Summary:
This PR extends the improvements in #3282 to also work when using Direct IO.
We see **4.5X performance improvement** in seekrandom benchmark doing long range scans, when using direct reads, on flash.
**Description:**
This change improves the performance of iterators doing long range scans (e.g. big/full index or table scans in MyRocks) by using readahead and prefetching additional data on each disk IO, and storing in a local buffer. 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.
**Implementation Details:**
- Used `FilePrefetchBuffer` as the underlying buffer to store the readahead data. `FilePrefetchBuffer` can now take file_reader, readahead_size and max_readahead_size as input to the constructor, and automatically do readahead.
- `FilePrefetchBuffer::TryReadFromCache` can now call `FilePrefetchBuffer::Prefetch` if readahead is enabled.
- `AlignedBuffer` (which is the underlying store for `FilePrefetchBuffer`) now takes a few additional args in `AlignedBuffer::AllocateNewBuffer` to allow copying data from the old buffer.
- Made sure not to re-read partial chunks of data that were already available in the buffer, from device again.
- Fixed a couple of cases where `AlignedBuffer::cursize_` was not being properly kept up-to-date.
**Constraints:**
- Similar to #3282, this gets currently enabled only when ReadOptions.readahead_size = 0 (which is the default value).
- Since the prefetched data is stored in a temporary buffer allocated on heap, this could increase the memory usage if you have many iterators doing long range scans simultaneously.
- Enabled only for user reads, and disabled for compactions. Compaction reads are controlled by the options `use_direct_io_for_flush_and_compaction` and `compaction_readahead_size`, and the current feature takes precautions not to mess with them.
**Benchmarks:**
I used the same benchmark as used in #3282.
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 -use_direct_reads -duration=60 -num=1000000000 -use_existing_db -seek_nexts=10000 -statistics -histogram
```
```
Before:
seekrandom : 37939.906 micros/op 26 ops/sec; 29.2 MB/s (1636 of 1999 found)
With this change:
seekrandom : 8527.720 micros/op 117 ops/sec; 129.7 MB/s (6530 of 7999 found)
```
~4.5X perf improvement. Taken on an average of 3 runs.
Closes https://github.com/facebook/rocksdb/pull/3884
Differential Revision: D8082143
Pulled By: sagar0
fbshipit-source-id: 4d7a8561cbac03478663713df4d31ad2620253bb
Summary:
Previously `DBOptions::use_direct_io_for_flush_and_compaction=true` combined with `DBOptions::use_direct_reads=false` could cause RocksDB to simultaneously read from two file descriptors for the same file, where background reads used direct I/O and foreground reads used buffered I/O. Our measurements found this mixed-mode I/O negatively impacted foreground read perf, compared to when only buffered I/O was used.
This PR makes the mixed-mode I/O situation impossible by repurposing `DBOptions::use_direct_io_for_flush_and_compaction` to only apply to background writes, and `DBOptions::use_direct_reads` to apply to all reads. There is no risk of direct background direct writes happening simultaneously with buffered reads since we never read from and write to the same file simultaneously.
Closes https://github.com/facebook/rocksdb/pull/3829
Differential Revision: D7915443
Pulled By: ajkr
fbshipit-source-id: 78bcbf276449b7e7766ab6b0db246f789fb1b279
Summary:
`ReadaheadRandomAccessFile` had an unwritten assumption, which was that its wrapped file's `Read()` function always copies into the provided scratch buffer. Actually this was not true when the wrapped file was `PosixMmapReadableFile`, whose `Read()` implementation does no copying and instead returns a `Slice` pointing directly into the `mmap`'d memory region. This PR:
- prevents `ReadaheadRandomAccessFile` from ever wrapping mmap readable files
- adds an assert for the assumption `ReadaheadRandomAccessFile` makes about the wrapped file's use of scratch buffer
Closes https://github.com/facebook/rocksdb/pull/3813
Differential Revision: D7891513
Pulled By: ajkr
fbshipit-source-id: dc64a55222d6af280c39a1852ee39e9e9d7cde7d
Summary:
Provide a block_align option in BlockBasedTableOptions to allow
alignment of SST file data blocks. This will avoid higher
IOPS/throughput load due to < 4KB data blocks spanning 2 4KB pages.
When this option is set to true, the block alignment is set to lower of
block size and 4KB.
Closes https://github.com/facebook/rocksdb/pull/3502
Differential Revision: D7400897
Pulled By: anand1976
fbshipit-source-id: 04cc3bd144e88e3431a4f97604e63ad7a0f06d44
Summary:
In attempting to build a static lib for use in iOS, I ran in to lots of type errors between uint64_t and size_t. This PR contains the changes I made to get `TARGET_OS=IOS make static_lib` to succeed while also getting Xcode to build successfully with the resulting `librocksdb.a` library imported.
This also compiles for me on macOS and tests fine, but I'm really not sure if I made the correct decisions about where to `static_cast` and where to change types.
Also up for discussion: is iOS worth supporting? Getting the static lib is just part one, we aren't providing any bridging headers or wrappers like the ObjectiveRocks project, it won't be a great experience.
Closes https://github.com/facebook/rocksdb/pull/3503
Differential Revision: D7106457
Pulled By: gfosco
fbshipit-source-id: 82ac2073de7e1f09b91f6b4faea91d18bd311f8e
Summary:
We don't do fsync() after truncate in direct I/O writeable file (in fact we don't do any fsync ever). This can cause metadata not persistent to disk after the file is generated. We call it instead.
Closes https://github.com/facebook/rocksdb/pull/3500
Differential Revision: D6981482
Pulled By: siying
fbshipit-source-id: 7e2b591b7e5dd1b96fc0775515b8b9e6092980ef
Summary:
`ReadaheadRandomAccessFile` is used by iterators for file reads in several cases, like in compaction when `compaction_readahead_size > 0` or `use_direct_io_for_flush_and_compaction == true`, or in user iterator when `ReadOptions::readahead_size > 0`. `ReadaheadRandomAccessFile` maintains an internal buffer for readahead data. It assumes that, if the buffer's length is less than `ReadaheadRandomAccessFile::readahead_size_`, which is fixed in the constructor, then EOF has been reached so it doesn't try reading further.
Recently, d938226af4 started calling `RandomAccessFile::Prefetch` with various lengths: 8KB, 16KB, etc. When the `RandomAccessFile` is a `ReadaheadRandomAccessFile`, it triggers the above condition and incorrectly determines EOF. If a block is partially in the readahead buffer and EOF is incorrectly decided, the result is a truncated data block.
The problem is reproducible:
```
TEST_TMPDIR=/data/compaction_bench ./db_bench -benchmarks=fillrandom -write_buffer_size=1048576 -target_file_size_base=1048576 -block_size=18384 -use_direct_io_for_flush_and_compaction=true
...
put error: Corruption: truncated block read from /data/compaction_bench/dbbench/000014.sst offset 20245, expected 10143 bytes, got 8427
```
Closes https://github.com/facebook/rocksdb/pull/3454
Differential Revision: D6869405
Pulled By: ajkr
fbshipit-source-id: 87001c299e7600a37c0dcccbd0368e0954c929cf
Summary:
FilePrefetchBuffer::Prefetch is currently rounds the offset up which does not fit its new use cases in prefetching index/filter blocks, as it would skips over some the offsets that were requested to be prefetched. This patch rounds down instead.
Fixes#3180
Closes https://github.com/facebook/rocksdb/pull/3413
Differential Revision: D6816392
Pulled By: maysamyabandeh
fbshipit-source-id: 3aaeaf59c55d72b61dacfae6d4a8e65eccb3c553
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:
Allow users to rate limit background work based on read bytes, written bytes, or sum of read and written bytes. Support these by changing the RateLimiter API, so no additional options were needed.
Closes https://github.com/facebook/rocksdb/pull/2433
Differential Revision: D5216946
Pulled By: ajkr
fbshipit-source-id: aec57a8357dbb4bfde2003261094d786d94f724e
Summary:
Currently for direct I/O, the large maximum buffer is always allocated. This will be wasteful if users flush the data in much smaller chunks. This diff fix this by changing the behavior of incremental buffer works. When we enlarge buffer, we try to copy the existing data in the buffer to the enlarged buffer, rather than flush the buffer first. This can make sure that no extra I/O is introduced because of buffer enlargement.
Closes https://github.com/facebook/rocksdb/pull/2403
Differential Revision: D5178403
Pulled By: siying
fbshipit-source-id: a8fe1e7304bdb8cab2973340022fe80ff83449fd
Summary:
unity test will fail even if we have the same function names in different anonymous namespaces in different files.
Closes https://github.com/facebook/rocksdb/pull/2321
Differential Revision: D5083783
Pulled By: lightmark
fbshipit-source-id: 1347aaf866900af30d23cdd4f29c1b96f17352af
Summary:
Remove double buffering on RandomRead on Windows.
With more logic appear in file reader/write Read no longer
obeys forwarding calls to Windows implementation.
Previously direct_io (unbuffered) was only available on Windows
but now is supported as generic.
We remove intermediate buffering on Windows.
Remove random_access_max_buffer_size option which was windows specific.
Non-zero values for that opton introduced unnecessary lock contention.
Remove Env::EnableReadAhead(), Env::ShouldForwardRawRequest() that are
no longer necessary.
Add aligned buffer reads for cases when requested reads exceed read ahead size.
Closes https://github.com/facebook/rocksdb/pull/2105
Differential Revision: D4847770
Pulled By: siying
fbshipit-source-id: 8ab48f8e854ab498a4fd398a6934859792a2788f
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:
Users usually set readahead buffer to a multiple of 4k, more than that, usually a multiple of blocks.
So previously we set real buffer size 512 * n + 4k, which may introduce an additional block reading.
Closes https://github.com/facebook/rocksdb/pull/2138
Differential Revision: D4871504
Pulled By: lightmark
fbshipit-source-id: b070faa51d92e976e8e8468c00692699e585e243
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:
in buffered io, the filesize_ is the real size.
Closes https://github.com/facebook/rocksdb/pull/1991
Differential Revision: D4711433
Pulled By: lightmark
fbshipit-source-id: ad604b9
Summary:
This is the metric I plan to use for adaptive rate limiting. The statistics are updated only if the rate limiter is drained by flush or compaction. I believe (but am not certain) that this is the normal case.
The Statistics object is passed in RateLimiter::Request() to avoid requiring changes to client code, which would've been necessary if we passed it in the RateLimiter constructor.
Closes https://github.com/facebook/rocksdb/pull/1946
Differential Revision: D4646489
Pulled By: ajkr
fbshipit-source-id: d8e0161
Summary:
Missing this function will cause RandomAccessFileReader not doing alignment in Direct IO mode, which introduce an IOError: invalid argument.
Closes https://github.com/facebook/rocksdb/pull/1900
Differential Revision: D4601261
Pulled By: lightmark
fbshipit-source-id: c3eadf1
Summary:
fix lite bugs
disable direct io in lite mode
Closes https://github.com/facebook/rocksdb/pull/1870
Differential Revision: D4559866
Pulled By: yiwu-arbug
fbshipit-source-id: 3761c51
Summary:
also change variable name `direct_io_` to `use_direct_io_` in WritableFile to make it consistent with read path.
Closes https://github.com/facebook/rocksdb/pull/1770
Differential Revision: D4416435
Pulled By: lightmark
fbshipit-source-id: 4143c53
Summary:
direct IO reads refactoring
remove unnecessary classes and unified interfaces
tested with db_bench
need more change for options and ON/OFF for different files.
Since disabled is default, it should be fine now
Closes https://github.com/facebook/rocksdb/pull/1636
Differential Revision: D4307189
Pulled By: lightmark
fbshipit-source-id: 6991e22
in 64-bit.
Currently, a signed off_t type is being used for the following
interfaces for both offset and the length in bytes:
* `Allocate`
* `RangeSync`
On Linux `off_t` is automatically either 32 or 64-bit depending on
the platform. On Windows it is always a 32-bit signed long which
limits file access and in particular space pre-allocation
to effectively 2 Gb.
Proposal is to replace off_t with uint64_t as a portable type
always access files with 64-bit interfaces.
May need to modify posix code but lack resources to test it.
in 64-bit.
Currently, a signed off_t type is being used for the following
interfaces for both offset and the length in bytes:
* `Allocate`
* `RangeSync`
On Linux `off_t` is automatically either 32 or 64-bit depending on
the platform. On Windows it is always a 32-bit signed long which
limits file access and in particular space pre-allocation
to effectively 2 Gb.
Proposal is to replace off_t with uint64_t as a portable type
always access files with 64-bit interfaces.
May need to modify posix code but lack resources to test it.
Summary: Run "make format" for some recent commits.
Test Plan: Build and run tests
Reviewers: IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D49707
Summary: It looks like WritableFileWriter::Append() was returning OK() even when there is an error
Test Plan: make check
Reviewers: sdong, yhchiang, anthony, rven, kradhakrishnan, igor
Reviewed By: igor
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D49569
introduce a new DBOption random_access_max_buffer_size to limit
the size of the random access buffer used for unbuffered access.
Implement read ahead buffering when enabled.
To that effect propagate compaction_readahead_size and the new option
to the env options to make it available for the implementation.
Add Hint() override so SetupForCompaction() call would call Hint()
readahead can now be setup from both Hint() and EnableReadAhead()
Add new option random_access_max_buffer_size support
db_bench, options_helper to make it string parsable
and the unit test.
Summary: We don't yet have a CI build for iOS, so our iOS compile gets broken sometimes. Most of the errors are from assumption that size_t is 64-bit, while it's actually 32-bit on some (all?) iOS platforms. This diff fixes the compile.
Test Plan:
TARGET_OS=IOS make static_lib
Observe there are no warnings
Reviewers: sdong, anthony, IslamAbdelRahman, kradhakrishnan, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D49029
Summary:
Add kill points in:
1. after creating a file
2. before writing a manifest record
3. before syncing manifest
4. before creating a new current file
5. after creating a new current file
Test Plan: Run all current tests.
Reviewers: yhchiang, igor, anthony, IslamAbdelRahman, rven, kradhakrishnan
Reviewed By: kradhakrishnan
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D48855
Summary:
Give a name for every kill point, and allow users to disable some kill points based on prefixes. The kill points can be passed by db_stress through a command line paramter. This provides a way for users to boost the chance of triggering low frequency kill points
This allow follow up changes in crash test scripts to improve crash test coverage.
Test Plan:
Manually run db_stress with variable values of --kill_random_test and --kill_prefix_blacklist. Like this:
--kill_random_test=2 --kill_prefix_blacklist=Posix,WritableFileWriter::Append,WritableFileWriter::WriteBuffered,WritableFileWriter::Sync
Reviewers: igor, kradhakrishnan, rven, IslamAbdelRahman, yhchiang
Reviewed By: yhchiang
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D48735
Summary. A change https://reviews.facebook.net/differential/diff/224721/
Has attempted to move common functionality out of platform dependent
code to a new facility called file_reader_writer.
This includes:
- perf counters
- Buffering
- RateLimiting
However, the change did not attempt to refactor Windows code.
To mitigate, we introduce new quering interfaces such as UseOSBuffer(),
GetRequiredBufferAlignment() and ReaderWriterForward()
for pure forwarding where required.
Introduce WritableFile got a new method Truncate(). This is to communicate
to the file as to how much data it has on close.
- When space is pre-allocated on Linux it is filled with zeros implicitly,
no such thing exist on Windows so we must truncate file on close.
- When operating in unbuffered mode the last page is filled with zeros but we still want to truncate.
Previously, Close() would take care of it but now buffer management is shifted to the wrappers and the file has
no idea about the file true size.
This means that Close() on the wrapper level must always include
Truncate() as well as wrapper __dtor should call Close() and
against double Close().
Move buffered/unbuffered write logic to the wrapper.
Utilize Aligned buffer class.
Adjust tests and implement Truncate() where necessary.
Come up with reasonable defaults for new virtual interfaces.
Forward calls for RandomAccessReadAhead class to avoid double
buffering and locking (double locking in unbuffered mode on WIndows).
Summary:
This patch adds the helper functions and variables to allow a backend
implementing WritableFile to support direct IO when persisting a
memtable.
Test Plan:
Since there is no upstream implementation of WritableFile supporting
direct IO, the new behavior is disabled.
Tests should be provided by the backend implementing WritableFile.
Summary:
ReadaheadRandomAccessFile acts as a transparent layer on top of RandomAccessFile. When a Read() request is issued, it issues a much bigger request to the OS and caches the result. When a new request comes in and we already have the data cached, it doesn't have to issue any requests to the OS.
We add ReadaheadRandomAccessFile layer only when file is read during compactions.
D45105 was incorrectly closed by Phabricator because I committed it to a separate branch (not master), so I'm resubmitting the diff.
Test Plan: make check
Reviewers: MarkCallaghan, sdong
Reviewed By: sdong
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D45123
Summary: In internal stats, remember read latency histogram, if statistics is enabled. It can be retrieved from DB::GetProperty() with "rocksdb.dbstats" property, if it is enabled.
Test Plan: Manually run db_bench and prints out "rocksdb.dbstats" by hand and make sure it prints out as expected
Reviewers: igor, IslamAbdelRahman, rven, kradhakrishnan, anthony, yhchiang
Reviewed By: yhchiang
Subscribers: MarkCallaghan, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D44193
Summary:
Add options.compaction_measure_io_stats to print out / pass to listener accumulated time spent on write calls. Example outputs in info logs:
2015/08/12-16:27:59.463944 7fd428bff700 (Original Log Time 2015/08/12-16:27:59.463922) EVENT_LOG_v1 {"time_micros": 1439422079463897, "job": 6, "event": "compaction_finished", "output_level": 1, "num_output_files": 4, "total_output_size": 6900525, "num_input_records": 111483, "num_output_records": 106877, "file_write_nanos": 15663206, "file_range_sync_nanos": 649588, "file_fsync_nanos": 349614797, "file_prepare_write_nanos": 1505812, "lsm_state": [2, 4, 0, 0, 0, 0, 0]}
Add two more counters in iostats_context.
Also add a parameter of db_bench.
Test Plan: Add a unit test. Also manually verify LOG outputs in db_bench
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D44115
Summary: RandomRWFile is not used anywhere in out code base, this patch remove RandomRWFile
Test Plan:
make check -j64
USE_CLANG=1 make all -j64
OPT=-DROCKSDB_LITE make release -j64
Reviewers: sdong, yhchiang, anthony, kradhakrishnan, rven, igor
Reviewed By: igor
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D44091
Summary:
Add two unit tests for SyncWAL(). One makes sure SyncWAL() doesn't block writes in the other thread. Another one makes sure SyncWAL() doesn't wait ongoing writes to finish before being executed.
Create a new test file db_wal_test and move two WAL related tests from db_test to here.
Test Plan: Run the new tests
Reviewers: IslamAbdelRahman, rven, kradhakrishnan, kolmike, tnovak, yhchiang
Reviewed By: yhchiang
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D43605
Summary: Measure read latency histogram and put in statistics. Compaction inputs are excluded from it when possible (unfortunately usually no possible as we usually take table reader from table cache.
Test Plan:
Run db_bench and it shows the stats, like:
rocksdb.sst.read.micros statistics Percentiles :=> 50 : 1.238522 95 : 2.529740 99 : 3.912180
Reviewers: kradhakrishnan, rven, anthony, IslamAbdelRahman, MarkCallaghan, yhchiang
Reviewed By: yhchiang
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D43275
Summary:
Subj. We really need this feature.
Previous diff D40899 has most of the changes to make this possible, this diff just adds the method.
Test Plan: `make check`, the new test fails without this diff; ran with ASAN, TSAN and valgrind.
Reviewers: igor, rven, IslamAbdelRahman, anthony, kradhakrishnan, tnovak, yhchiang, sdong
Reviewed By: sdong
Subscribers: MarkCallaghan, maykov, hermanlee4, yoshinorim, tnovak, dhruba
Differential Revision: https://reviews.facebook.net/D40905
Summary:
From other ones' investigation:
"sync_file_range() behavior highly depends on kernel version and filesystem.
xfs does neighbor page flushing outside of the specified ranges. For example, sync_file_range(fd, 8192, 16384) does not only trigger flushing page #3 to #4, but also flushing many more dirty pages (i.e. up to page#16)... Ranges of the sync_file_range() should be far enough from write() offset (at least 1MB)."
Test Plan: make all check
Reviewers: igor, rven, kradhakrishnan, yhchiang, IslamAbdelRahman, anthony
Reviewed By: anthony
Subscribers: yoshinorim, MarkCallaghan, sumeet, domas, dhruba, leveldb, ljin
Differential Revision: https://reviews.facebook.net/D15807
Sagar Vemuri