This is just a draft PR for a first insight on memory mapping improvements in JDK 16+.
Some background information: Starting with JDK-14, there is a new incubating module "jdk.incubator.foreign" that has a new, not yet stable API for accessing off-heap memory (and later it will also support calling functions using classical MethodHandles that are located in libraries like .so or .dll files). This incubator module has several versions:
- first version: https://openjdk.java.net/jeps/370 (slow, very buggy and thread confinement, so making it unuseable with Lucene)
- second version: https://openjdk.java.net/jeps/383 (still thread confinement, but now allows transfer of "ownership" to other threads; this is still impossible to use with Lucene.
- third version in JDK 16: https://openjdk.java.net/jeps/393 (this version has included "Support for shared segments"). This now allows us to safely use the same external mmaped memory from different threads and also unmap it!
This module more or less overcomes several problems:
- ByteBuffer API is limited to 32bit (in fact MMapDirectory has to chunk in 1 GiB portions)
- There is no official way to unmap ByteBuffers when the file is no longer used. There is a way to use
sun.misc.Unsafe
and forcefully unmap segments, but any IndexInput accessing the file from another thread will crush the JVM with SIGSEGV or SIGBUS. We learned to live with that and we happily apply the unsafe unmapping, but that's the main issue.
@uschindler had many discussions with the team at OpenJDK and finally with the third incubator, we have an API that works with Lucene. It was very fruitful discussions (thanks to @mcimadamore !)
With the third incubator we are now finally able to do some tests (especially performance). As this is an incubating module, this PR first changes a bit the build system:
- disable
-Werror
for :lucene:core
- add the incubating module to compiler of
:lucene:core
and enable it for all test builds. This is important, as you have to pass --add-modules jdk.incubator.foreign
also at runtime!
The code basically just modifies MMapDirectory
to use LONG instead of INT for the chunk size parameter. In addition it adds MemorySegmentIndexInput
that is a copy of our ByteBufferIndexInput
(still there, but unused), but using MemorySegment instead of ByteBuffer behind the scenes. It works in exactly the same way, just the try/catch blocks for supporting EOFException or moving to another segment were rewritten.
The openInput code uses MemorySegment.mapFile()
to get a memory mapping. This method is unfortunately a bit buggy in JDK-16-ea-b30, so I added some workarounds. See JDK issues: https://bugs.openjdk.java.net/browse/JDK-8259027, https://bugs.openjdk.java.net/browse/JDK-8259028, https://bugs.openjdk.java.net/browse/JDK-8259032, https://bugs.openjdk.java.net/browse/JDK-8259034. The bugs with alignment and zero byte mmaps are fixed in b32, this PR was adapted (hacks removed).
It passes all tests and it looks like you can use it to read indexes. The default chunk size is now 16 GiB (but you can raise or lower it as you like; tests are doing this). Of course you can set it to Long.MAX_VALUE, in that case every index file is always mapped to one big memory mapping. My testing with Windows 10 have shown, that this is not a good idea!!!. Huge mappings fragment address space over time and as we can only use like 43 or 46 bits (depending on OS), the fragmentation will at some point kill you. So 16 GiB looks like a good compromise: Most files will be smaller than 6 GiB anyways (unless you optimize your index to one huge segment). So for most Lucene installations, the number of segments will equal the number of open files, so Elasticsearch huge user consumers will be very happy. The sysctl max_map_count may not need to be touched anymore.
In addition, this implements readLELongs
in a better way than @jpountz did (no caching or arbitrary objects). Nevertheless, as the new MemorySegment API relies on final, unmodifiable classes and coping memory from a MemorySegment to a on-heap Java array, it requires us to wrap all those arrays using a MemorySegment each time (e.g. in readBytes()
or readLELongs
), there may be some overhead du to short living object allocations (those are NOT reuseable!!!). In short: In future we should throw away on coping/loading our stuff to heap and maybe throw away IndexInput completely and base our code fully on random access. The new foreign-vector APIs will in future also be written with MemorySegment in its focus. So you can allocate a vector view on a MemorySegment and let the vectorizer fully work outside java heap inside our mmapped files! :-)
It would be good if you could checkout this branch and try it in production.
But be aware:
- You need JDK 11 to run Gradle (set
JAVA_HOME
to it)
- You need JDK 16-ea-b32 (set
RUNTIME_JAVA_HOME
to it)
- The lucene-core.jar will be JDK16 class files and requires JDK-16 to execute.
- Also you need to add
--add-modules jdk.incubator.foreign
to the command line of your Java program/Solr server/Elasticsearch server
It would be good to get some benchmarks, especially by @rmuir or @mikemccand. Take your time and enjoy the complexity of setting this up! ;-)
My plan is the following:
- report any bugs or slowness, especially with Hotspot optimizations. The last time I talked to Maurizio, he taked about Hotspot not being able to fully optimize for-loops with long instead of int, so it may take some time until the full performance is there.
- wait until the final version of project PANAMA-foreign goes into Java's Core Library (no module needed anymore)
- add a MR-JAR for lucene-core.jar and compile the MemorySegmentIndexInput and maybe some helper classes with JDK 17/18/19 (hopefully?).
~~In addition there are some comments in the code talking about safety (e.g., we need IOUtils.close()
taking AutoCloseable
instead of just Closeable
, so we can also enfoce that all memory segments are closed after usage.~~ In addition, by default all VarHandles are aligned. By default it refuses to read a LONG from an address which is not a multiple of 8. I had to disable this feature, as all our index files are heavily unaliged. We should in meantime not only convert our files to little endian, but also make all non-compressed types (like long[]
arrays or non-encoded integers be aligned to the correct boundaries in files). The most horrible thing I have seen is that our CFS file format starts the "inner" files totally unaligned. We should fix the CFSWriter to start new files always at multiples of 8 bytes. I will open an issue about this.
enhancement optimization