Related issue: #2028
Motivation:
Some copiedBuffer() methods in Unpooled allocated a direct buffer. An
allocation of a direct buffer is an expensive operation, and thus should
be avoided for unpooled buffers.
Modifications:
- Use heap buffers in all copiedBuffer() methods
Result:
Unpooled.copiedBuffers() are less expensive now.
Motivation:
While porting some changes from 4.0 to 4.1 and master branch I changed the default allocator from pooled to unpooled by mistake. This should be reverted. The guilty commit is 4a3ef90381.
Thanks to @blucas for spotting this.
Modifications:
Revert changes related to allocator.
Result:
Use the correct default allocator again.
Motivation:
We did various changes related to the ChannelOutboundBuffer in 4.0 branch. This commit port all of them over and so make sure our branches are synced in terms of these changes.
Related to [#2734], [#2709], [#2729], [#2710] and [#2693] .
Modification:
Port all changes that was done on the ChannelOutboundBuffer.
This includes the port of the following commits:
- 73dfd7c01b
- 997d8c32d2
- e282e504f1
- 5e5d1a58fd
- 8ee3575e72
- d6f0d12a86
- 16e50765d1
- 3f3e66c31a
Result:
- Less memory usage by ChannelOutboundBuffer
- Same code as in 4.0 branch
- Make it possible to use ChannelOutboundBuffer with Channel implementation that not extends AbstractChannel
Modifications:
- Added a static modifier for CompositeByteBuf.Component.
This class is an inner class, but does not use its embedded reference to the object which created it. This reference makes the instances of the class larger, and may keep the reference to the creator object alive longer than necessary.
- Removed unnecessary boxing/unboxing operations in HttpResponseDecoder, RtspResponseDecoder, PerMessageDeflateClientExtensionHandshaker and PerMessageDeflateServerExtensionHandshaker
A boxed primitive is created from a String, just to extract the unboxed primitive value.
- Removed unnecessary 3 times calculations in DiskAttribute.addContent(...).
- Removed unnecessary checks if file exists before call mkdirs() in NativeLibraryLoader and PlatformDependent.
Because the method mkdirs() has this check inside.
- Removed unnecessary `instanceof AsciiString` check in StompSubframeAggregator.contentLength(StompHeadersSubframe) and StompSubframeDecoder.getContentLength(StompHeaders, long).
Because StompHeaders.get(CharSequence) always returns java.lang.String.
Motivation:
I introduced ensureAccessible() class as part of 6c47cc9711 in some places. Unfortunally I also added some where these are not needed and so caused a performance regression.
Modification:
Remove calls where not needed.
Result:
Fixed performance regression.
Motivation:
I introduced range checks as part of 6c47cc9711 in some places. Unfortunally I also added some where these are not needed and so caused a performance regression.
Modification:
Remove range checks where not needed
Result:
Fixed performance regression.
Motivation:
CompositeByteBuf.deallocate generates unnecessary GC pressure when using the 'foreach' loop, as a 'foreach' loop creates an iterator when looping.
Modification:
Convert 'foreach' loop into regular 'for' loop.
Result:
Less GC pressure (and possibly more throughput) as the 'for' loop does not create an iterator
Motivation:
AbstractByteBufTest.testInternalBuffer() uses writeByte() operations to
populate the sample data. Usually, this isn't a problem, but it starts
to take a lot of time when the resource leak detection level gets
higher.
In our CI machine, testInternalBuffer() takes more than 30 minutes,
causing the build timeout when the 'leak' profile is active (paranoid
level resource detection.)
Modification:
Populate the sample data using ThreadLocalRandom.nextBytes() instead of
using millions of writeByte() operations.
Result:
Test runs much faster when leak detection level is high.
Motivation:
Because of how we use reference counting we need to check for the reference count before each operation that touches the underlying memory. This is especially true as we use sun.misc.Cleaner.clean() to release the memory ASAP when possible. Because of this the user may cause a SEGFAULT if an operation is called that tries to access the backing memory after it was released.
Modification:
Correctly check the reference count on all methods that access the underlying memory or expose it via a ByteBuffer.
Result:
Safer usage of ByteBuf
- Using short[] for memoryMap did not improve performance. Reverting
back to the original dual-byte[] structure in favor of simplicity.
- Optimize allocateRun() which yields small performence improvement
- Use local variable when member fields are accessed more than once
Motivation:
Depth-first search is not always efficient for buddy allocation.
Modification:
Employ a new faster search algorithm with different memoryMap layout.
Result:
With thread-local cache disabled, we see a lot of performance
improvment, especially when the size of the allocation is as small as
the page size, which had the largest search space previously.
Motivation:
Persuit for the consistency in method naming
Modifications:
- Remove the 'get' prefix from all HTTP/SPDY message classes
- Fix some inspector warnings
Result:
Consistency
Motivation:
MessageToByteEncoder always starts with ByteBuf that use initalCapacity == 0 when preferDirect is used. This is really wasteful in terms of performance as every first write into the buffer will cause an expand of the buffer itself.
Modifications:
- Change ByteBufAllocator.ioBuffer() use the same default initialCapacity as heapBuffer() and directBuffer()
- Add new allocateBuffer method to MessageToByteEncoder that allow the user to do some smarter allocation based on the message that will be encoded.
Result:
Less expanding of buffer and more flexibilty when allocate the buffer for encoding.
Motivation:
Depth-first search is not always efficient for buddy allocation.
Modification:
Employ a new faster search algorithm with different memoryMap layout.
Result:
With thread-local cache disabled, we see a lot of performance
improvment, especially when the size of the allocation is as small as
the page size, which had the largest search space previously:
-- master head --
Benchmark (size) Mode Score Error Units
pooledDirectAllocAndFree 8192 thrpt 215.392 1.565 ops/ms
pooledDirectAllocAndFree 16384 thrpt 594.625 2.154 ops/ms
pooledDirectAllocAndFree 65536 thrpt 1221.520 18.965 ops/ms
pooledHeapAllocAndFree 8192 thrpt 217.175 1.653 ops/ms
pooledHeapAllocAndFree 16384 thrpt 587.250 14.827 ops/ms
pooledHeapAllocAndFree 65536 thrpt 1217.023 44.963 ops/ms
-- changes --
Benchmark (size) Mode Score Error Units
pooledDirectAllocAndFree 8192 thrpt 3656.744 94.093 ops/ms
pooledDirectAllocAndFree 16384 thrpt 4087.152 22.921 ops/ms
pooledDirectAllocAndFree 65536 thrpt 4058.814 29.276 ops/ms
pooledHeapAllocAndFree 8192 thrpt 3640.355 44.418 ops/ms
pooledHeapAllocAndFree 16384 thrpt 4030.206 24.365 ops/ms
pooledHeapAllocAndFree 65536 thrpt 4103.991 70.991 ops/ms
Motivation:
To improve the speed of ByteBuf with order LITTLE_ENDIAN and where the native order is also LITTLE_ENDIAN (intel) we introduces a new special SwappedByteBuf before in commit 4ad3984c8b. Unfortunally the commit has a flaw which does not handle correctly the case when a ByteBuf expands. This was caused because the memoryAddress was cached and never changed again even if the underlying buffer expanded. This can lead to corrupt data or even to SEGFAULT the JVM if you are lucky enough.
Modification:
Always lookup the actual memoryAddress of the wrapped ByteBuf.
Result:
No more data-corruption for ByteBuf with order LITTLE_ENDIAN and no JVM crashes.
Motivation:
When Netty runs in a managed environment such as web application server,
Netty needs to provide an explicit way to remove the thread-local
variables it created to prevent class loader leaks.
FastThreadLocal uses different execution paths for storing a
thread-local variable depending on the type of the current thread.
It increases the complexity of thread-local removal.
Modifications:
- Moved FastThreadLocal and FastThreadLocalThread out of the internal
package so that a user can use it.
- FastThreadLocal now keeps track of all thread local variables it has
initialized, and calling FastThreadLocal.removeAll() will remove all
thread-local variables of the caller thread.
- Added FastThreadLocal.size() for diagnostics and tests
- Introduce InternalThreadLocalMap which is a mixture of hard-wired
thread local variable fields and extensible indexed variables
- FastThreadLocal now uses InternalThreadLocalMap to implement a
thread-local variable.
- Added ThreadDeathWatcher.unwatch() so that PooledByteBufAllocator
tells it to stop watching when its thread-local cache has been freed
by FastThreadLocal.removeAll().
- Added FastThreadLocalTest to ensure that removeAll() works
- Added microbenchmark for FastThreadLocal and JDK ThreadLocal
- Upgraded to JMH 0.9
Result:
- A user can remove all thread-local variables Netty created, as long as
he or she did not exit from the current thread. (Note that there's no
way to remove a thread-local variable from outside of the thread.)
- FastThreadLocal exposes more useful operations such as isSet() because
we always implement a thread local variable via InternalThreadLocalMap
instead of falling back to JDK ThreadLocal.
- FastThreadLocalBenchmark shows that this change improves the
performance of FastThreadLocal even more.
Motivation:
Currently we have the algorithm of calculate the new capacity of a ByteBuf implemented in AbstractByteBuf. The problem with this is that it is impossible for a user to change it if it not fits well it's use-case. We should better move it to ByteBufAllocator and so let the user implement it's own by either write his/her own ByteBufAllocator or just override the default implementation in one of our provided ByteBufAllocators.
Modifications:
Move calculateNewCapacity(...) to ByteBufAllocator and move the implementation (which was part of AbstractByteBuf) to AbstractByteBufAllocator.
Result:
The user can now override the default calculation algorithm when needed.
Motivation:
UnpooledUnsafeDirectByteBuf.setBytes(int,ByteBuf,int,int) fails to use fast-path when src uses an array as backing storage. This is because the if else uses the wrong ByteBuf for its check.
Modifications:
- Use correct ByteBuf when check for array as backing storage
- Also eliminate unecessary check in UnpooledDirectByteBuf which always fails anyway
Result:
Faster setBytes(...) when src ByteBuf is backed by an array.
No more IndexOutOfBoundsException or data-corruption.
Motivation:
Provide a faster ThreadLocal implementation
Modification:
Add a "FastThreadLocal" which uses an EnumMap and a predefined fixed set of possible thread locals (all of the static instances created by netty) that is around 10-20% faster than standard ThreadLocal in my benchmarks (and can be seen having an effect in the direct PooledByteBufAllocator benchmark that uses the DEFAULT ByteBufAllocator which uses this FastThreadLocal, as opposed to normal instantiations that do not, and in the new RecyclableArrayList benchmark);
Result:
Improved performance
Motivation:
Our Unsafe*ByteBuf implementation always invert bytes when the native ByteOrder is LITTLE_ENDIAN (this is true on intel), even when the user calls order(ByteOrder.LITTLE_ENDIAN). This is not optimal for performance reasons as the user should be able to set the ByteOrder to LITTLE_ENDIAN and so write bytes without the extra inverting.
Modification:
- Introduce a new special SwappedByteBuf (called UnsafeDirectSwappedByteBuf) that is used by all the Unsafe*ByteBuf implementation and allows to write without inverting the bytes.
- Add benchmark
- Upgrade jmh to 0.8
Result:
The user is be able to get the max performance even on servers that have ByteOrder.LITTLE_ENDIAN as their native ByteOrder.
Motivation:
PooledByteBufAllocator's thread local cache and
ReferenceCountUtil.releaseLater() are in need of a way to run an
arbitrary logic when a certain thread is terminated.
Modifications:
- Add ThreadDeathWatcher, which spawns a low-priority daemon thread
that watches a list of threads periodically (every second) and
invokes the specified tasks when the associated threads are not alive
anymore
- Start-stop logic based on CAS operation proposed by @tea-dragon
- Add debug-level log messages to see if ThreadDeathWatcher works
Result:
- Fixes#2519 because we don't use GlobalEventExecutor anymore
- Cleaner code
Motivation:
If we make allocateRun/SubpageSimple() always try the left node first and make allocateRun/Subpage() always tries the right node first, it is more likely that allocateRun/Subpage() will find a node with ST_UNUSED sooner.
Modifications:
- Make allocateRunSimple() and allocateSubpageSimple() always try the left node first.
- Make allocateRun() and allocateSubpage() always try the right node first.
- Remove randome
Result:
We get the same performance without using random numbers.
Motivation:
We still have a room for improvement in PoolChunk.allocateRun() and
Subpage.allocate().
Modifications:
- Unroll the recursion in PoolChunk.allocateRun()
- Subpage.allocate() makes use of the 'nextAvail' value set by previous
free().
Result:
- PoolChunk.allocateRun() optimization yields 10%+ improvements in
allocation throughput for non-subpage allocations.
- Subpage.allocate() optimization makes the subpage allocations for
tiny buffers as fast as non-tiny buffers even when the pageSize is
huge (e.g. 1048576) because it doesn't need to perform a linear search
in most cases.
Motivation:
4 and 5 were diverged long time ago and we recently reverted some of the
early commits in master. We must make sure 4.1 and master are not very
different now.
Modification:
Fix found differences
Result:
4.1 and master got closer.
Motivation:
PoolArena's 'normalizeCapacity' function was micro-optimized some
time ago to remove a while loop. However, there was a change of
behavior in the function as a result. Capacities passed into it
that are already powers of 2 (and >= 512) are doubled in size. So
if I ask for a buffer with a capacity of 1024, I will get back one
that actually uses 2048 bytes (stored in maxLength).
Aligning to powers of two for book keeping ease is reasonable,
and if someone tries to expand a buffer, you might as well use some
of the previously wasted space. However, since this distinction
between 'easily expanded' and 'costly to expand' space is not
supported at all by the APIs, I cannot imagine this change to
doubling is desirable or intentional.
This is especially costly when using composite buffers. They
frequently allocate components with a capacity that is a power of
2, and they never attempt to expand components themselves. The end
result is that heavy use of pool-backed composite buffers wastes
almost half of the memory pool (the smaller / initial components are
<512 and so are not affected by the off-by-one bug).
Modifications:
Although I find it difficult to believe that such an optimization
is really helpful, I left it in and fixed the off-by-one issue by
decrementing the value at the start.
I also added a simple test to both attempt to verify that the
decrement fixes the issue without introducing any other change, and
to make it easy for a reviewer to test the existing behavior. PoolArena
does not seem to have much testing or testability support though so
the test is kind of a hack and will break for unrelated changes. I
suggest either removing it or factoring out the single non-static
portion of normalizeCapacity so that the fragile dummy PoolArena is
not required.
Result:
Pooled allocators will allocate less resources to the highly
inefficient and undocumented buffer section between length and
maxLength.
Composite buffers of non-trivial size that are backed by pooled
allocators will use about half as much memory.
Motivation:
At the moment we create new ThreadPoolCache whenever a Thread tries either allocate or release something on the PooledByteBufAllocator. When something is released we put it then in its ThreadPoolCache. The problem is we never check if a Thread is not alive anymore and so we may end up with memory that is never freed again if a user create many short living Threads that use the PooledByteBufAllocator.
Modifications:
Periodically check if the Thread is still alive that has a ThreadPoolCache assinged and if not free it.
Result:
Memory is freed up correctly even for short living Threads.
Motivation:
Remove the synchronization bottleneck in PoolArena and so speed up things
Modifications:
This implementation uses kind of the same technics as outlined in the jemalloc paper and jemalloc
blogpost https://www.facebook.com/notes/facebook-engineering/scalable-memory-allocation-using-jemalloc/480222803919.
At the moment we only cache for "known" Threads (that powers EventExecutors) and not for others to keep the overhead
minimal when need to free up unused buffers in the cache and free up cached buffers once the Thread completes. Here
we use multi-level caches for tiny, small and normal allocations. Huge allocations are not cached at all to keep the
memory usage at a sane level. All the different cache configurations can be adjusted via system properties or the constructor
directly where it makes sense.
Result:
Less conditions as most allocations can be served by the cache itself
Motivation:
6e8ba291cf introduced a regression in Android because Android does not have sun.nio.ch.DirectBuffer (see #2330.) I also found PlatformDependent0.freeDirectBuffer() and freeDirectBufferUnsafe() are pretty much same after the commit and the unsafe version should be removed.
Modifications:
- Do not use the pooled allocator in Android because it's too resource hungry for Androids.
- Merge PlatformDependent0.freeDirectBuffer() and freeDirectBufferUnsafe() into one method.
- Make the Unsafe unavailable when sun.nio.ch.DirectBuffer is unavailable. We could keep the Unsafe available and handle the sun.nio.ch.DirectBuffer case separately, but I don't want to complicate our code just because of that. All supported JDK versions have sun.nio.ch.DirectBuffer if the Unsafe is available.
Result:
Simpler code. Fixes Android support (#2330)
Motivation:
I was studying the code and thought this was simpler and easier to
understand.
Modifications:
Replaced the for loop and if conditions, with a simple implementation.
Result:
Code is easier to understand.
Motivation:
When starting with a read-only NIO buffer, wrapping it in a ByteBuf,
and then later retrieving a re-wrapped NIO buffer the limit was getting
too short.
Modifications:
Changed ReadOnlyByteBufferBuf.nioBuffer(int,int) to compute the
limit in the same manner as the internalNioBuffer method.
Result:
Round-trip conversion from NIO to ByteBuf to NIO will work reliably.
