Summary:
This change improves the performance of iterators doing long range scans (e.g. big/full table scans in MyRocks) by using readahead and prefetching additional data on each disk IO. This prefetching is automatically enabled on noticing more than 2 IOs for the same table file during iteration. The readahead size starts with 8KB and is exponentially increased on each additional sequential IO, up to a max of 256 KB. This helps in cutting down the number of IOs needed to complete the range scan.
Constraints:
- The prefetched data is stored by the OS in page cache. So this currently works only for non direct-reads use-cases i.e applications which use page cache. (Direct-I/O support will be enabled in a later PR).
- This gets currently enabled only when ReadOptions.readahead_size = 0 (which is the default value).
Thanks to siying for the original idea and implementation.
**Benchmarks:**
Data fill:
```
TEST_TMPDIR=/data/users/$USER/benchmarks/iter ./db_bench -benchmarks=fillrandom -num=1000000000 -compression_type="none" -level_compaction_dynamic_level_bytes
```
Do a long range scan: Seekrandom with large number of nexts
```
TEST_TMPDIR=/data/users/$USER/benchmarks/iter ./db_bench -benchmarks=seekrandom -duration=60 -num=1000000000 -use_existing_db -seek_nexts=10000 -statistics -histogram
```
Page cache was cleared before each experiment with the command:
```
sudo sh -c "echo 3 > /proc/sys/vm/drop_caches"
```
```
Before:
seekrandom : 34020.945 micros/op 29 ops/sec; 32.5 MB/s (1636 of 1999 found)
With this change:
seekrandom : 8726.912 micros/op 114 ops/sec; 126.8 MB/s (5702 of 6999 found)
```
~3.9X performance improvement.
Also verified with strace and gdb that the readahead size is increasing as expected.
```
strace -e readahead -f -T -t -p <db_bench process pid>
```
Closes https://github.com/facebook/rocksdb/pull/3282
Differential Revision: D6586477
Pulled By: sagar0
fbshipit-source-id: 8a118a0ed4594fbb7f5b1cafb242d7a4033cb58c
Summary:
This is a pre-cleaning up before a major block based table iterator refactoring. BlockBasedTable::NewDataBlockIterator() will always return BlockIter. This simplifies the logic and code and enable further refactoring and optimization.
Closes https://github.com/facebook/rocksdb/pull/3398
Differential Revision: D6780165
Pulled By: siying
fbshipit-source-id: 273f7dc896724f682c0118fb69a359d9cc4418b4
Summary:
Re-use metadata for reading Compression Dictionary on BlockBased
table open, this saves two reads from disk.
This helps to our 999 percentile in 5.6.1 where prefetch buffer is not present.
Closes https://github.com/facebook/rocksdb/pull/3354
Differential Revision: D6695753
Pulled By: ajkr
fbshipit-source-id: bb8acd9e9e66e65b89c548ab8940570ae360333c
Summary:
Divide ReadBlockContents() to multiple sub-functions. Maintaining the input and intermediate data in a new class BlockFetcher.
I hope in general it makes the code easier to maintain.
Another motivation to do it is to clearly divide the logic before file reading and after file reading. The refactor will help us evaluate how can we make I/O async in the future.
Closes https://github.com/facebook/rocksdb/pull/3244
Differential Revision: D6520983
Pulled By: siying
fbshipit-source-id: 338d90bc0338472d46be7a7682028dc9114b12e9
Summary:
The ASCII output is almost always useless to me as the first '\0' byte in the key or value causes it to stop printing. Since all characters are already surrounded by spaces, "\ 0" (how we display a backslash followed by a zero) and "\0" (how this PR displays a null terminator) are distinguishable. My assumption is the value of seeing all the bytes outweighs the value of the alignment we had before, where we always had one character followed by one space.
Closes https://github.com/facebook/rocksdb/pull/3203
Differential Revision: D6428651
Pulled By: ajkr
fbshipit-source-id: aafc978a51e9ea029cfe3e763e2bb0e1751b9ccf
Summary:
When using with compressed cache it is possible that the status is ok but the block is not actually added to the block cache. The patch takes this case into account.
Closes https://github.com/facebook/rocksdb/pull/2945
Differential Revision: D5937613
Pulled By: maysamyabandeh
fbshipit-source-id: 5428cf1115e5046b3d01ab78d26cb181122af4c6
Summary:
A user encountered segfault on the call to `CacheDependencies()`, probably because `NewIndexIterator()` failed before populating `*index_entry`. Let's avoid the call in that case.
Closes https://github.com/facebook/rocksdb/pull/2939
Differential Revision: D5928611
Pulled By: ajkr
fbshipit-source-id: 484be453dbb00e5e160e9c6a1bc933df7d80f574
Summary:
This is the warning that clang considers a bug and has been causing it to fail:
```
table/block_based_table_reader.cc:240:27: warning: Potential leak of memory pointed to by 'block.value'
for (; biter.Valid(); biter.Next()) {
^~~~~
```
Actually clang just doesn't have enough knowledge to statically determine it's safe. We can teach it using an assert.
Closes https://github.com/facebook/rocksdb/pull/2779
Differential Revision: D5691225
Pulled By: ajkr
fbshipit-source-id: 3f0d545bf44636953b30ee5243c63239e8f16d8e
Summary:
This is the continuation of https://github.com/facebook/rocksdb/pull/2661 for filter partitions. When pin_l0 is set (along with cache_xxx), then open table open the filter partitions are loaded into the cache and pinned there.
Closes https://github.com/facebook/rocksdb/pull/2766
Differential Revision: D5671098
Pulled By: maysamyabandeh
fbshipit-source-id: 174f24018f1d7f1129621e7380287b65b67d2115
Summary:
This fixes the existing logic for pinning l0 index partitions. The patch preloads the partitions into block cache and pin them if they belong to level 0 and pin_l0 is set.
The drawback is that it does many small IOs when preloading all the partitions into the cache is direct io is enabled. Working for a solution for that.
Closes https://github.com/facebook/rocksdb/pull/2661
Differential Revision: D5554010
Pulled By: maysamyabandeh
fbshipit-source-id: 1e6f32a3524d71355c77d4138516dcfb601ca7b2
Summary:
Right now, if direct I/O is enabled, prefetching the last 512KB cannot be applied, except compaction inputs or readahead is enabled for iterators. This can create a lot of I/O for HDD cases. To solve the problem, the 512KB is prefetched in block based table if direct I/O is enabled. The prefetched buffer is passed in totegher with random access file reader, so that we try to read from the buffer before reading from the file. This can be extended in the future to support flexible user iterator readahead too.
Closes https://github.com/facebook/rocksdb/pull/2708
Differential Revision: D5593091
Pulled By: siying
fbshipit-source-id: ee36ff6d8af11c312a2622272b21957a7b5c81e7
Summary:
We need a tool to check any sst file corruption in the db.
It will check all the sst files in current version and read all the blocks (data, meta, index) with checksum verification. If any verification fails, the function will return non-OK status.
Closes https://github.com/facebook/rocksdb/pull/2498
Differential Revision: D5324269
Pulled By: lightmark
fbshipit-source-id: 6f8a272008b722402a772acfc804524c9d1a483b
Summary:
Breaking commit: d12691b86f
In the above commit, I moved the `TableCache` cleanup logic from `Version` destructor into `PurgeObsoleteFiles`. I missed cleaning up `TableCache` entries for the current `Version` during DB destruction.
This PR adds that logic to `VersionSet` destructor. One unfortunate side effect is now we're potentially deleting `TableReader`s after `column_family_set_.reset()`, which means we can't call `BlockBasedTableReader::Close` a second time as the block cache might already be destroyed.
Closes https://github.com/facebook/rocksdb/pull/2662
Differential Revision: D5515108
Pulled By: ajkr
fbshipit-source-id: 2cb820e19aa813e0d258d17f76b2d7b6b7ee0b18
Summary:
This reverts the previous commit 1d7048c598, which broke the build.
Did a `git revert 1d7048c`.
Closes https://github.com/facebook/rocksdb/pull/2627
Differential Revision: D5476473
Pulled By: sagar0
fbshipit-source-id: 4756ff5c0dfc88c17eceb00e02c36176de728d06
Summary: This uses `clang-tidy` to comment out unused parameters (in functions, methods and lambdas) in fbcode. Cases that the tool failed to handle are fixed manually.
Reviewed By: igorsugak
Differential Revision: D5454343
fbshipit-source-id: 5dee339b4334e25e963891b519a5aa81fbf627b2
Summary:
When Partitioning index/filter is enabled the user might need to check the index block size as well as the top-level index size via sst_dump. This patch records i) number of partitions, ii) top-level index size and make it accessible through sst_dump. The number of partitions for filters is the same as that of indexes. The top-level index for filters has a similar size to top-level index for indexes, so it is not repeated.
Closes https://github.com/facebook/rocksdb/pull/2437
Differential Revision: D5224225
Pulled By: maysamyabandeh
fbshipit-source-id: 5324598c75793523aef1bb7ee225a5475e95a9cb
Summary:
BlockBasedTable::compaction_optimized_ is never used but can cause TSAN warning. Remove it.
Closes https://github.com/facebook/rocksdb/pull/2324
Differential Revision: D5085533
Pulled By: siying
fbshipit-source-id: 2feefce6806d559dfb4ab2989aa3db36752fe25d
Summary:
Based on my experience with linkbench, We should not skip loading bloom filter blocks when they are not available in block cache when using Iterator::Seek
Actually I am not sure why this behavior existed in the first place
Closes https://github.com/facebook/rocksdb/pull/2255
Differential Revision: D5010721
Pulled By: maysamyabandeh
fbshipit-source-id: 0af545a06ac4baeecb248706ec34d009c2480ca4
Summary:
Any non-raw-data dependent object must be destructed before the table
closes. There was a bug of not doing that for filter object. This patch
fixes the bug and adds a unit test to prevent such bugs in future.
Closes https://github.com/facebook/rocksdb/pull/2246
Differential Revision: D5001318
Pulled By: maysamyabandeh
fbshipit-source-id: 6d8772e58765485868094b92964da82ef9730b6d
Summary:
Now if we have iterate_upper_bound set, we continue read until get a key >= upper_bound. For a lot of cases that neighboring data blocks have a user key gap between them, our index key will be a user key in the middle to get a shorter size. For example, if we have blocks:
[a b c d][f g h]
Then the index key for the first block will be 'e'.
then if upper bound is any key between 'd' and 'e', for example, d1, d2, ..., d99999999999, we don't have to read the second block and also know that we have done our iteration by reaching the last key that smaller the upper bound already.
This diff can reduce RA in most cases.
Closes https://github.com/facebook/rocksdb/pull/2239
Differential Revision: D4990693
Pulled By: lightmark
fbshipit-source-id: ab30ea2e3c6edf3fddd5efed3c34fcf7739827ff
Summary:
Some filters such as partitioned filter have pointers to the table for which they are created. Therefore is they are stored in the block cache, the should be forcibly erased from block cache before closing the table, which would result into deleting the object. Otherwise the destructor will be called later when the cache is lazily erasing the object, which having the parent table no longer existent it could result into undefined behavior.
Update: there will be still cases the filter is not removed from the cache since the table has not kept a pointer to the cache handle to be able to forcibly release it later. We make sure that the filter destructor does not access the table pointer to get around such cases.
Closes https://github.com/facebook/rocksdb/pull/2207
Differential Revision: D4941591
Pulled By: maysamyabandeh
fbshipit-source-id: 56fbab2a11cf447e1aa67caa30b58d7bd7ce5bbd
Summary:
prefetch some data from the end of the file for each compaction to reduce IO.
Closes https://github.com/facebook/rocksdb/pull/2149
Differential Revision: D4880576
Pulled By: lightmark
fbshipit-source-id: aa767cd1afc84c541837fbf1ad6c0d45b34d3932
Summary:
Move some files under util/ to new directories env/, monitoring/ options/ and cache/
Closes https://github.com/facebook/rocksdb/pull/2090
Differential Revision: D4833681
Pulled By: siying
fbshipit-source-id: 2fd8bef
Summary:
Fixes#1961 which causes a segfault when filter_policy is nullptr and both
pin_l0_filter_and_index_blocks_in_cache/cache_index_and_filter_blocks
are set.
Closes https://github.com/facebook/rocksdb/pull/2029
Differential Revision: D4764862
Pulled By: maysamyabandeh
fbshipit-source-id: 05bd695
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:
Partition Index blocks and use a Partition-index as a 2nd level index.
The two-level index can be used by setting
BlockBasedTableOptions::kTwoLevelIndexSearch as the index type and
configuring BlockBasedTableOptions::index_per_partition
t15539501
Closes https://github.com/facebook/rocksdb/pull/1814
Differential Revision: D4473535
Pulled By: maysamyabandeh
fbshipit-source-id: bffb87e
Summary:
I added the Cache::Ref() function a couple weeks ago (#1761) to make this feature possible. Like other meta-blocks, rep_->range_del_entry holds a cache handle to pin the range deletion block in uncompressed block cache for the duration of the table reader's lifetime. We can reuse this cache handle to create an iterator over this meta-block without any cache lookup. Ref() is used to increment the cache handle's refcount in case the returned iterator outlives the table reader.
Closes https://github.com/facebook/rocksdb/pull/1801
Differential Revision: D4458782
Pulled By: ajkr
fbshipit-source-id: 2883f10
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:
Cleanable objects will perform the registered cleanups when
they are destructed. We however rather to delay this cleaning like when
we are gathering the merge operands. Current approach is to create the
Cleanable object on heap (instead of on stack) and delay deleting it.
By allowing Cleanables to delegate their cleanups to another cleanable
object we can delay the cleaning without however the need to craete the
cleanable object on heap and keeping it around. This patch applies this
technique for the cleanups of BlockIter and shows improved performance
for some in-memory benchmarks:
+1.8% for merge worklaod, +6.4% for non-merge workload when the merge
operator is specified.
https://our.intern.facebook.com/intern/tasks?t=15168163
Non-merge benchmark:
TEST_TMPDIR=/dev/shm/v100nocomp/ ./db_bench --benchmarks=fillrandom
--num=1000000 -value_size=100 -compression_type=none
Reading random with no merge operator specified:
TEST_TMPDIR=/dev/shm/v100nocomp/ ./db_bench
--benchmarks="read
Closes https://github.com/facebook/rocksdb/pull/1711
Differential Revision: D4361163
Pulled By: maysamyabandeh
fbshipit-source-id: 9801e07
Summary:
- Made RangeDelAggregator's InternalKeyComparator member a reference-to-const so we don't need to copy-construct it. Also added InternalKeyComparator to ImmutableCFOptions so we don't need to construct one for each DBIter.
- Made MemTable::NewRangeTombstoneIterator and the table readers' NewRangeTombstoneIterator() functions return nullptr instead of NewEmptyInternalIterator to avoid the allocation. Updated callers accordingly.
Closes https://github.com/facebook/rocksdb/pull/1548
Differential Revision: D4208169
Pulled By: ajkr
fbshipit-source-id: 2fd65cf
Summary:
Change DumpTable() so we can see the range deletion meta-block.
Closes https://github.com/facebook/rocksdb/pull/1505
Differential Revision: D4172227
Pulled By: ajkr
fbshipit-source-id: ae35665
Summary:
This handles two issues: (1) range deletion iterator sometimes outlives
the table reader that created it, in which case the block must not be destroyed
during table reader destruction; and (2) we prefer to read these range tombstone
meta-blocks from file fewer times.
- Extracted cache-populating logic from NewDataBlockIterator() into a separate function: MaybeLoadDataBlockToCache()
- Use MaybeLoadDataBlockToCache() to load range deletion meta-block and pin it through the reader's lifetime. This code reuse works since range deletion meta-block has same format as data blocks.
- Use NewDataBlockIterator() to create range deletion iterators, which uses block cache if enabled, otherwise reads the block from file. Either way, the underlying block won't disappear until after the iterator is destroyed.
Closes https://github.com/facebook/rocksdb/pull/1459
Differential Revision: D4123175
Pulled By: ajkr
fbshipit-source-id: 8f64281
Summary:
reland https://reviews.facebook.net/D62523
- Update SstFileWriter to include a property for a global sequence number in the SST file `rocksdb.external_sst_file.global_seqno`
- Update TableProperties to be aware of the offset of each property in the file
- Update BlockBasedTableReader and Block to be able to honor the sequence number in `rocksdb.external_sst_file.global_seqno` property and use it to overwrite all sequence number in the file
Something worth mentioning is that we don't update the seqno in the index block since and when doing a binary search, the reason for that is that it's guaranteed that SST files with global seqno will have only one user_key and each key will have seqno=0 encoded in it, This mean that this key is greater than any other key with seqno> 0. That mean that we can actually keep the current logic for these blocks
Test Plan: unit tests
Reviewers: sdong, yhchiang
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D65211
Summary:
- Update SstFileWriter to include a property for a global sequence number in the SST file `rocksdb.external_sst_file.global_seqno`
- Update TableProperties to be aware of the offset of each property in the file
- Update BlockBasedTableReader and Block to be able to honor the sequence number in `rocksdb.external_sst_file.global_seqno` property and use it to overwrite all sequence number in the file
Something worth mentioning is that we don't update the seqno in the index block since and when doing a binary search, the reason for that is that it's guaranteed that SST files with global seqno will have only one user_key and each key will have seqno=0 encoded in it, This mean that this key is greater than any other key with seqno> 0. That mean that we can actually keep the current logic for these blocks
Test Plan: unit tests
Reviewers: andrewkr, yhchiang, yiwu, sdong
Reviewed By: sdong
Subscribers: hcz, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D62523
Sagar Vemuri