Summary:
New memory technologies are being developed by various hardware vendors (Intel DCPMM is one such technology currently available). These new memory types require different libraries for allocation and management (such as PMDK and memkind). The high capacities available make it possible to provision large caches (up to several TBs in size), beyond what is achievable with DRAM.
The new allocator provided in this PR uses the memkind library to allocate memory on different media.
**Performance**
We tested the new allocator using db_bench.
- For each test, we vary the size of the block cache (relative to the size of the uncompressed data in the database).
- The database is filled sequentially. Throughput is then measured with a readrandom benchmark.
- We use a uniform distribution as a worst-case scenario.
The plot shows throughput (ops/s) relative to a configuration with no block cache and default allocator.
For all tests, p99 latency is below 500 us.
**Changes**
- Add MemkindKmemAllocator
- Add --use_cache_memkind_kmem_allocator db_bench option (to create an LRU block cache with the new allocator)
- Add detection of memkind library with KMEM DAX support
- Add test for MemkindKmemAllocator
**Minimum Requirements**
- kernel 5.3.12
- ndctl v67 - https://github.com/pmem/ndctl
- memkind v1.10.0 - https://github.com/memkind/memkind
**Memory Configuration**
The allocator uses the MEMKIND_DAX_KMEM memory kind. Follow the instructions on[ memkind’s GitHub page](https://github.com/memkind/memkind) to set up NVDIMM memory accordingly.
Note on memory allocation with NVDIMM memory exposed as system memory.
- The MemkindKmemAllocator will only allocate from NVDIMM memory (using memkind_malloc with MEMKIND_DAX_KMEM kind).
- The default allocator is not restricted to RAM by default. Based on NUMA node latency, the kernel should allocate from local RAM preferentially, but it’s a kernel decision. numactl --preferred/--membind can be used to allocate preferentially/exclusively from the local RAM node.
**Usage**
When creating an LRU cache, pass a MemkindKmemAllocator object as argument.
For example (replace capacity with the desired value in bytes):
```
#include "rocksdb/cache.h"
#include "memory/memkind_kmem_allocator.h"
NewLRUCache(
capacity /*size_t*/,
6 /*cache_numshardbits*/,
false /*strict_capacity_limit*/,
false /*cache_high_pri_pool_ratio*/,
std::make_shared<MemkindKmemAllocator>());
```
Refer to [RocksDB’s block cache documentation](https://github.com/facebook/rocksdb/wiki/Block-Cache) to assign the LRU cache as block cache for a database.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6214
Reviewed By: cheng-chang
Differential Revision: D19292435
fbshipit-source-id: 7202f47b769e7722b539c86c2ffd669f64d7b4e1
Summary:
In the `.travis.yml` file the `jdk: openjdk7` element is ignored when `language: cpp`. So whatever version of the JDK that was installed in the Travis container was used - typically JDK 11.
To ensure our RocksJava builds are working, we now instead install and use OpenJDK 8. Ideally we would use OpenJDK 7, as RocksJava supports Java 7, but many of the newer Travis containers don't support Java 7, so Java 8 is the next best thing.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6512
Differential Revision: D20388296
Pulled By: pdillinger
fbshipit-source-id: 8bbe6b59b70cfab7fe81ff63867d907fefdd2df1
Summary:
Check for sys/auxv.h and getauxval before using them as they are not
always available (for example on uclibc)
Signed-off-by: Fabrice Fontaine <fontaine.fabrice@gmail.com>
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6359
Differential Revision: D20239797
fbshipit-source-id: 175a098094d81545628c2372e7c388e70a32fd48
Summary:
We realized bugs related to IO Uring. Turn it off by default.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6405
Test Plan: Manually run build_tools/build_detect_platform and observe outputs.
Differential Revision: D19862792
fbshipit-source-id: 5d5e8e2762997b72a145ae59389ef3d7e4ccd060
Summary:
I set up a mirror of our Java deps on github so we can download
them through github URLs rather than maven.org, which is proving
terribly unreliable from Travis builds.
Also sanitized calls to curl, so they are easier to read and
appropriately fail on download failure.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6348
Test Plan: CI
Differential Revision: D19633621
Pulled By: pdillinger
fbshipit-source-id: 7eb3f730953db2ead758dc94039c040f406790f3
Summary:
Difficult to root cause crash test failures without archiving
db dir. Now all crash test configurations should save the db dir.
Also exit with error code on bad command.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6344
Test Plan:
Hmm, how about this:
for TARGET in stress_crash asan_crash ubsan_crash tsan_crash; do EMAIL=email ONCALL=oncall TRIGGER=all SUBSCRIBER=sub build_tools/rocksdb-lego-determinator $TARGET > tmp && node -c tmp && grep -q Upload tmp || echo Bad; done
Differential Revision: D19625605
Pulled By: pdillinger
fbshipit-source-id: cb84aa93ee80b4534f4c61b90f0e0f99a41155d5
Summary:
While the instruction of installing "make format" dependencies works on some platforms, it is hard to use for some others. Improve it a little bit.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6162
Test Plan: Run "make format" on an envrionment missing the dependencies and see the instructions printed out
Differential Revision: D18970773
fbshipit-source-id: fd21b31053407cc171a6675f781a556a1c3e8945
Summary:
Right now, PosixRandomAccessFile::MultiRead() executes read requests in parallel. In this PR, it leverages I/O Uring library to run it in parallel, even when page cache is enabled. This function will fall back if the kernel version doesn't support it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5881
Test Plan: Run the unit test on a kernel version supporting it and make sure all tests pass, and run a unit test on kernel version supporting it and see it pass. Before merging, will also run stress test and see it passes.
Differential Revision: D17742266
fbshipit-source-id: e05699c925ac04fdb42379456a4e23e4ebcb803a
Summary:
Adds an improved, replacement Bloom filter implementation (FastLocalBloom) for full and partitioned filters in the block-based table. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single filter.
Speed
The improved speed, at least on recent x86_64, comes from
* Using fastrange instead of modulo (%)
* Using our new hash function (XXH3 preview, added in a previous commit), which is much faster for large keys and only *slightly* slower on keys around 12 bytes if hashing the same size many thousands of times in a row.
* Optimizing the Bloom filter queries with AVX2 SIMD operations. (Added AVX2 to the USE_SSE=1 build.) Careful design was required to support (a) SIMD-optimized queries, (b) compatible non-SIMD code that's simple and efficient, (c) flexible choice of number of probes, and (d) essentially maximized accuracy for a cache-local Bloom filter. Probes are made eight at a time, so any number of probes up to 8 is the same speed, then up to 16, etc.
* Prefetching cache lines when building the filter. Although this optimization could be applied to the old structure as well, it seems to balance out the small added cost of accumulating 64 bit hashes for adding to the filter rather than 32 bit hashes.
Here's nominal speed data from filter_bench (200MB in filters, about 10k keys each, 10 bits filter data / key, 6 probes, avg key size 24 bytes, includes hashing time) on Skylake DE (relatively low clock speed):
$ ./filter_bench -quick -impl=2 -net_includes_hashing # New Bloom filter
Build avg ns/key: 47.7135
Mixed inside/outside queries...
Single filter net ns/op: 26.2825
Random filter net ns/op: 150.459
Average FP rate %: 0.954651
$ ./filter_bench -quick -impl=0 -net_includes_hashing # Old Bloom filter
Build avg ns/key: 47.2245
Mixed inside/outside queries...
Single filter net ns/op: 63.2978
Random filter net ns/op: 188.038
Average FP rate %: 1.13823
Similar build time but dramatically faster query times on hot data (63 ns to 26 ns), and somewhat faster on stale data (188 ns to 150 ns). Performance differences on batched and skewed query loads are between these extremes as expected.
The only other interesting thing about speed is "inside" (query key was added to filter) vs. "outside" (query key was not added to filter) query times. The non-SIMD implementations are substantially slower when most queries are "outside" vs. "inside". This goes against what one might expect or would have observed years ago, as "outside" queries only need about two probes on average, due to short-circuiting, while "inside" always have num_probes (say 6). The problem is probably the nastily unpredictable branch. The SIMD implementation has few branches (very predictable) and has pretty consistent running time regardless of query outcome.
Accuracy
The generally improved accuracy (re: Issue https://github.com/facebook/rocksdb/issues/5857) comes from a better design for probing indices
within a cache line (re: Issue https://github.com/facebook/rocksdb/issues/4120) and improved accuracy for millions of keys in a single filter from using a 64-bit hash function (XXH3p). Design details in code comments.
Accuracy data (generalizes, except old impl gets worse with millions of keys):
Memory bits per key: FP rate percent old impl -> FP rate percent new impl
6: 5.70953 -> 5.69888
8: 2.45766 -> 2.29709
10: 1.13977 -> 0.959254
12: 0.662498 -> 0.411593
16: 0.353023 -> 0.0873754
24: 0.261552 -> 0.0060971
50: 0.225453 -> ~0.00003 (less than 1 in a million queries are FP)
Fixes https://github.com/facebook/rocksdb/issues/5857
Fixes https://github.com/facebook/rocksdb/issues/4120
Unlike the old implementation, this implementation has a fixed cache line size (64 bytes). At 10 bits per key, the accuracy of this new implementation is very close to the old implementation with 128-byte cache line size. If there's sufficient demand, this implementation could be generalized.
Compatibility
Although old releases would see the new structure as corrupt filter data and read the table as if there's no filter, we've decided only to enable the new Bloom filter with new format_version=5. This provides a smooth path for automatic adoption over time, with an option for early opt-in.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6007
Test Plan: filter_bench has been used thoroughly to validate speed, accuracy, and correctness. Unit tests have been carefully updated to exercise new and old implementations, as well as the logic to select an implementation based on context (format_version).
Differential Revision: D18294749
Pulled By: pdillinger
fbshipit-source-id: d44c9db3696e4d0a17caaec47075b7755c262c5f
Summary:
- Updated our included xxhash implementation to version 0.7.2 (== the latest dev version as of 2019-10-09).
- Using XXH_NAMESPACE (like other fb projects) to avoid potential name collisions.
- Added fastrange64, and unit tests for it and fastrange32. These are faster alternatives to hash % range.
- Use preview version of XXH3 instead of MurmurHash64A for NPHash64
-- Had to update cache_test to increase probability of passing for any given hash function.
- Use fastrange64 instead of % with uses of NPHash64
-- Had to fix WritePreparedTransactionTest.CommitOfDelayedPrepared to avoid deadlock apparently caused by new hash collision.
- Set default seed for NPHash64 because specifying a seed rarely makes sense for it.
- Removed unnecessary include xxhash.h in a popular .h file
- Rename preview version of XXH3 to XXH3p for clarity and to ease backward compatibility in case final version of XXH3 is integrated.
Relying on existing unit tests for NPHash64-related changes. Each new implementation of fastrange64 passed unit tests when manipulating my local build to select it. I haven't done any integration performance tests, but I consider the improved performance of the pieces being swapped in to be well established.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5909
Differential Revision: D18125196
Pulled By: pdillinger
fbshipit-source-id: f6bf83d49d20cbb2549926adf454fd035f0ecc0d
Summary:
Some dependency path is not correct so that ASAN cannot run with CLANG. Fix it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5946
Test Plan: Run ASAN with CLANG
Differential Revision: D18040933
fbshipit-source-id: 1d82be9d350485cf1df1c792dad765188958641f
Summary:
format-diff.sh, a.k.a. 'make format', would use 'master'
to decide which commits are probably unpublished. Much better to use
facebook remote master since local master may not be caught up and may
have its own unpublished commits. Script now tries to compare against
facebook remote master branch (branch pointer is updated with any fetch
or pull), because those differences are what would be considered the
differences for a pull request.
Also, script would compare against *parent* of merge-base with that
reference point, which is just wrong since that includes the last
published commit.
In case of problems, you can now customize the reference point, by
setting the FORMAT_UPSTREAM variable.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5831
Test Plan: manual
Differential Revision: D17528462
Pulled By: pdillinger
fbshipit-source-id: 50fdb8795d683bf3c14d449669c1a5299e0dfa8b
Summary:
Update version of dependencies.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5777
Test Plan: make release
Differential Revision: D17269421
fbshipit-source-id: e76dbe5389e1d7f811739d3bc1e404b482dfce34
Summary:
In preparing to utilize a new Intel instruction extension, I
noticed problems with the existing build script in regard to the
existing utilized extensions, either with USE_SSE or PORTABLE flags.
* PORTABLE=0 was interpreted the same as PORTABLE=1. Now empty and 0
mean the same. (I guess you were not supposed to set PORTABLE= if you
wanted non-portable--except that...)
* The Facebook build script extensions would set PORTABLE=1 even if
it's already set in a make var or environment. Now it does not override
a non-empty setting, so use PORTABLE=0 for fully optimized build,
overriding Facebook environment default.
* Put in an explanation of the USE_SSE flag where it's used by
build_detect_platform, and cleaned up some confusing/redundant
associated logic.
* If USE_SSE was set and expected intrinsics were not available,
build_detect_platform would exit early but build would proceed with
broken, incomplete configuration. Now warning is gracefully recovered.
* If USE_SSE was set and expected intrinsics were not available,
build would still try to use flags like -msse4.2 etc. which could lead
to unexpected compilation failure or binary incompatibility. Now those
flags are not used if the warning is issued.
This should not break or change existing, valid build scripts.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5800
Test Plan: manual case testing
Differential Revision: D17369543
Pulled By: pdillinger
fbshipit-source-id: 4ee244911680ae71144d272c40aceea548e3ce88
Summary:
This ports `folly::DistributedMutex` into RocksDB. The PR includes everything else needed to compile and use DistributedMutex as a component within folly. Most files are unchanged except for some portability stuff and includes.
For now, I've put this under `rocksdb/third-party`, but if there is a better folder to put this under, let me know. I also am not sure how or where to put unit tests for third-party stuff like this. It seems like gtest is included already, but I need to link with it from another third-party folder.
This also includes some other common components from folly
- folly/Optional
- folly/ScopeGuard (In particular `SCOPE_EXIT`)
- folly/synchronization/ParkingLot (A portable futex-like interface)
- folly/synchronization/AtomicNotification (The standard C++ interface for futexes)
- folly/Indestructible (For singletons that don't get destroyed without allocations)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5642
Differential Revision: D16544439
fbshipit-source-id: 179b98b5dcddc3075926d31a30f92fd064245731
Summary:
Add an extra cleanup step so that db directory can be saved and uploaded.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5554
Reviewed By: yancouto
Differential Revision: D16168844
Pulled By: riversand963
fbshipit-source-id: ec7b2cee5f11c7d388c36531f8b076d648e2fb19
Summary:
This property is needed to run the child jobs on the same host and thus propagate the child job status back to the parent's.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5456
Reviewed By: yancouto
Differential Revision: D15824382
Pulled By: maysamyabandeh
fbshipit-source-id: 42f2efbedaa3a8b399281105f0ce793c1c9a6191
Summary:
Special characters like slashes and parentheses are not supported.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5424
Differential Revision: D15708067
Pulled By: ltamasi
fbshipit-source-id: 90527ec3ee882a0cdd1249c3946f5eff2ff7c115
Summary:
This change adds a Dynamic Library class to the RocksDB Env. Dynamic libraries are populated via the Env::LoadLibrary method.
The addition of dynamic library support allows for a few different features to be developed:
1. The compression code can be changed to use dynamic library support. This would allow RocksDB to determine at run-time what compression packages were installed. This change would eliminate the need to make sure the build-time and run-time environment had the same library set. It would also simplify some of the Java build issues (where it attempts to build and include various packages inside the RocksDB jars).
2. Along with other features (to be provided in a subsequent PR), this change would allow code/configurations to be added to RocksDB at run-time. For example, the build system includes code for building an "rados" environment and adding "Cassandra" features. Instead of these extensions being built into the base RocksDB code, these extensions could be loaded at run-time as required/appropriate, either by configuration or explicitly.
We intend to push out other changes in support of the extending RocksDB at run-time via configurations.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5281
Differential Revision: D15447613
Pulled By: riversand963
fbshipit-source-id: 452cd4f54511c0bceee18f6d9d919aae9fd25fef
Summary:
Fix some hdfs-related code so that it can compile and run 'db_stress'
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5122
Differential Revision: D14675495
Pulled By: riversand963
fbshipit-source-id: cac280479efcf5451982558947eac1732e8bc45a
Summary:
The compiler flag `-DROCKSDB_FALLOCATE_PRESENT` was only set when
`fallocate`, `FALLOC_FL_KEEP_SIZE`, and `FALLOC_FL_PUNCH_HOLE` were all
present. However, the last of the three is not really necessary for the
primary `fallocate` use case; furthermore, it was introduced only in later
Linux kernel versions (2.6.38+).
This PR changes the flag `-DROCKSDB_FALLOCATE_PRESENT` to only require
`fallocate` and `FALLOC_FL_KEEP_SIZE` to be present. There is a separate
check for `FALLOC_FL_PUNCH_HOLE` only in the place where it is used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5023
Differential Revision: D14248487
Pulled By: siying
fbshipit-source-id: a10ed0b902fa755988e957bd2dcec9081ec0502e
Summary:
Currently crash test covers cases with and without atomic flush, but takes too
long to finish. Therefore it may be a better idea to put crash test with atomic
flush in a separate set of tests.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4945
Differential Revision: D13947548
Pulled By: riversand963
fbshipit-source-id: 177c6de865290fd650b0103408339eaa3f801d8c
Summary:
We should strip `-DZSTD` to prevent ZSTD from being used in the no compression tests, similarly to how we prevent all other compression libraries from being used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4982
Differential Revision: D14075349
Pulled By: ajkr
fbshipit-source-id: 8bd861516cf28a568c2b701ad33d0bb658db93b2
Summary:
- When building with internal dependencies, specify this toolchain by setting `ROCKSDB_FBCODE_BUILD_WITH_PLATFORM007=1`
- It is not enabled by default. However, it is enabled for TSAN builds in CI since there is a known problem with TSAN in gcc-5: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71090
- I did not add support for Lua since (1) we agreed to deprecate it, and (2) we only have an internal build for v5.3 with this toolchain while that has breaking changes compared to our current version (v5.2).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4923
Differential Revision: D13827226
Pulled By: ajkr
fbshipit-source-id: 9aa3388ed3679777cfb15ef8cbcb83c07f62f947
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:
Found a callsite for `moo_translate` in the Scuba warnings and realized we have a few calls to `define()` left in fbcode.
- I ran the `DefineCodemod` script against fbcode
- Fixed broken tests, and ensured that tests that are explicitly testing the behaviour of `define()` were not changed.
bypass-lint
Reviewed By: kmeht
Differential Revision: D12968447
fbshipit-source-id: d8fd3649a2ce9868b8938d293e1bebf1a6d2fad8
Summary:
In fbcode when we build with clang7++, although -faligned-new is available in compile phase, we link with an older version of libstdc++.a and it doesn't come with aligned-new support (e.g. `nm libstdc++.a | grep align_val_t` return empty). In this case the previous -faligned-new detection can pass but will end up with link error. Fixing it by only have the detection for non-fbcode build.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4576
Differential Revision: D10500008
Pulled By: yiwu-arbug
fbshipit-source-id: b375de4fbb61d2a08e54ab709441aa8e7b4b08cf
Summary:
I noticed we were building against zstd 1.3.0 which is missing optimizations that our fbcode customers have (they're on zstd 1.3.5). Ran `./build_tools/update_dependencies.sh` to catch us up. Omitted the changes it made for gcc-4.8 since it's broken.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4244
Differential Revision: D9230111
Pulled By: ajkr
fbshipit-source-id: 3e8ec1d8a961f98ec77c8c6580bde4caacf2d437
Summary:
The error_filter.py script parses the output of the "Build and run" stage of continuous tests to check for errors. It is currently only detecting compile errors and not link errors. This change fixes that.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4247
Differential Revision: D9233735
Pulled By: anand1976
fbshipit-source-id: 16e5a04950891cd9aba5cb3efcb6abc2a2e0d5ae
Summary:
TSAN requires the code is built with -fPIC. This PR links against a libzstd built with -fPIC when necessary, which enables ZSTD compression to be used in TSAN builds.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4249
Differential Revision: D9244746
Pulled By: ajkr
fbshipit-source-id: 8c6a8fadd6c8643b2077afcbc3626779e1d73b63
Summary:
Fix the nested quotes for CRASH_TEST_EXT_ARGS, as the generated json could not be parsed by the sandcastle job.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4243
Differential Revision: D9228429
Pulled By: anand1976
fbshipit-source-id: 3c2bcac34870e377949d8a79c55e33b8363b25dd
Summary:
- Add `--compression_max_dict_bytes` and `--compression_zstd_max_train_bytes` flags to stress test
- Randomly enable/disable the above flags in crash test
- Set `--compression_type=zstd` in FB-specific crash test runs
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4234
Differential Revision: D9187207
Pulled By: ajkr
fbshipit-source-id: 8d78cf8d8e1165f2cd1c32e069b73726b5bc1fd2
Summary:
clang compilation is failing due to a4fb1f8c04. In that commit I added a call to `std::atomic::is_lock_free` which was evidently relying on a compiler builtin only present in gcc.
Drawbacks to this fix are:
- users may need to install libatomic
- there might be cases where clang is used even though USE_CLANG is unset (e.g., when clang is the only available compiler). I didn't figure out how to add -latomic in those cases...
An alternative fix mentioned in http://lists.llvm.org/pipermail/llvm-bugs/2017-August/057263.html is using -stdlib=libc++ with clang.
Closes https://github.com/facebook/rocksdb/pull/3769
Differential Revision: D7756261
Pulled By: ajkr
fbshipit-source-id: 26888300683fa9970ab5950239d1aa217e8efd49
Luca Giacchino
