6 Commits (7343eb4a74ab7a6f48ce6d7f2a34c806e146dfaa)

Author	SHA1	Message	Date
Peter Dillinger	e4f1e64c30	Add prefetching (batched MultiGet) for experimental Ribbon filter (#7889) ... Summary: Adds support for prefetching data in Ribbon queries, which especially optimizes batched Ribbon queries for MultiGet (~222ns/key to ~97ns/key) but also single key queries on cold memory (~333ns to ~226ns) because many queries span more than one cache line. This required some refactoring of the query algorithm, and there does not appear to be a noticeable regression in "hot memory" query times (perhaps from 48ns to 50ns). Pull Request resolved: https://github.com/facebook/rocksdb/pull/7889 Test Plan: existing unit tests, plus performance validation with filter_bench: Each data point is the best of two runs. I saturated the machine CPUs with other filter_bench runs in the background. Before: $ ./filter_bench -impl=3 -m_keys_total_max=200 -average_keys_per_filter=100000 -m_queries=50 WARNING: Assertions are enabled; benchmarks unnecessarily slow Building... Build avg ns/key: 125.86 Number of filters: 1993 Total size (MB): 168.166 Reported total allocated memory (MB): 183.211 Reported internal fragmentation: 8.94626% Bits/key stored: 7.05341 Prelim FP rate %: 0.951827 ---------------------------- Mixed inside/outside queries... Single filter net ns/op: 48.0111 Batched, prepared net ns/op: 222.384 Batched, unprepared net ns/op: 343.908 Skewed 50% in 1% net ns/op: 252.916 Skewed 80% in 20% net ns/op: 320.579 Random filter net ns/op: 332.957 After: $ ./filter_bench -impl=3 -m_keys_total_max=200 -average_keys_per_filter=100000 -m_queries=50 WARNING: Assertions are enabled; benchmarks unnecessarily slow Building... Build avg ns/key: 128.117 Number of filters: 1993 Total size (MB): 168.166 Reported total allocated memory (MB): 183.211 Reported internal fragmentation: 8.94626% Bits/key stored: 7.05341 Prelim FP rate %: 0.951827 ---------------------------- Mixed inside/outside queries... Single filter net ns/op: 49.8812 Batched, prepared net ns/op: 97.1514 Batched, unprepared net ns/op: 222.025 Skewed 50% in 1% net ns/op: 197.48 Skewed 80% in 20% net ns/op: 212.457 Random filter net ns/op: 226.464 Bloom comparison, for reference: $ ./filter_bench -impl=2 -m_keys_total_max=200 -average_keys_per_filter=100000 -m_queries=50 WARNING: Assertions are enabled; benchmarks unnecessarily slow Building... Build avg ns/key: 35.3042 Number of filters: 1993 Total size (MB): 238.488 Reported total allocated memory (MB): 262.875 Reported internal fragmentation: 10.2255% Bits/key stored: 10.0029 Prelim FP rate %: 0.965327 ---------------------------- Mixed inside/outside queries... Single filter net ns/op: 9.09931 Batched, prepared net ns/op: 34.21 Batched, unprepared net ns/op: 88.8564 Skewed 50% in 1% net ns/op: 139.75 Skewed 80% in 20% net ns/op: 181.264 Random filter net ns/op: 173.88 Reviewed By: jay-zhuang Differential Revision: D26378710 Pulled By: pdillinger fbshipit-source-id: 058428967c55ed763698284cd3b4bbe3351b6e69	4 years ago
Peter Dillinger	239d17a19c	Support optimize_filters_for_memory for Ribbon filter (#7774) ... Summary: Primarily this change refactors the optimize_filters_for_memory code for Bloom filters, based on malloc_usable_size, to also work for Ribbon filters. This change also replaces the somewhat slow but general BuiltinFilterBitsBuilder::ApproximateNumEntries with implementation-specific versions for Ribbon (new) and Legacy Bloom (based on a recently deleted version). The reason is to emphasize speed in ApproximateNumEntries rather than 100% accuracy. Justification: ApproximateNumEntries (formerly CalculateNumEntry) is only used by RocksDB for range-partitioned filters, called each time we start to construct one. (In theory, it should be possible to reuse the estimate, but the abstractions provided by FilterPolicy don't really make that workable.) But this is only used as a heuristic estimate for hitting a desired partitioned filter size because of alignment to data blocks, which have various numbers of unique keys or prefixes. The two factors lead us to prioritize reasonable speed over 100% accuracy. optimize_filters_for_memory adds extra complication, because precisely calculating num_entries for some allowed number of bytes depends on state with optimize_filters_for_memory enabled. And the allocator-agnostic implementation of optimize_filters_for_memory, using malloc_usable_size, means we would have to actually allocate memory, many times, just to precisely determine how many entries (keys) could be added and stay below some size budget, for the current state. (In a draft, I got this working, and then realized the balance of speed vs. accuracy was all wrong.) So related to that, I have made CalculateSpace, an internal-only API only used for testing, non-authoritative also if optimize_filters_for_memory is enabled. This simplifies some code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7774 Test Plan: unit test updated, and for FilterSize test, range of tested values is greatly expanded (still super fast) Also tested `db_bench -benchmarks=fillrandom,stats -bloom_bits=10 -num=1000000 -partition_index_and_filters -format_version=5 [-optimize_filters_for_memory] [-use_ribbon_filter]` with temporary debug output of generated filter sizes. Bloom+optimize_filters_for_memory: 1 Filter size: 197 (224 in memory) 134 Filter size: 3525 (3584 in memory) 107 Filter size: 4037 (4096 in memory) Total on disk: 904,506 Total in memory: 918,752 Ribbon+optimize_filters_for_memory: 1 Filter size: 3061 (3072 in memory) 110 Filter size: 3573 (3584 in memory) 58 Filter size: 4085 (4096 in memory) Total on disk: 633,021 (-30.0%) Total in memory: 634,880 (-30.9%) Bloom (no offm): 1 Filter size: 261 (320 in memory) 1 Filter size: 3333 (3584 in memory) 240 Filter size: 3717 (4096 in memory) Total on disk: 895,674 (-1% on disk vs. +offm; known tolerable overhead of offm) Total in memory: 986,944 (+7.4% vs. +offm) Ribbon (no offm): 1 Filter size: 2949 (3072 in memory) 1 Filter size: 3381 (3584 in memory) 167 Filter size: 3701 (4096 in memory) Total on disk: 624,397 (-30.3% vs. Bloom) Total in memory: 690,688 (-30.0% vs. Bloom) Note that optimize_filters_for_memory is even more effective for Ribbon filter than for cache-local Bloom, because it can close the unused memory gap even tighter than Bloom filter, because of 16 byte increments for Ribbon vs. 64 byte increments for Bloom. Reviewed By: jay-zhuang Differential Revision: D25592970 Pulled By: pdillinger fbshipit-source-id: 606fdaa025bb790d7e9c21601e8ea86e10541912	4 years ago
Peter Dillinger	60af964372	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
Peter Dillinger	8b8a2e9f05	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
Peter Dillinger	746909ceda	Ribbon: InterleavedSolutionStorage (#7598) ... Summary: The core algorithms for InterleavedSolutionStorage and the implementation SerializableInterleavedSolution make Ribbon fast for filter queries. Example output from new unit test: Simple outside query, hot, incl hashing, ns/key: 117.796 Interleaved outside query, hot, incl hashing, ns/key: 42.2655 Bloom outside query, hot, incl hashing, ns/key: 24.0071 Also includes misc cleanup of previous Ribbon code and comments. Some TODOs and FIXMEs remain for futher work / investigation. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7598 Test Plan: unit tests included (integration work and tests coming later) Reviewed By: jay-zhuang Differential Revision: D24559209 Pulled By: pdillinger fbshipit-source-id: fea483cd354ba782aea3e806f2bc96e183d59441	4 years ago
Peter Dillinger	25d54c799c	Ribbon: initial (general) algorithms and basic unit test (#7491) ... Summary: This is intended as the first commit toward a near-optimal alternative to static Bloom filters for SSTs. Stephan Walzer and I have agreed upon the name "Ribbon" for a PHSF based on his linear system construction in "Efficient Gauss Elimination for Near-Quadratic Matrices with One Short Random Block per Row, with Applications" ("SGauss") and my much faster "on the fly" algorithm for gaussian elimination (or for this linear system, "banding"), which can be faster than peeling while also more compact and flexible. See util/ribbon_alg.h for more detailed introduction and background. RIBBON = Rapid Incremental Boolean Banding ON-the-fly This commit just adds generic (templatized) core algorithms and a basic unit test showing some features, including the ability to construct structures within 2.5% space overhead vs. information theoretic lower bound. (Compare to cache-local Bloom filter's ~50% space overhead -> ~30% reduction anticipated.) This commit does not include the storage scheme necessary to make queries fast, especially for filter queries, nor fractional "result bits", but there is some description already and those implementations will come soon. Nor does this commit add FilterPolicy support, for use in SST files, but that will also come soon. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7491 Reviewed By: jay-zhuang Differential Revision: D24517954 Pulled By: pdillinger fbshipit-source-id: 0119ee597e250d7e0edd38ada2ba50d755606fa7	4 years ago