Motivation:
Thread-confinement ends up being too confusing to code for, and also prevents some legitimate use cases.
Additionally, thread-confinement exposed implementation specific behavioural differences of buffers, where we would ideally like all buffers to always behave the same, regardless of implementation.
Modification:
All MemorySegment based buffers now always use shared segments.
For heap-based segments, we avoid the overhead associated with the closing of shared segments, by just not closing them, and instead just leave the whole thing for the GC to deal with.
Result:
Buffers can now always be accessed from multiple different threads at the same time.
Motivation:
Although having a cleaner attached adds a bit of overhead when allocating or closing buffers,
it is more important to make our systems, libraries and frameworks misuse resistant and safe by default.
Modification:
Remove the ability to allocate a buffer that does not have a cleaner attached.
Reference counting and the ability to explicitly release memory remains.
This just makes sure that we always have a safety net to fall back on.
Result:
This will make systems less prone to crashes through running out of memory, native or otherwise, even in the face of true memory leaks.
(Leaks through retained strong references cannot be fixed in any way)
Looks like the overhead is not too bad, so I think we can just always do that:
```
Benchmark (workload) Mode Cnt Score Error Units
explicitPooledClose light avgt 150 1,094 ± 0,017 us/op
pooledWithCleanerExplicitClose light avgt 150 1,181 ± 0,009 us/op
```
Motivation:
The main use case with Buf.compact is in conjunction with ensureWritable.
It turns out we can get a simpler API, and faster methods, by combining those two operations, because it allows us to relax some guarantees and skip some steps in certain cases, which wouldn't be as neat or clean if they were two separate steps.
Modification:
Add a new Buf.ensureWritable method, which takes an allowCompaction argument.
In MemSegBuf, we can just delegate to compact() when applicable.
In CompositeBuf, we can sometimes get away with just reorganising the bufs array.
Result:
We can now do ensureWritable without allocating in some cases, and this can in particular make the operation faster for CompositeBuf.
Motivation:
Compaction makes more space available at the end of a buffer, by discarding bytes at the beginning that have already been processed.
Modification:
Add a copying compact method to Buf.
Result:
It is now possible to discard read bytes by calling `compact()`.
Motivation:
There are many use cases where other objects will have fields that are buffers.
Since buffers are reference counted, their life cycle needs to be managed carefully.
Modification:
Add the abstract BufHolder, and the concrete sub-class BufRef, as neat building blocks for building other classes that contain field references to buffers.
The behaviours of closed/sent buffers have also been specified in tests, and tightened up in the code.
Result:
It is now easier to create classes/objects that wrap buffers.
Motivation:
There are use cases that involve accumulating data into a buffer, then carving out prefix slices and sending them off on their own journey for further processing.
Modification:
Add a Buf.bifurcate API, that split a buffer, and its ownership, in two.
Internally, the API will inject and maintain an atomically reference counted Drop instance, so that the original memory segment is not released until all bifurcated parts are closed.
This works particularly well for composite buffers, where only the buffer (if any) wherein the bifurcation point lands, will actually have its memory split. A composite buffer can otherwise just crack its buffer array in two.
Result:
We now have a safe way of breaking the single ownership of some memory into multiple parts, that can be sent and owned independently.
Motivation:
Cursors are better than iterators in that they only need to check boundary conditions once per iteration, when processed in a loop.
This should make them easier for the compiler to optimise.
Modification:
Change the ByteIterator to a ByteCursor. The API is almost the same, but with a few subtle differences in semantics.
The primary difference is that the cursor movement and boundary condition checking and position movement happen at the same time, and do not need to occur when the values are fetched out of the cursor.
An iterator, on the other hand, needs to throw an exception if "next" is called too many times.
Result:
Simpler code, and hopefully faster code as well.
Motivation:
We were seeing rare test failures where a cleaner had raced to close a memory segment we were using or closing.
The cause was that a single MemorySegment ended up used in multiple Buf instances.
When the SizeClassedMemoryPool was closed, the memory segments could be disposed without closing the gate in the NativeMemoryCleanerDrop.
The gate is important because it prevents double-frees of the memory segment.
Modification:
The fix is to change how the SizeClassedMemoryPool is closed, such that it always releases memory by calling `close()` on its buffers, which in turn will close the gate. The program will then proceed through the SizeClassedMemoryPool.drop implementation, which in turn will observe that the allocator is closed, and *then* dispose of the memory.
Result:
We should hopefully not see any more random test failures, but if we do, they would at least indicate a different bug.
This particular one was mostly showing up inside the cleaner threads, which were ignoring the exception, but occasionally I think the race went the other way, causing a test failure.
Motivation:
This will likely be a somewhat common operation, as buffers move between eventloop and worker threads, so it's important to have an understanding of how it performs.
Modification:
Add a benchmark that specifically targets the send() operation on buffers.
Result:
We got benchmark numbers that clearly show the cost of confinement transfer
Motivation:
Composite buffers are uniquely positioned to be able to extend their underlying storage relatively cheaply.
This fact is relied upon in a couple of buffer use cases within Netty, that we wish to support.
Modification:
Add a static `extend` method to Allocator, so that the CompositeBuf class can remain internal.
The `extend` method inserts the extension buffer at the end of the composite buffer as if it had been included from the start.
This involves checking offsets and byte order invariants.
We also require that the composite buffer be in an owned state.
Result:
It's now possible to extend a composite buffer with a specific buffer, after the composite buffer has been created.
Motivation:
Capture the performance characteristics of this primitive for various buffer implementations.
Modification:
Add a benchmark that iterate 4KiB buffers forwards, and backwards, on various buffer implementations.
Result:
Another aspect of the implementation covered by benchmarks.
Turns out the composite iterators a somewhat slow.
Motivation:
When a build fails, it's desirable to have the build artifacts available
so the build failure can be inspected, investigated and hopefully fixed.
Modification:
Change the Makefile and CI workflow such that the build artifacts are
captured and uploaded when the build fails.
Result:
A "target.zip" file is now available for download on failed GitHub
Actions builds.
Motivation:
Pooled buffers are a very important use case, and they change the cost dynamics around shared memory segments, so it's worth looking into in detail.
Modification:
Add another explicit close of pooled direct buffers to MemorySegmentClosedByCleanerBenchmark
Result:
Explicitly closing of pooled buffers is even out-performing cleaner close on the "heavy" workload, so this is currently the fastest way to run that workload:
Benchmark (workload) Mode Cnt Score Error Units
MemorySegmentClosedByCleanerBenchmark.cleanerClose heavy avgt 150 14,194 ± 0,558 us/op
MemorySegmentClosedByCleanerBenchmark.explicitClose heavy avgt 150 40,496 ± 0,414 us/op
MemorySegmentClosedByCleanerBenchmark.explicitPooledClose heavy avgt 150 12,723 ± 0,134 us/op
When send() a confined buffer, we had to first turn it into a shard buffer, so that it could be claimed by an arbitrary recipient thread.
As we've learned, however, closing shared segments is expensive.
We can speed up the send() call by simply leaving the segment shared.
This weakens the confinement of the received segment, though.
Currently no tests fails on that, but in the future we should re-implement confinement checking inside the Buf implementations themselves any, because pooled buffers also violate the confinement restriction, and we have a guiding principle that all buffers, regardless of implementation, should always behave the same.
The results of this change can be observed in the MemorySegmentClosedByCleanerBenchmark, with the heavy workload.
Explicitly closed segments now run the workload twice as fast, and the cleaner based closing is now 3 times faster.
Before:
Benchmark (workload) Mode Cnt Score Error Units
MemorySegmentClosedByCleanerBenchmark.cleanerClose heavy avgt 150 42,221 ± 0,943 us/op
MemorySegmentClosedByCleanerBenchmark.explicitClose heavy avgt 150 65,215 ± 0,761 us/op
After:
Benchmark (workload) Mode Cnt Score Error Units
MemorySegmentClosedByCleanerBenchmark.cleanerClose heavy avgt 150 13,871 ± 0,544 us/op
MemorySegmentClosedByCleanerBenchmark.explicitClose heavy avgt 150 37,516 ± 0,426 us/op
Also schedule a build to run at 06:30 in the morning, every Monday.
This way, the JDK and Netty 5 snapshots will be updated in the build cache every week.
