Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
Summary:
- If block cache disabled or not used for meta-blocks, `BlockBasedTableReader::Rep::uncompression_dict` owns the `UncompressionDict`. It is preloaded during `PrefetchIndexAndFilterBlocks`.
- If block cache is enabled and used for meta-blocks, block cache owns the `UncompressionDict`, which holds dictionary and digested dictionary when needed. It is never prefetched though there is a TODO for this in the code. The cache key is simply the compression dictionary block handle.
- New stats for compression dictionary accesses in block cache: "BLOCK_CACHE_COMPRESSION_DICT_*" and "compression_dict_block_read_count"
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4881
Differential Revision: D13663801
Pulled By: ajkr
fbshipit-source-id: bdcc54044e180855cdcc57639b493b0e016c9a3f
Summary:
Previously, range tombstones were accumulated from every level, which
was necessary if a range tombstone in a higher level covered a key in a lower
level. However, RangeDelAggregator::AddTombstones's complexity is based on
the number of tombstones that are currently stored in it, which is wasteful in
the Get case, where we only need to know the highest sequence number of range
tombstones that cover the key from higher levels, and compute the highest covering
sequence number at the current level. This change introduces this optimization, and
removes the use of RangeDelAggregator from the Get path.
In the benchmark results, the following command was used to initialize the database:
```
./db_bench -db=/dev/shm/5k-rts -use_existing_db=false -benchmarks=filluniquerandom -write_buffer_size=1048576 -compression_type=lz4 -target_file_size_base=1048576 -max_bytes_for_level_base=4194304 -value_size=112 -key_size=16 -block_size=4096 -level_compaction_dynamic_level_bytes=true -num=5000000 -max_background_jobs=12 -benchmark_write_rate_limit=20971520 -range_tombstone_width=100 -writes_per_range_tombstone=100 -max_num_range_tombstones=50000 -bloom_bits=8
```
...and the following command was used to measure read throughput:
```
./db_bench -db=/dev/shm/5k-rts/ -use_existing_db=true -benchmarks=readrandom -disable_auto_compactions=true -num=5000000 -reads=100000 -threads=32
```
The filluniquerandom command was only run once, and the resulting database was used
to measure read performance before and after the PR. Both binaries were compiled with
`DEBUG_LEVEL=0`.
Readrandom results before PR:
```
readrandom : 4.544 micros/op 220090 ops/sec; 16.9 MB/s (63103 of 100000 found)
```
Readrandom results after PR:
```
readrandom : 11.147 micros/op 89707 ops/sec; 6.9 MB/s (63103 of 100000 found)
```
So it's actually slower right now, but this PR paves the way for future optimizations (see #4493).
----
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4449
Differential Revision: D10370575
Pulled By: abhimadan
fbshipit-source-id: 9a2e152be1ef36969055c0e9eb4beb0d96c11f4d
Summary:
Currently in `Version::Get` when reporting ticker stats stored in `GetContext`, there is a big for-loop through all `Ticker` which adds unnecessary cost to overall CPU usage. We can optimize by storing only ticker values that are used in `Get()` calls in a new struct `GetContextStats` since only a small fraction of all tickers are used in `Get()` calls. For comparison, with the new approach we only need to visit 17 values while old approach will require visiting 100+ `Ticker`
Pull Request resolved: https://github.com/facebook/rocksdb/pull/3490
Differential Revision: D6969154
Pulled By: miasantreble
fbshipit-source-id: fc27072965a3a94125a3e6883d20dafcf5b84029
Summary:
This is implemented by extending ReadCallback with another function `MaxUnpreparedSequenceNumber` which returns the largest visible sequence number for the current transaction, if there is uncommitted data written to DB. Otherwise, it returns zero, indicating no uncommitted data.
There are the places where reads had to be modified.
- Get and Seek/Next was just updated to seek to max(snapshot_seq, MaxUnpreparedSequenceNumber()) instead, and iterate until a key was visible.
- Prev did not need need updates since it did not use the Seek to sequence number optimization. Assuming that locks were held when writing unprepared keys, and ValidateSnapshot runs, there should only be committed keys and unprepared keys of the current transaction, all of which are visible. Prev will simply iterate to get the last visible key.
- Reseeking to skip keys optimization was also disabled for write unprepared, since it's possible to hit the max_skip condition even while reseeking. There needs to be some way to resolve infinite looping in this case.
Closes https://github.com/facebook/rocksdb/pull/3955
Differential Revision: D8286688
Pulled By: lth
fbshipit-source-id: 25e42f47fdeb5f7accea0f4fd350ef35198caafe
Summary:
Adds two stats to allow us measuring the false positive rate of full filters:
- The total count of positives: rocksdb.bloom.filter.full.positive
- The total count of true positives: rocksdb.bloom.filter.full.true.positive
Not the term "full" in the stat name to indicate that they are meaningful in full filters. block-based filters are to be deprecated soon and supporting it is not worth the the additional cost of if-then-else branches.
Closes#3680
Tested by:
$ ./db_bench -benchmarks=fillrandom -db /dev/shm/rocksdb-tmpdb --num=1000000 -bloom_bits=10
$ ./db_bench -benchmarks="readwhilewriting" -db /dev/shm/rocksdb-tmpdb --statistics -bloom_bits=10 --duration=60 --num=2000000 --use_existing_db 2>&1 > /tmp/full.log
$ grep filter.full /tmp/full.log
rocksdb.bloom.filter.full.positive COUNT : 3628593
rocksdb.bloom.filter.full.true.positive COUNT : 3536026
which gives the false positive rate of 2.5%
Closes https://github.com/facebook/rocksdb/pull/3681
Differential Revision: D7517570
Pulled By: maysamyabandeh
fbshipit-source-id: 630ab1a473afdce404916d297035b6318de4c052
Summary:
Add kTypeBlobIndex value type, which will be used by blob db only, to insert a (key, blob_offset) KV pair. The purpose is to
1. Make it possible to open existing rocksdb instance as blob db. Existing value will be of kTypeIndex type, while value inserted by blob db will be of kTypeBlobIndex.
2. Make rocksdb able to detect if the db contains value written by blob db, if so return error.
3. Make it possible to have blob db optionally store value in SST file (with kTypeValue type) or as a blob value (with kTypeBlobIndex type).
The root db (DBImpl) basically pretended kTypeBlobIndex are normal value on write. On Get if is_blob is provided, return whether the value read is of kTypeBlobIndex type, or return Status::NotSupported() status if is_blob is not provided. On scan allow_blob flag is pass and if the flag is true, return wether the value is of kTypeBlobIndex type via iter->IsBlob().
Changes on blob db side will be in a separate patch.
Closes https://github.com/facebook/rocksdb/pull/2886
Differential Revision: D5838431
Pulled By: yiwu-arbug
fbshipit-source-id: 3c5306c62bc13bb11abc03422ec5cbcea1203cca
Summary:
This patch instruments the read path to verify each read value against an optional ReadCallback class. If the value is rejected, the reader moves on to the next value. The WritePreparedTxn makes use of this feature to skip sequence numbers that are not in the read snapshot.
Closes https://github.com/facebook/rocksdb/pull/2850
Differential Revision: D5787375
Pulled By: maysamyabandeh
fbshipit-source-id: 49d808b3062ab35e7ae98ad388f659757794184c
Summary:
This patch enables using PinnableSlice for RowCache, changes include
not releasing the cache handle immediately after lookup in TableCache::Get, instead pass a Cleanble function which does Cache::RleaseHandle.
Closes https://github.com/facebook/rocksdb/pull/2492
Differential Revision: D5316216
Pulled By: maysamyabandeh
fbshipit-source-id: d2a684bd7e4ba73772f762e58a82b5f4fbd5d362
Summary:
We estimate number of reads per SST files, by updating the counter per file in sampled read requests. This information can later be used to trigger compactions to improve read performacne.
Closes https://github.com/facebook/rocksdb/pull/2417
Differential Revision: D5193528
Pulled By: siying
fbshipit-source-id: b4241c5ad0eaf444b61afb53f8e6290d9f5da2df
Summary:
PinnableSlice
Summary:
Currently the point lookup values are copied to a string provided by the
user. This incures an extra memcpy cost. This patch allows doing point lookup
via a PinnableSlice which pins the source memory location (instead of
copying their content) and releases them after the content is consumed
by the user. The old API of Get(string) is translated to the new API
underneath.
Here is the summary for improvements:
value 100 byte: 1.8% regular, 1.2% merge values
value 1k byte: 11.5% regular, 7.5% merge values
value 10k byte: 26% regular, 29.9% merge values
The improvement for merge could be more if we extend this approach to
pin the merge output and delay the full merge operation until the user
actually needs it. We have put that for future work.
PS:
Sometimes we observe a small decrease in performance when switching from
t5452014 to this patch but with the old Get(string) API. The d
Closes https://github.com/facebook/rocksdb/pull/1756
Differential Revision: D4391738
Pulled By: maysamyabandeh
fbshipit-source-id: 6f3edd3
Summary:
Currently the point lookup values are copied to a string provided by the user.
This incures an extra memcpy cost. This patch allows doing point lookup
via a PinnableSlice which pins the source memory location (instead of
copying their content) and releases them after the content is consumed
by the user. The old API of Get(string) is translated to the new API
underneath.
Here is the summary for improvements:
1. value 100 byte: 1.8% regular, 1.2% merge values
2. value 1k byte: 11.5% regular, 7.5% merge values
3. value 10k byte: 26% regular, 29.9% merge values
The improvement for merge could be more if we extend this approach to
pin the merge output and delay the full merge operation until the user
actually needs it. We have put that for future work.
PS:
Sometimes we observe a small decrease in performance when switching from
t5452014 to this patch but with the old Get(string) API. The difference
is a little and could be noise. More importantly it is safely
cancelled
Closes https://github.com/facebook/rocksdb/pull/1732
Differential Revision: D4374613
Pulled By: maysamyabandeh
fbshipit-source-id: a077f1a
Summary:
Previously we used TableCache::NewIterator() for multiple purposes (data
block iterator and range deletion iterator), and returned non-ok status in
the data block iterator. In one case where the caller only used the range
deletion block iterator (9e7cf3469b/db/version_set.cc (L965-L973)),
we didn't check/free the data block iterator containing non-ok status, which
caused a valgrind error.
So, this diff decouples creation of data block and range deletion block iterators,
and updates the callers accordingly. Both functions can return non-ok status
in an InternalIterator. Since the non-ok status is returned in an iterator that the
callers will definitely use, it should be more usable/less error-prone.
Closes https://github.com/facebook/rocksdb/pull/1513
Differential Revision: D4181423
Pulled By: ajkr
fbshipit-source-id: 835b8f5
Summary:
During Get()/MultiGet(), build up a RangeDelAggregator with range
tombstones as we search through live memtable, immutable memtables, and
SST files. This aggregator is then used by memtable.cc's SaveValue() and
GetContext::SaveValue() to check whether keys are covered.
added tests for Get on memtables/files; end-to-end tests mainly in https://reviews.facebook.net/D64761
Closes https://github.com/facebook/rocksdb/pull/1456
Differential Revision: D4111271
Pulled By: ajkr
fbshipit-source-id: 6e388d4
Summary:
Currently, transactions can fail even if there is no actual write conflict. This is due to relying on only the memtables to check for write-conflicts. Users have to tune memtable settings to try to avoid this, but it's hard to figure out exactly how to tune these settings.
With this diff, TransactionDB will use both memtables and SST files to determine if there are any write conflicts. This relies on the fact that BlockBasedTable stores sequence numbers for all writes that happen after any open snapshot. Also, D50295 is needed to prevent SingleDelete from disappearing writes (the TODOs in this test code will be fixed once the other diff is approved and merged).
Note that Optimistic transactions will still rely on tuning memtable settings as we do not want to read from SST while on the write thread. Also, memtable settings can still be used to reduce how often TransactionDB needs to read SST files.
Test Plan: unit tests, db bench
Reviewers: rven, yhchiang, kradhakrishnan, IslamAbdelRahman, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb, yoshinorim
Differential Revision: https://reviews.facebook.net/D50475
Summary:
Implementation of a table-level row cache.
It only caches point queries done through the `DB::Get` interface, queries done through the `Iterator` interface will completely skip the cache.
Supports snapshots and merge operations.
Test Plan: Ran `make valgrind_check commit-prereq`
Reviewers: igor, philipp, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D39849
Summary:
We have addded new stats and perf_context for measuring the merge and filter operation time consumption.
We have bounded all the merge operations within the GUARD statment and collected the total time for these operations in the DB.
Test Plan: WIP
Reviewers: rven, yhchiang, kradhakrishnan, igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D34377
Summary:
Intead of passing callback function pointer and its arg on Table::Get()
interface, passing GetContext. This makes the interface cleaner and
possible better perf. Also adding a fast pass for SaveValue()
Test Plan: make all check
Reviewers: igor, yhchiang, sdong
Reviewed By: sdong
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D24057
anand76