overhead_ratio, (unsigned)batch_size);
|
|
|
|
}
|
|
|
|
|
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
|
|
|
// Vary bytes for InterleavedSoln to use number of solution columns
|
|
|
|
// from 0 to max allowed by ResultRow type (and used by SimpleSoln).
|
|
|
|
// Specifically include 0 and max, and otherwise skew toward max.
|
|
|
|
uint32_t max_ibytes =
|
|
|
|
static_cast<uint32_t>(sizeof(ResultRow) * num_slots);
|
|
|
|
size_t ibytes;
|
|
|
|
if (i == 0) {
|
|
|
|
ibytes = 0;
|
|
|
|
} else if (i == 1) {
|
|
|
|
ibytes = max_ibytes;
|
|
|
|
} else {
|
|
|
|
// Skewed
|
|
|
|
ibytes =
|
|
|
|
std::max(rnd.Uniformish(max_ibytes), rnd.Uniformish(max_ibytes));
|
|
|
|
}
|
|
|
|
std::unique_ptr<char[]> idata(new char[ibytes]);
|
|
|
|
InterleavedSoln isoln(idata.get(), ibytes);
|
|
|
|
|
|
|
|
SimpleSoln soln;
|
|
|
|
Hasher hasher;
|
|
|
|
bool first_single;
|
|
|
|
bool second_single;
|
|
|
|
bool batch_success;
|
|
|
|
{
|
|
|
|
Banding banding;
|
|
|
|
// Traditional solve for a fixed set.
|
|
|
|
ASSERT_TRUE(
|
|
|
|
banding.ResetAndFindSeedToSolve(num_slots, keys_begin, keys_end));
|
|
|
|
|
|
|
|
Index occupied_count = banding.GetOccupiedCount();
|
|
|
|
Index more_added = 0;
|
|
|
|
|
|
|
|
if (TypeParam::kHomogeneous || overhead_ratio < 1.01 ||
|
|
|
|
batch_size == 0) {
|
|
|
|
// Homogeneous not compatible with backtracking because add
|
|
|
|
// doesn't fail. Small overhead ratio too packed to expect more
|
|
|
|
first_single = false;
|
|
|
|
second_single = false;
|
|
|
|
batch_success = false;
|
|
|
|
} else {
|
|
|
|
// Now to test backtracking, starting with guaranteed fail. By using
|
|
|
|
// the keys that will be used to test FP rate, we are then doing an
|
|
|
|
// extra check that after backtracking there are no remnants (e.g. in
|
|
|
|
// result side of banding) of these entries.
|
|
|
|
KeyGen other_keys_too_big_end = other_keys_begin;
|
|
|
|
other_keys_too_big_end += num_to_add;
|
|
|
|
banding.EnsureBacktrackSize(std::max(num_to_add, batch_size));
|
|
|
|
EXPECT_FALSE(banding.AddRangeOrRollBack(other_keys_begin,
|
|
|
|
other_keys_too_big_end));
|
|
|
|
EXPECT_EQ(occupied_count, banding.GetOccupiedCount());
|
|
|
|
|
|
|
|
// Check that we still have a good chance of adding a couple more
|
|
|
|
// individually
|
|
|
|
first_single = banding.Add(*one_more);
|
|
|
|
second_single = banding.Add(*two_more);
|
|
|
|
more_added += (first_single ? 1 : 0) + (second_single ? 1 : 0);
|
|
|
|
total_singles += 2U;
|
|
|
|
total_single_failures += 2U - more_added;
|
|
|
|
|
|
|
|
// Or as a batch
|
|
|
|
batch_success = banding.AddRangeOrRollBack(batch_begin, batch_end);
|
|
|
|
++total_batch;
|
|
|
|
if (batch_success) {
|
|
|
|
more_added += batch_size;
|
|
|
|
++total_batch_successes;
|
|
|
|
}
|
|
|
|
EXPECT_LE(banding.GetOccupiedCount(), occupied_count + more_added);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Also verify that redundant adds are OK (no effect)
|
|
|
|
ASSERT_TRUE(
|
|
|
|
banding.AddRange(keys_begin, KeyGen(added_str, num_to_add / 8)));
|
|
|
|
EXPECT_LE(banding.GetOccupiedCount(), occupied_count + more_added);
|
|
|
|
|
|
|
|
// Now back-substitution
|
|
|
|
soln.BackSubstFrom(banding);
|
|
|
|
if (test_interleaved) {
|
|
|
|
isoln.BackSubstFrom(banding);
|
|
|
|
}
|
|
|
|
|
|
|
|
Seed reseeds = banding.GetOrdinalSeed();
|
|
|
|
total_reseeds += reseeds;
|
|
|
|
|
|
|
|
EXPECT_LE(reseeds, 8 + log2_thoroughness);
|
|
|
|
if (reseeds > log2_thoroughness + 1) {
|
|
|
|
fprintf(
|
|
|
|
stderr, "%s high reseeds at %u, %u/%u: %u\n",
|
|
|
|
reseeds > log2_thoroughness + 8 ? "ERROR Extremely" : "Somewhat",
|
|
|
|
static_cast<unsigned>(i), static_cast<unsigned>(num_to_add),
|
|
|
|
static_cast<unsigned>(num_slots), static_cast<unsigned>(reseeds));
|
|
|
|
}
|
|
|
|
|
|
|
|
if (reseeds > 0) {
|
|
|
|
// "Expand" test: given a failed construction, how likely is it to
|
|
|
|
// pass with same seed and more slots. At each step, we increase
|
|
|
|
// enough to ensure there is at least one shift within each coeff
|
|
|
|
// block.
|
|
|
|
++total_expand_trials;
|
|
|
|
Index expand_count = 0;
|
|
|
|
Index ex_slots = num_slots;
|
|
|
|
banding.SetOrdinalSeed(0);
|
|
|
|
for (;; ++expand_count) {
|
|
|
|
ASSERT_LE(expand_count, log2_thoroughness);
|
|
|
|
ex_slots += ex_slots / kCoeffBits;
|
|
|
|
if (test_interleaved) {
|
|
|
|
ex_slots = InterleavedSoln::RoundUpNumSlots(ex_slots);
|
|
|
|
}
|
|
|
|
banding.Reset(ex_slots);
|
|
|
|
bool success = banding.AddRange(keys_begin, keys_end);
|
|
|
|
if (success) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
total_expand_failures += expand_count;
|
|
|
|
total_expand_overhead += 1.0 * (ex_slots - num_slots) / num_slots;
|
|
|
|
}
|
|
|
|
|
|
|
|
hasher.SetOrdinalSeed(reseeds);
|
|
|
|
}
|
|
|
|
// soln and hasher now independent of Banding object
|
|
|
|
|
|
|
|
// Verify keys added
|
|
|
|
KeyGen cur = keys_begin;
|
|
|
|
while (cur != keys_end) {
|
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
|
|
|
ASSERT_TRUE(soln.FilterQuery(*cur, hasher));
|
|
|
|
ASSERT_TRUE(!test_interleaved || isoln.FilterQuery(*cur, hasher));
|
|
|
|
++cur;
|
|
|
|
}
|
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
|
|
|
// We (maybe) snuck these in!
|
|
|
|
if (first_single) {
|
|
|
|
ASSERT_TRUE(soln.FilterQuery(*one_more, hasher));
|
|
|
|
ASSERT_TRUE(!test_interleaved || isoln.FilterQuery(*one_more, hasher));
|
|
|
|
}
|
|
|
|
if (second_single) {
|
|
|
|
ASSERT_TRUE(soln.FilterQuery(*two_more, hasher));
|
|
|
|
ASSERT_TRUE(!test_interleaved || isoln.FilterQuery(*two_more, hasher));
|
|
|
|
}
|
|
|
|
if (batch_success) {
|
|
|
|
cur = batch_begin;
|
|
|
|
while (cur != batch_end) {
|
|
|
|
ASSERT_TRUE(soln.FilterQuery(*cur, hasher));
|
|
|
|
ASSERT_TRUE(!test_interleaved || isoln.FilterQuery(*cur, hasher));
|
|
|
|
++cur;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
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
|
|
|
// Check FP rate (depends only on number of result bits == solution
|
|
|
|
// columns)
|
|
|
|
Index fp_count = 0;
|
|
|
|
cur = other_keys_begin;
|
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
|
|
|
{
|
|
|
|
ROCKSDB_NAMESPACE::StopWatchNano timer(
|
|
|
|
ROCKSDB_NAMESPACE::SystemClock::Default(), true);
|
|
|
|
while (cur != other_keys_end) {
|
|
|
|
bool fp = soln.FilterQuery(*cur, hasher);
|
|
|
|
fp_count += fp ? 1 : 0;
|
|
|
|
++cur;
|
|
|
|
}
|
|
|
|
soln_query_nanos += timer.ElapsedNanos();
|
|
|
|
soln_query_count += FLAGS_max_check;
|
|
|
|
}
|
|
|
|
{
|
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 expected_fp_count = soln.ExpectedFpRate() * FLAGS_max_check;
|
|
|
|
// For expected FP rate, also include false positives due to collisions
|
|
|
|
// in Hash value. (Negligible for 64-bit, can matter for 32-bit.)
|
|
|
|
double correction =
|
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
|
|
|
FLAGS_max_check * ExpectedCollisionFpRate(hasher, num_to_add);
|
|
|
|
|
|
|
|
// NOTE: rare violations expected with kHomogeneous
|
|
|
|
EXPECT_LE(fp_count,
|
|
|
|
FrequentPoissonUpperBound(expected_fp_count + correction));
|
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
|
|
|
EXPECT_GE(fp_count,
|
|
|
|
FrequentPoissonLowerBound(expected_fp_count + correction));
|
|
|
|
}
|
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
|
|
|
total_fp_count += fp_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
|
|
|
// And also check FP rate for isoln
|
|
|
|
if (test_interleaved) {
|
|
|
|
Index ifp_count = 0;
|
|
|
|
cur = other_keys_begin;
|
|
|
|
ROCKSDB_NAMESPACE::StopWatchNano timer(
|
|
|
|
ROCKSDB_NAMESPACE::SystemClock::Default(), true);
|
|
|
|
while (cur != other_keys_end) {
|
|
|
|
ifp_count += isoln.FilterQuery(*cur, hasher) ? 1 : 0;
|
|
|
|
++cur;
|
|
|
|
}
|
|
|
|
isoln_query_nanos += timer.ElapsedNanos();
|
|
|
|
isoln_query_count += FLAGS_max_check;
|
|
|
|
{
|
|
|
|
double expected_fp_count = isoln.ExpectedFpRate() * FLAGS_max_check;
|
|
|
|
// For expected FP rate, also include false positives due to
|
|
|
|
// collisions in Hash value. (Negligible for 64-bit, can matter for
|
|
|
|
// 32-bit.)
|
|
|
|
double correction =
|
|
|
|
FLAGS_max_check * ExpectedCollisionFpRate(hasher, num_to_add);
|
|
|
|
|
|
|
|
// NOTE: rare violations expected with kHomogeneous
|
|
|
|
EXPECT_LE(ifp_count,
|
|
|
|
FrequentPoissonUpperBound(expected_fp_count + correction));
|
|
|
|
|
|
|
|
// FIXME: why sometimes can we slightly "beat the odds"?
|
|
|
|
// (0.95 factor should not be needed)
|
|
|
|
EXPECT_GE(ifp_count, FrequentPoissonLowerBound(
|
|
|
|
0.95 * expected_fp_count + correction));
|
|
|
|
}
|
|
|
|
// Since the bits used in isoln are a subset of the bits used in soln,
|
|
|
|
// it cannot have fewer FPs
|
|
|
|
EXPECT_GE(ifp_count, fp_count);
|
|
|
|
}
|
|
|
|
|
|
|
|
// And compare to Bloom time, for fun
|
|
|
|
if (ibytes >= /* minimum Bloom impl bytes*/ 64) {
|
|
|
|
Index bfp_count = 0;
|
|
|
|
cur = other_keys_begin;
|
|
|
|
ROCKSDB_NAMESPACE::StopWatchNano timer(
|
|
|
|
ROCKSDB_NAMESPACE::SystemClock::Default(), true);
|
|
|
|
while (cur != other_keys_end) {
|
|
|
|
uint64_t h = hasher.GetHash(*cur);
|
|
|
|
uint32_t h1 = ROCKSDB_NAMESPACE::Lower32of64(h);
|
|
|
|
uint32_t h2 = sizeof(Hash) >= 8 ? ROCKSDB_NAMESPACE::Upper32of64(h)
|
|
|
|
: h1 * 0x9e3779b9;
|
|
|
|
bfp_count +=
|
|
|
|
ROCKSDB_NAMESPACE::FastLocalBloomImpl::HashMayMatch(
|
|
|
|
h1, h2, static_cast<uint32_t>(ibytes), 6, idata.get())
|
|
|
|
? 1
|
|
|
|
: 0;
|
|
|
|
++cur;
|
|
|
|
}
|
|
|
|
bloom_query_nanos += timer.ElapsedNanos();
|
|
|
|
// ensure bfp_count is used
|
|
|
|
ASSERT_LT(bfp_count, FLAGS_max_check);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
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
|
|
|
// "outside" == key not in original set so either negative or false positive
|
|
|
|
fprintf(stderr,
|
|
|
|
"Simple outside query, hot, incl hashing, ns/key: %g\n",
|
|
|
|
1.0 * soln_query_nanos / soln_query_count);
|
|
|
|
fprintf(stderr,
|
|
|
|
"Interleaved outside query, hot, incl hashing, ns/key: %g\n",
|
|
|
|
1.0 * isoln_query_nanos / isoln_query_count);
|
|
|
|
fprintf(stderr,
|
|
|
|
"Bloom outside query, hot, incl hashing, ns/key: %g\n",
|
|
|
|
1.0 * bloom_query_nanos / soln_query_count);
|
|
|
|
|
|
|
|
if (TypeParam::kHomogeneous) {
|
|
|
|
EXPECT_EQ(total_reseeds, 0U);
|
|
|
|
} else {
|
|
|
|
double average_reseeds = 1.0 * total_reseeds / FLAGS_thoroughness;
|
|
|
|
fprintf(stderr, "Average re-seeds: %g\n", average_reseeds);
|
|
|
|
// Values above were chosen to target around 50% chance of encoding
|
|
|
|
// success rate (average of 1.0 re-seeds) or slightly better. But 1.15 is
|
|
|
|
// also close enough.
|
|
|
|
EXPECT_LE(total_reseeds,
|
|
|
|
InfrequentPoissonUpperBound(1.15 * expected_reseeds *
|
|
|
|
FLAGS_thoroughness));
|
|
|
|
// Would use 0.85 here instead of 0.75, but
|
|
|
|
// TypesAndSettings_Hash32_SmallKeyGen can "beat the odds" because of
|
|
|
|
// sequential keys with a small, cheap hash function. We accept that
|
|
|
|
// there are surely inputs that are somewhat bad for this setup, but
|
|
|
|
// these somewhat good inputs are probably more likely.
|
|
|
|
EXPECT_GE(total_reseeds,
|
|
|
|
InfrequentPoissonLowerBound(0.75 * expected_reseeds *
|
|
|
|
FLAGS_thoroughness));
|
|
|
|
}
|
|
|
|
|
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 (total_expand_trials > 0) {
|
|
|
|
double average_expand_failures =
|
|
|
|
1.0 * total_expand_failures / total_expand_trials;
|
|
|
|
fprintf(stderr, "Average expand failures, and overhead: %g, %g\n",
|
|
|
|
average_expand_failures,
|
|
|
|
total_expand_overhead / total_expand_trials);
|
|
|
|
// Seems to be a generous allowance
|
|
|
|
EXPECT_LE(total_expand_failures,
|
|
|
|
InfrequentPoissonUpperBound(1.0 * total_expand_trials));
|
|
|
|
} else {
|
|
|
|
fprintf(stderr, "Average expand failures: N/A\n");
|
|
|
|
}
|
|
|
|
|
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 (total_singles > 0) {
|
|
|
|
double single_failure_rate = 1.0 * total_single_failures / total_singles;
|
|
|
|
fprintf(stderr, "Add'l single, failure rate: %g\n", single_failure_rate);
|
|
|
|
// A rough bound (one sided) based on nothing in particular
|
|
|
|
double expected_single_failures =
|
|
|
|
1.0 * total_singles /
|
|
|
|
(sizeof(CoeffRow) == 16 ? 128 : TypeParam::kUseSmash ? 64 : 32);
|
|
|
|
EXPECT_LE(total_single_failures,
|
|
|
|
InfrequentPoissonUpperBound(expected_single_failures));
|
|
|
|
}
|
|
|
|
|
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 (total_batch > 0) {
|
|
|
|
// Counting successes here for Poisson to approximate the Binomial
|
|
|
|
// distribution.
|
|
|
|
// A rough bound (one sided) based on nothing in particular.
|
|
|
|
double expected_batch_successes = 1.0 * total_batch / 2;
|
|
|
|
uint64_t lower_bound =
|
|
|
|
InfrequentPoissonLowerBound(expected_batch_successes);
|
|
|
|
fprintf(stderr, "Add'l batch, success rate: %g (>= %g)\n",
|
|
|
|
1.0 * total_batch_successes / total_batch,
|
|
|
|
1.0 * lower_bound / total_batch);
|
|
|
|
EXPECT_GE(total_batch_successes, lower_bound);
|
|
|
|
}
|
|
|
|
|
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
|
|
|
{
|
|
|
|
uint64_t total_checked = uint64_t{FLAGS_max_check} * FLAGS_thoroughness;
|
|
|
|
double expected_total_fp_count =
|
|
|
|
total_checked * std::pow(0.5, 8U * sizeof(ResultRow));
|
|
|
|
// For expected FP rate, also include false positives due to collisions
|
|
|
|
// in Hash value. (Negligible for 64-bit, can matter for 32-bit.)
|
|
|
|
double average_added = 1.0 * total_added / FLAGS_thoroughness;
|
|
|
|
expected_total_fp_count +=
|
|
|
|
total_checked * ExpectedCollisionFpRate(Hasher(), average_added);
|
|
|
|
|
|
|
|
uint64_t upper_bound =
|
|
|
|
InfrequentPoissonUpperBound(expected_total_fp_count);
|
|
|
|
uint64_t lower_bound =
|
|
|
|
InfrequentPoissonLowerBound(expected_total_fp_count);
|
|
|
|
fprintf(stderr, "Average FP rate: %g (~= %g, <= %g, >= %g)\n",
|
|
|
|
1.0 * total_fp_count / total_checked,
|
|
|
|
expected_total_fp_count / total_checked,
|
|
|
|
1.0 * upper_bound / total_checked,
|
|
|
|
1.0 * lower_bound / total_checked);
|
|
|
|
EXPECT_LE(total_fp_count, upper_bound);
|
|
|
|
EXPECT_GE(total_fp_count, lower_bound);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
TYPED_TEST(RibbonTypeParamTest, Extremes) {
|
|
|
|
IMPORT_RIBBON_TYPES_AND_SETTINGS(TypeParam);
|
|
|
|
IMPORT_RIBBON_IMPL_TYPES(TypeParam);
|
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
|
|
|
using KeyGen = typename TypeParam::KeyGen;
|
|
|
|
|
|
|
|
size_t bytes = 128 * 1024;
|
|
|
|
std::unique_ptr<char[]> buf(new char[bytes]);
|
|
|
|
InterleavedSoln isoln(buf.get(), bytes);
|
|
|
|
SimpleSoln soln;
|
|
|
|
Hasher hasher;
|
|
|
|
Banding banding;
|
|
|
|
|
|
|
|
// ########################################
|
|
|
|
// Add zero keys to minimal number of slots
|
|
|
|
KeyGen begin_and_end("foo", 123);
|
|
|
|
ASSERT_TRUE(banding.ResetAndFindSeedToSolve(
|
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
|
|
|
/*slots*/ kCoeffBits, begin_and_end, begin_and_end, /*first seed*/ 0,
|
|
|
|
/* seed mask*/ 0));
|
|
|
|
|
|
|
|
soln.BackSubstFrom(banding);
|
|
|
|
isoln.BackSubstFrom(banding);
|
|
|
|
|
|
|
|
// Because there's plenty of memory, we expect the interleaved solution to
|
|
|
|
// use maximum supported columns (same as simple solution)
|
|
|
|
ASSERT_EQ(isoln.GetUpperNumColumns(), 8U * sizeof(ResultRow));
|
|
|
|
ASSERT_EQ(isoln.GetUpperStartBlock(), 0U);
|
|
|
|
|
|
|
|
// Somewhat oddly, we expect same FP rate as if we had essentially filled
|
|
|
|
// up the slots.
|
|
|
|
KeyGen other_keys_begin("not", 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
|
|
|
KeyGen other_keys_end("not", FLAGS_max_check);
|
|
|
|
|
|
|
|
Index fp_count = 0;
|
|
|
|
KeyGen cur = other_keys_begin;
|
|
|
|
while (cur != other_keys_end) {
|
|
|
|
bool isoln_query_result = isoln.FilterQuery(*cur, hasher);
|
|
|
|
bool soln_query_result = soln.FilterQuery(*cur, hasher);
|
|
|
|
// Solutions are equivalent
|
|
|
|
ASSERT_EQ(isoln_query_result, soln_query_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
|
|
|
if (!TypeParam::kHomogeneous) {
|
|
|
|
// And in fact we only expect an FP when ResultRow is 0
|
|
|
|
// (except Homogeneous)
|
|
|
|
ASSERT_EQ(soln_query_result, hasher.GetResultRowFromHash(
|
|
|
|
hasher.GetHash(*cur)) == ResultRow{0});
|
|
|
|
}
|
|
|
|
fp_count += soln_query_result ? 1 : 0;
|
|
|
|
++cur;
|
|
|
|
}
|
|
|
|
{
|
|
|
|
ASSERT_EQ(isoln.ExpectedFpRate(), soln.ExpectedFpRate());
|
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 expected_fp_count = isoln.ExpectedFpRate() * FLAGS_max_check;
|
|
|
|
EXPECT_LE(fp_count, InfrequentPoissonUpperBound(expected_fp_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
|
|
|
if (TypeParam::kHomogeneous) {
|
|
|
|
// Pseudorandom garbage in Homogeneous filter can "beat the odds" if
|
|
|
|
// nothing added
|
|
|
|
} else {
|
|
|
|
EXPECT_GE(fp_count, InfrequentPoissonLowerBound(expected_fp_count));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// ######################################################
|
|
|
|
// Use zero bytes for interleaved solution (key(s) added)
|
|
|
|
|
|
|
|
// Add one key
|
|
|
|
KeyGen key_begin("added", 0);
|
|
|
|
KeyGen key_end("added", 1);
|
|
|
|
ASSERT_TRUE(banding.ResetAndFindSeedToSolve(
|
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
|
|
|
/*slots*/ kCoeffBits, key_begin, key_end, /*first seed*/ 0,
|
|
|
|
/* seed mask*/ 0));
|
|
|
|
|
|
|
|
InterleavedSoln isoln2(nullptr, /*bytes*/ 0);
|
|
|
|
|
|
|
|
isoln2.BackSubstFrom(banding);
|
|
|
|
|
|
|
|
ASSERT_EQ(isoln2.GetUpperNumColumns(), 0U);
|
|
|
|
ASSERT_EQ(isoln2.GetUpperStartBlock(), 0U);
|
|
|
|
|
|
|
|
// All queries return true
|
|
|
|
ASSERT_TRUE(isoln2.FilterQuery(*other_keys_begin, hasher));
|
|
|
|
ASSERT_EQ(isoln2.ExpectedFpRate(), 1.0);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(RibbonTest, AllowZeroStarts) {
|
|
|
|
IMPORT_RIBBON_TYPES_AND_SETTINGS(TypesAndSettings_AllowZeroStarts);
|
|
|
|
IMPORT_RIBBON_IMPL_TYPES(TypesAndSettings_AllowZeroStarts);
|
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
|
|
|
using KeyGen = StandardKeyGen;
|
|
|
|
|
|
|
|
InterleavedSoln isoln(nullptr, /*bytes*/ 0);
|
|
|
|
SimpleSoln soln;
|
|
|
|
Hasher hasher;
|
|
|
|
Banding banding;
|
|
|
|
|
|
|
|
KeyGen begin("foo", 0);
|
|
|
|
KeyGen end("foo", 1);
|
|
|
|
// Can't add 1 entry
|
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
|
|
|
ASSERT_FALSE(banding.ResetAndFindSeedToSolve(/*slots*/ 0, begin, end));
|
|
|
|
|
|
|
|
KeyGen begin_and_end("foo", 123);
|
|
|
|
// Can add 0 entries
|
|
|
|
ASSERT_TRUE(banding.ResetAndFindSeedToSolve(/*slots*/ 0, begin_and_end,
|
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
|
|
|
begin_and_end));
|
|
|
|
|
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 reseeds = banding.GetOrdinalSeed();
|
|
|
|
ASSERT_EQ(reseeds, 0U);
|
|
|
|
hasher.SetOrdinalSeed(reseeds);
|
|
|
|
|
|
|
|
// Can construct 0-slot solutions
|
|
|
|
isoln.BackSubstFrom(banding);
|
|
|
|
soln.BackSubstFrom(banding);
|
|
|
|
|
|
|
|
// Should always return false
|
|
|
|
ASSERT_FALSE(isoln.FilterQuery(*begin, hasher));
|
|
|
|
ASSERT_FALSE(soln.FilterQuery(*begin, hasher));
|
|
|
|
|
|
|
|
// And report that in FP rate
|
|
|
|
ASSERT_EQ(isoln.ExpectedFpRate(), 0.0);
|
|
|
|
ASSERT_EQ(soln.ExpectedFpRate(), 0.0);
|
|
|
|
}
|
|
|
|
|
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
|
|
|
TEST(RibbonTest, RawAndOrdinalSeeds) {
|
|
|
|
StandardHasher<TypesAndSettings_Seed64> hasher64;
|
|
|
|
StandardHasher<DefaultTypesAndSettings> hasher64_32;
|
|
|
|
StandardHasher<TypesAndSettings_Hash32> hasher32;
|
|
|
|
StandardHasher<TypesAndSettings_Seed8> hasher8;
|
|
|
|
|
|
|
|
for (uint32_t limit : {0xffU, 0xffffU}) {
|
|
|
|
std::vector<bool> seen(limit + 1);
|
|
|
|
for (uint32_t i = 0; i < limit; ++i) {
|
|
|
|
hasher64.SetOrdinalSeed(i);
|
|
|
|
auto raw64 = hasher64.GetRawSeed();
|
|
|
|
hasher32.SetOrdinalSeed(i);
|
|
|
|
auto raw32 = hasher32.GetRawSeed();
|
|
|
|
hasher8.SetOrdinalSeed(static_cast<uint8_t>(i));
|
|
|
|
auto raw8 = hasher8.GetRawSeed();
|
|
|
|
{
|
|
|
|
hasher64_32.SetOrdinalSeed(i);
|
|
|
|
auto raw64_32 = hasher64_32.GetRawSeed();
|
|
|
|
ASSERT_EQ(raw64_32, raw32); // Same size seed
|
|
|
|
}
|
|
|
|
if (i == 0) {
|
|
|
|
// Documented that ordinal seed 0 == raw seed 0
|
|
|
|
ASSERT_EQ(raw64, 0U);
|
|
|
|
ASSERT_EQ(raw32, 0U);
|
|
|
|
ASSERT_EQ(raw8, 0U);
|
|
|
|
} else {
|
|
|
|
// Extremely likely that upper bits are set
|
|
|
|
ASSERT_GT(raw64, raw32);
|
|
|
|
ASSERT_GT(raw32, raw8);
|
|
|
|
}
|
|
|
|
// Hashers agree on lower bits
|
|
|
|
ASSERT_EQ(static_cast<uint32_t>(raw64), raw32);
|
|
|
|
ASSERT_EQ(static_cast<uint8_t>(raw32), raw8);
|
|
|
|
|
|
|
|
// The translation is one-to-one for this size prefix
|
|
|
|
uint32_t v = static_cast<uint32_t>(raw32 & limit);
|
|
|
|
ASSERT_EQ(raw64 & limit, v);
|
|
|
|
ASSERT_FALSE(seen[v]);
|
|
|
|
seen[v] = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
|
|
|
|
struct PhsfInputGen {
|
|
|
|
PhsfInputGen(const std::string& prefix, uint64_t id) : id_(id) {
|
|
|
|
val_.first = prefix;
|
|
|
|
ROCKSDB_NAMESPACE::PutFixed64(&val_.first, /*placeholder*/ 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Prefix (only one required)
|
|
|
|
PhsfInputGen& operator++() {
|
|
|
|
++id_;
|
|
|
|
return *this;
|
|
|
|
}
|
|
|
|
|
|
|
|
const std::pair<std::string, uint8_t>& operator*() {
|
|
|
|
// Use multiplication to mix things up a little in the key
|
|
|
|
ROCKSDB_NAMESPACE::EncodeFixed64(&val_.first[val_.first.size() - 8],
|
|
|
|
id_ * uint64_t{0x1500000001});
|
|
|
|
// Occasionally repeat values etc.
|
|
|
|
val_.second = static_cast<uint8_t>(id_ * 7 / 8);
|
|
|
|
return val_;
|
|
|
|
}
|
|
|
|
|
|
|
|
const std::pair<std::string, uint8_t>* operator->() { return &**this; }
|
|
|
|
|
|
|
|
bool operator==(const PhsfInputGen& other) {
|
|
|
|
// Same prefix is assumed
|
|
|
|
return id_ == other.id_;
|
|
|
|
}
|
|
|
|
bool operator!=(const PhsfInputGen& other) {
|
|
|
|
// Same prefix is assumed
|
|
|
|
return id_ != other.id_;
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t id_;
|
|
|
|
std::pair<std::string, uint8_t> val_;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct PhsfTypesAndSettings : public DefaultTypesAndSettings {
|
|
|
|
static constexpr bool kIsFilter = false;
|
|
|
|
};
|
|
|
|
} // namespace
|
|
|
|
|
|
|
|
TEST(RibbonTest, PhsfBasic) {
|
|
|
|
IMPORT_RIBBON_TYPES_AND_SETTINGS(PhsfTypesAndSettings);
|
|
|
|
IMPORT_RIBBON_IMPL_TYPES(PhsfTypesAndSettings);
|
|
|
|
|
|
|
|
Index num_slots = 12800;
|
|
|
|
Index num_to_add = static_cast<Index>(num_slots / 1.02);
|
|
|
|
|
|
|
|
PhsfInputGen begin("in", 0);
|
|
|
|
PhsfInputGen end("in", num_to_add);
|
|
|
|
|
|
|
|
std::unique_ptr<char[]> idata(new char[/*bytes*/ num_slots]);
|
|
|
|
InterleavedSoln isoln(idata.get(), /*bytes*/ num_slots);
|
|
|
|
SimpleSoln soln;
|
|
|
|
Hasher hasher;
|
|
|
|
|
|
|
|
{
|
|
|
|
Banding banding;
|
|
|
|
ASSERT_TRUE(banding.ResetAndFindSeedToSolve(num_slots, begin, end));
|
|
|
|
|
|
|
|
soln.BackSubstFrom(banding);
|
|
|
|
isoln.BackSubstFrom(banding);
|
|
|
|
|
|
|
|
hasher.SetOrdinalSeed(banding.GetOrdinalSeed());
|
|
|
|
}
|
|
|
|
|
|
|
|
for (PhsfInputGen cur = begin; cur != end; ++cur) {
|
|
|
|
ASSERT_EQ(cur->second, soln.PhsfQuery(cur->first, hasher));
|
|
|
|
ASSERT_EQ(cur->second, isoln.PhsfQuery(cur->first, hasher));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
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
|
|
|
// Not a real test, but a tool used to build APIs in ribbon_config.h
|
|
|
|
TYPED_TEST(RibbonTypeParamTest, FindOccupancy) {
|
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
|
|
|
IMPORT_RIBBON_TYPES_AND_SETTINGS(TypeParam);
|
|
|
|
IMPORT_RIBBON_IMPL_TYPES(TypeParam);
|
|
|
|
using KeyGen = typename TypeParam::KeyGen;
|
|
|
|
|
|
|
|
if (!FLAGS_find_occ) {
|
|
|
|
fprintf(stderr, "Tool disabled during unit test runs\n");
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
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
|
|
|
KeyGen cur(ROCKSDB_NAMESPACE::ToString(
|
|
|
|
testing::UnitTest::GetInstance()->random_seed()),
|
|
|
|
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
|
|
|
|
|
|
|
Banding banding;
|
|
|
|
Index num_slots = InterleavedSoln::RoundUpNumSlots(FLAGS_find_min_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
|
|
|
Index max_slots = InterleavedSoln::RoundUpNumSlots(FLAGS_find_max_slots);
|
|
|
|
while (num_slots <= max_slots) {
|
|
|
|
std::map<int32_t, uint32_t> rem_histogram;
|
|
|
|
std::map<Index, uint32_t> slot_histogram;
|
|
|
|
if (FLAGS_find_slot_occ) {
|
|
|
|
for (Index i = 0; i < kCoeffBits; ++i) {
|
|
|
|
slot_histogram[i] = 0;
|
|
|
|
slot_histogram[num_slots - 1 - i] = 0;
|
|
|
|
slot_histogram[num_slots / 2 - kCoeffBits / 2 + i] = 0;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
uint64_t total_added = 0;
|
|
|
|
for (uint32_t i = 0; i < FLAGS_find_iters; ++i) {
|
|
|
|
banding.Reset(num_slots);
|
|
|
|
uint32_t j = 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
|
|
|
KeyGen end = cur;
|
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
|
|
|
end += num_slots + num_slots / 10;
|
|
|
|
for (; cur != end; ++cur) {
|
|
|
|
if (banding.Add(*cur)) {
|
|
|
|
++j;
|
|
|
|
} else {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
total_added += j;
|
|
|
|
for (auto& slot : slot_histogram) {
|
|
|
|
slot.second += banding.IsOccupied(slot.first);
|
|
|
|
}
|
|
|
|
|
|
|
|
int32_t bucket =
|
|
|
|
static_cast<int32_t>(num_slots) - static_cast<int32_t>(j);
|
|
|
|
rem_histogram[bucket]++;
|
|
|
|
if (FLAGS_verbose) {
|
|
|
|
fprintf(stderr, "num_slots: %u i: %u / %u avg_overhead: %g\r",
|
|
|
|
static_cast<unsigned>(num_slots), static_cast<unsigned>(i),
|
|
|
|
static_cast<unsigned>(FLAGS_find_iters),
|
|
|
|
1.0 * (i + 1) * num_slots / total_added);
|
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
|
|
|
}
|
|
|
|
}
|
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 (FLAGS_verbose) {
|
|
|
|
fprintf(stderr, "\n");
|
|
|
|
}
|
|
|
|
|
|
|
|
uint32_t cumulative = 0;
|
|
|
|
|
|
|
|
double p50_rem = 0;
|
|
|
|
double p95_rem = 0;
|
|
|
|
double p99_9_rem = 0;
|
|
|
|
|
|
|
|
for (auto& h : rem_histogram) {
|
|
|
|
double before = 1.0 * cumulative / FLAGS_find_iters;
|
|
|
|
double not_after = 1.0 * (cumulative + h.second) / FLAGS_find_iters;
|
|
|
|
if (FLAGS_verbose) {
|
|
|
|
fprintf(stderr, "overhead: %g before: %g not_after: %g\n",
|
|
|
|
1.0 * num_slots / (num_slots - h.first), before, not_after);
|
|
|
|
}
|
|
|
|
cumulative += h.second;
|
|
|
|
if (before < 0.5 && 0.5 <= not_after) {
|
|
|
|
// fake it with linear interpolation
|
|
|
|
double portion = (0.5 - before) / (not_after - before);
|
|
|
|
p50_rem = h.first + portion;
|
|
|
|
} else if (before < 0.95 && 0.95 <= not_after) {
|
|
|
|
// fake it with linear interpolation
|
|
|
|
double portion = (0.95 - before) / (not_after - before);
|
|
|
|
p95_rem = h.first + portion;
|
|
|
|
} else if (before < 0.999 && 0.999 <= not_after) {
|
|
|
|
// fake it with linear interpolation
|
|
|
|
double portion = (0.999 - before) / (not_after - before);
|
|
|
|
p99_9_rem = h.first + portion;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
for (auto& slot : slot_histogram) {
|
|
|
|
fprintf(stderr, "slot[%u] occupied: %g\n", (unsigned)slot.first,
|
|
|
|
1.0 * slot.second / FLAGS_find_iters);
|
|
|
|
}
|
|
|
|
|
|
|
|
double mean_rem =
|
|
|
|
(1.0 * FLAGS_find_iters * num_slots - total_added) / FLAGS_find_iters;
|
|
|
|
fprintf(
|
|
|
|
stderr,
|
|
|
|
"num_slots: %u iters: %u mean_ovr: %g p50_ovr: %g p95_ovr: %g "
|
|
|
|
"p99.9_ovr: %g mean_rem: %g p50_rem: %g p95_rem: %g p99.9_rem: %g\n",
|
|
|
|
static_cast<unsigned>(num_slots),
|
|
|
|
static_cast<unsigned>(FLAGS_find_iters),
|
|
|
|
1.0 * num_slots / (num_slots - mean_rem),
|
|
|
|
1.0 * num_slots / (num_slots - p50_rem),
|
|
|
|
1.0 * num_slots / (num_slots - p95_rem),
|
|
|
|
1.0 * num_slots / (num_slots - p99_9_rem), mean_rem, p50_rem, p95_rem,
|
|
|
|
p99_9_rem);
|
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
|
|
|
|
|
|
|
num_slots = std::max(
|
|
|
|
num_slots + 1, static_cast<Index>(num_slots * FLAGS_find_next_factor));
|
|
|
|
num_slots = InterleavedSoln::RoundUpNumSlots(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
|
|
|
// Not a real test, but a tool to understand Homogeneous Ribbon
|
|
|
|
// behavior (TODO: configuration APIs & tests)
|
|
|
|
TYPED_TEST(RibbonTypeParamTest, OptimizeHomogAtScale) {
|
|
|
|
IMPORT_RIBBON_TYPES_AND_SETTINGS(TypeParam);
|
|
|
|
IMPORT_RIBBON_IMPL_TYPES(TypeParam);
|
|
|
|
using KeyGen = typename TypeParam::KeyGen;
|
|
|
|
|
|
|
|
if (!FLAGS_optimize_homog) {
|
|
|
|
fprintf(stderr, "Tool disabled during unit test runs\n");
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!TypeParam::kHomogeneous) {
|
|
|
|
fprintf(stderr, "Only for Homogeneous Ribbon\n");
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
KeyGen cur(ROCKSDB_NAMESPACE::ToString(
|
|
|
|
testing::UnitTest::GetInstance()->random_seed()),
|
|
|
|
0);
|
|
|
|
|
|
|
|
Banding banding;
|
|
|
|
Index num_slots = SimpleSoln::RoundUpNumSlots(FLAGS_optimize_homog_slots);
|
|
|
|
banding.Reset(num_slots);
|
|
|
|
|
|
|
|
// This and "band_ovr" is the "allocated overhead", or slots over added.
|
|
|
|
// It does not take into account FP rates.
|
|
|
|
double target_overhead = 1.20;
|
|
|
|
uint32_t num_added = 0;
|
|
|
|
|
|
|
|
do {
|
|
|
|
do {
|
|
|
|
(void)banding.Add(*cur);
|
|
|
|
++cur;
|
|
|
|
++num_added;
|
|
|
|
} while (1.0 * num_slots / num_added > target_overhead);
|
|
|
|
|
|
|
|
SimpleSoln soln;
|
|
|
|
soln.BackSubstFrom(banding);
|
|
|
|
|
|
|
|
std::array<uint32_t, 8U * sizeof(ResultRow)> fp_counts_by_cols;
|
|
|
|
fp_counts_by_cols.fill(0U);
|
|
|
|
for (uint32_t i = 0; i < FLAGS_optimize_homog_check; ++i) {
|
|
|
|
ResultRow r = soln.PhsfQuery(*cur, banding);
|
|
|
|
++cur;
|
|
|
|
for (size_t j = 0; j < fp_counts_by_cols.size(); ++j) {
|
|
|
|
if ((r & 1) == 1) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
fp_counts_by_cols[j]++;
|
|
|
|
r /= 2;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
fprintf(stderr, "band_ovr: %g ", 1.0 * num_slots / num_added);
|
|
|
|
for (unsigned j = 0; j < fp_counts_by_cols.size(); ++j) {
|
|
|
|
double inv_fp_rate =
|
|
|
|
1.0 * FLAGS_optimize_homog_check / fp_counts_by_cols[j];
|
|
|
|
double equiv_cols = std::log(inv_fp_rate) * 1.4426950409;
|
|
|
|
// Overhead vs. information-theoretic minimum based on observed
|
|
|
|
// FP rate (subject to sampling error, especially for low FP rates)
|
|
|
|
double actual_overhead =
|
|
|
|
1.0 * (j + 1) * num_slots / (equiv_cols * num_added);
|
|
|
|
fprintf(stderr, "ovr_%u: %g ", j + 1, actual_overhead);
|
|
|
|
}
|
|
|
|
fprintf(stderr, "\n");
|
|
|
|
target_overhead -= FLAGS_optimize_homog_granularity;
|
|
|
|
} while (target_overhead > 1.0);
|
|
|
|
}
|
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
|
|
|
|
|
|
|
int main(int argc, char** argv) {
|
|
|
|
::testing::InitGoogleTest(&argc, argv);
|
|
|
|
#ifdef GFLAGS
|
|
|
|
ParseCommandLineFlags(&argc, &argv, true);
|
|
|
|
#endif // GFLAGS
|
|
|
|
return RUN_ALL_TESTS();
|
|
|
|
}
|