Motivation:
At the moment we use 1 * cores as default mimimum for pool arenas. This can easily lead to conditions as we use 2 * cores as default for EventLoop's when using NIO or EPOLL. If we choose a smaller number we will run into hotspots as allocation and deallocation needs to be synchronized on the PoolArena.
Modifications:
Change the default number of arenas to 2 * cores.
Result:
Less conditions when using the default settings.
Motivation:
Dumping the content of a ByteBuf in a hex format is very useful.
Modifications:
Move code into ByteBufUtil so its easy to reuse.
Result:
Easy to reuse dumping code.
Motivation:
PoolThreadCache did only cache allocations if the allocation and deallocation Thread were the same. This is not optimal as often people write from differen thread then the actual EventLoop thread.
Modification:
- Add MpscArrayQueue which was forked from jctools and lightly modified.
- Use MpscArrayQueue for caches and always add buffer back to the cache that belongs to the allocation thread.
Result:
ThreadPoolCache is now also usable and so gives performance improvements when allocation and deallocation thread are different.
Performance when using same thread for allocation and deallocation is noticable worse then before.
Motivation:
Currently we hold a lock on the PoolArena when we allocate / free PoolSubpages, which is wasteful as this also affects "normal" allocations. The same is true vice-verse.
Modifications:
Ensure we synchronize on the head of the PoolSubPages pool. This is done per size and so it is possible to concurrently allocate / deallocate PoolSubPages with different sizes, and also normal allocations.
Result:
Less condition and so faster allocation/deallocation.
Before this commit:
xxx:~/wrk $ ./wrk -H 'Connection: keep-alive' -d 120 -c 256 -t 16 -s scripts/pipeline-many.lua http://xxx:8080/plaintext
Running 2m test @ http://xxx:8080/plaintext
16 threads and 256 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 17.61ms 29.52ms 689.73ms 97.27%
Req/Sec 278.93k 41.97k 351.04k 84.83%
530527460 requests in 2.00m, 71.64GB read
Requests/sec: 4422226.13
Transfer/sec: 611.52MB
After this commit:
xxx:~/wrk $ ./wrk -H 'Connection: keep-alive' -d 120 -c 256 -t 16 -s scripts/pipeline-many.lua http://xxx:8080/plaintext
Running 2m test @ http://xxx:8080/plaintext
16 threads and 256 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 15.85ms 24.50ms 681.61ms 97.42%
Req/Sec 287.14k 38.39k 360.33k 85.88%
547902773 requests in 2.00m, 73.99GB read
Requests/sec: 4567066.11
Transfer/sec: 631.55MB
This is reproducable every time.
Motiviation:
At the moment we sometimes hold the lock on the PoolArena during destroy a PoolChunk. This is not needed.
Modification:
- Ensure we not hold the lock during destroy a PoolChunk
- Move all synchronized usage in PoolArena
- Cleanup
Result:
Less condition.
Motivation:
The PooledByteBufAllocator is more or less a black-box atm. We need to expose some metrics to allow the user to get a better idea how to tune it.
Modifications:
- Expose different metrics via PooledByteBufAllocator
- Add *Metrics interfaces
Result:
It is now easy to gather metrics and detail about the PooledByteBufAllocator and so get a better understanding about resource-usage etc.
Motivation:
At the moment when calling slice(...) or duplicate(...) on a Pooled*ByteBuf a new SlicedByteBuf or DuplicatedByteBuf. This can create a lot of GC.
Modifications:
Add PooledSlicedByteBuf and PooledDuplicatedByteBuf which will be used when a PooledByteBuf is used.
Result:
Less GC.
Motivation:
From the javadocs of ByteBuf.duplicate() it is not clear if the reader and writer marks will be duplicated.
Modifications:
Add sentence to clarify that marks will not be duplicated.
Result:
Clear semantics.
Motivation:
When allocate a PooledByteBuf we need to ensure to also reset the markers for the readerIndex and writerIndex.
Modifications:
- Correct reset the markers
- Add test-case for it
Result:
Correctly reset markers.
Motiviation:
When tried to allocate tiny and small sized and failed to serve these out of the PoolSubPage we exit the synchronization
block just to enter it again when call allocateNormal(...).
Modification:
Not exit the synchronized block until allocateNormal(...) is done.
Result:
Better performance.
Motivation:
The way of firstIndexOf and lastIndexOf iterating the ByteBuf is similar to forEachByte and forEachByteDesc, but have many range checks.
Modifications:
Use forEachByte and a IndexOfProcessor to find occurrence.
Result:
eliminate range checks
Motivation:
The ByteString class currently assumes the underlying array will be a complete representation of data. This is limiting as it does not allow a subsection of another array to be used. The forces copy operations to take place to compensate for the lack of API support.
Modifications:
- add arrayOffset method to ByteString
- modify all ByteString and AsciiString methods that loop over or index into the underlying array to use this offset
- update all code that uses ByteString.array to ensure it accounts for the offset
- add unit tests to test the implementation respects the offset
Result:
ByteString and AsciiString can represent a sub region of a byte[].
Motivation:
CompositeByteBuf.iterator() currently creates a new ArrayList and fill it with the ByteBufs, which is more expensive then it needs to be.
Modifications:
- Use special Iterator implementation
Result:
Less overhead when calling iterator()
Motivation:
The usage and code within AsciiString has exceeded the original design scope for this class. Its usage as a binary string is confusing and on the verge of violating interface assumptions in some spots.
Modifications:
- ByteString will be created as a base class to AsciiString. All of the generic byte handling processing will live in ByteString and all the special character encoding will live in AsciiString.
Results:
The AsciiString interface will be clarified. Users of AsciiString can now be clear of the limitations the class imposes while users of the ByteString class don't have to live with those limitations.
Motivation:
When create NormalMemoryRegionCache for PoolThreadCache, we overbooked
cache array size. This means unnecessary overhead for thread local cache
as we will create multi cache enties for each element in cache array.
Modifications:
change:
int arraySize = Math.max(1, max / area.pageSize);
to:
int arraySize = Math.max(1, log2(max / area.pageSize) + 1);
Result:
Now arraySize won't introduce unnecessary overhead.
Changes to be committed:
modified: buffer/src/main/java/io/netty/buffer/PoolThreadCache.java
Motivation:
CompositeByteBuf has an iterator() method but fails to implement Iterable
Modifications:
Let CompositeByteBuf implement Iterable<ByteBuf>
Result:
Easier usage
Motivation:
We missed to dereference the chunk and tmpNioBuf when calling deallocate(). This means the GC can not collect these as we still hold a reference while have the PooledByteBuf in the recycler stack.
Modifications:
Dereference chunk and tmpNioBuf.
Result:
GC can collect things.
Motiviation:
At the moment we use FIFO for the PoolThreadCache which is sub-optimal as this may reduce the changes to have the cached memory actual still in the cpu-cache.
Modification:
- Change to use LIFO as this increase the chance to be able to serve buffers from the cpu-cache
Results:
Faster allocation out of the ThreadLocal cache.
Before the commit:
[xxx wrk]$ ./wrk -H 'Connection: keep-alive' -d 120 -c 256 -t 16 -s scripts/pipeline-many.lua http://xxx:8080/plaintext
Running 2m test @ http://xxx:8080/plaintext
16 threads and 256 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 14.69ms 10.06ms 131.43ms 80.10%
Req/Sec 283.89k 40.37k 433.69k 66.81%
533859742 requests in 2.00m, 72.09GB read
Requests/sec: 4449510.51
Transfer/sec: 615.29MB
After the commit:
[xxx wrk]$ ./wrk -H 'Connection: keep-alive' -d 120 -c 256 -t 16 -s scripts/pipeline-many.lua http://xxx:8080/plaintext
Running 2m test @ http://xxx:8080/plaintext
16 threads and 256 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 16.38ms 26.32ms 734.06ms 97.38%
Req/Sec 283.86k 39.31k 361.69k 83.38%
540836511 requests in 2.00m, 73.04GB read
Requests/sec: 4508150.18
Transfer/sec: 623.40MB
Motivation:
The DefaultHttp2ConnectionDecoder class is calling verifyPrefaceReceived() for almost every frame event at all times.
The Http2ConnectionHandler class is calling readClientPrefaceString() on every decode event.
Modifications:
- DefaultHttp2ConnectionDecoder should not have to continuously call verifyPrefaceReceived() because it transitions boolean state 1 time for each connection.
- Http2ConnectionHandler should not have to continuously call readClientPrefaceString() because it transitions boolean state 1 time for each connection.
Result:
- Less conditional checks for the mainstream usage of the connection.
Motivation:
`Unpooled` javadoc's mentioned the generation of hex dump and swapping an integer's byte order,
which are actually provided by `ByteBufUtil`.
Modifications:
Sentence moved to `ByteBufUtil` javadoc.
Result:
`Unpooled` javadoc is correct.
Motivation:
At the moment we have two problems:
- CompositeByteBuf.addComponent(...) will not add the supplied buffer to the CompositeByteBuf if its empty, which means it will not be released on CompositeByteBuf.release() call. This is a problem as a user will expect everything added will be released (the user not know we not added it).
- CompositeByteBuf.addComponents(...) will either add no buffers if none is readable and so has the same problem as addComponent(...) or directly release the ByteBuf if at least one ByteBuf is readable. Again this gives inconsistent handling and may lead to memory leaks.
Modifications:
- Always add the buffer to the CompositeByteBuf and so release it on release call.
Result:
Consistent handling and no buffer leaks.
Motivation:
When a CompositeByteBuf is empty (i.e. has no component), its internal
memory access operations do not always behave as expected.
Modifications:
Check if the nunmber of components is zero. If so, return an empty
array or an empty NIO buffer, etc.
Result:
More robustness
- Ensure an EmptyByteBuf has an array, an NIO buffer, and a memory
address at the same time
- Add an assertion that checks if EMPTY_BUFFER is an EmptyByteBuf,
just in case we make a mistake in the future
Motivation:
We expose no methods in ByteBuf to directly write a CharSequence into it. This leads to have the user either convert the CharSequence first to a byte array or use CharsetEncoder. Both cases have some overheads and we can do a lot better for well known Charsets like UTF-8 and ASCII.
Modifications:
Add ByteBufUtil.writeAscii(...) and ByteBufUtil.writeUtf8(...) which can do the task in an optimized way. This is especially true if the passed in ByteBuf extends AbstractByteBuf which is true for all of our implementations which not wrap another ByteBuf.
Result:
Writing an ASCII and UTF-8 CharSequence into a AbstractByteBuf is a lot faster then what the user could do by himself as we can make use of some package private methods and so eliminate reference and range checks. When the Charseq is not ASCII or UTF-8 we can still do a very good job and are on par in most of the cases with what the user would do.
The following benchmark shows the improvements:
Result: 2456866.966 ?(99.9%) 59066.370 ops/s [Average]
Statistics: (min, avg, max) = (2297025.189, 2456866.966, 2586003.225), stdev = 78851.914
Confidence interval (99.9%): [2397800.596, 2515933.336]
Benchmark Mode Samples Score Score error Units
i.n.m.b.ByteBufUtilBenchmark.writeAscii thrpt 50 9398165.238 131503.098 ops/s
i.n.m.b.ByteBufUtilBenchmark.writeAsciiString thrpt 50 9695177.968 176684.821 ops/s
i.n.m.b.ByteBufUtilBenchmark.writeAsciiStringViaArray thrpt 50 4788597.415 83181.549 ops/s
i.n.m.b.ByteBufUtilBenchmark.writeAsciiStringViaArrayWrapped thrpt 50 4722297.435 98984.491 ops/s
i.n.m.b.ByteBufUtilBenchmark.writeAsciiStringWrapped thrpt 50 4028689.762 66192.505 ops/s
i.n.m.b.ByteBufUtilBenchmark.writeAsciiViaArray thrpt 50 3234841.565 91308.009 ops/s
i.n.m.b.ByteBufUtilBenchmark.writeAsciiViaArrayWrapped thrpt 50 3311387.474 39018.933 ops/s
i.n.m.b.ByteBufUtilBenchmark.writeAsciiWrapped thrpt 50 3379764.250 66735.415 ops/s
i.n.m.b.ByteBufUtilBenchmark.writeUtf8 thrpt 50 5671116.821 101760.081 ops/s
i.n.m.b.ByteBufUtilBenchmark.writeUtf8String thrpt 50 5682733.440 111874.084 ops/s
i.n.m.b.ByteBufUtilBenchmark.writeUtf8StringViaArray thrpt 50 3564548.995 55709.512 ops/s
i.n.m.b.ByteBufUtilBenchmark.writeUtf8StringViaArrayWrapped thrpt 50 3621053.671 47632.820 ops/s
i.n.m.b.ByteBufUtilBenchmark.writeUtf8StringWrapped thrpt 50 2634029.071 52304.876 ops/s
i.n.m.b.ByteBufUtilBenchmark.writeUtf8ViaArray thrpt 50 3397049.332 57784.119 ops/s
i.n.m.b.ByteBufUtilBenchmark.writeUtf8ViaArrayWrapped thrpt 50 3318685.262 35869.562 ops/s
i.n.m.b.ByteBufUtilBenchmark.writeUtf8Wrapped thrpt 50 2473791.249 46423.114 ops/s
Tests run: 1, Failures: 0, Errors: 0, Skipped: 0, Time elapsed: 1,387.417 sec - in io.netty.microbench.buffer.ByteBufUtilBenchmark
Results :
Tests run: 1, Failures: 0, Errors: 0, Skipped: 0
Results :
Tests run: 1, Failures: 0, Errors: 0, Skipped: 0
The *ViaArray* benchmarks are basically doing a toString().getBytes(Charset) which the others are using ByteBufUtil.write*(...).
Motivation:
CompositeByteBuf.nioBuffers(...) returns an empty ByteBuffer array if the specified length is 0. This is not consistent with other ByteBuf implementations which return an ByteBuffer array of size 1 with an empty ByteBuffer included.
Modifications:
Make CompositeByteBuf.nioBuffers(...) consistent with other ByteBuf implementations.
Result:
Consistent and correct behaviour of nioBufffers(...)
Motivation:
When calling slice(...) on a ByteBuf the returned ByteBuf should be the slice of a ByteBuf and shares it's reference count. This is important as it is perfect legal to use buf.slice(...).release() and have both, the slice and the original ByteBuf released. At the moment this is only the case if the requested slice size is > 0. This makes the behavior inconsistent and so may lead to a memory leak.
Modifications:
- Never return Unpooled.EMPTY_BUFFER when calling slice(...).
- Adding test case for buffer.slice(...).release() and buffer.duplicate(...).release()
Result:
Consistent behaviour and so no more leaks possible.
Motivation:
Before we missed to check if a buffer was released before we return the backing byte array or memoryaddress. This could lead to JVM crashes when someone tried various bulk operations on the Unsafe*ByteBuf implementations.
Modifications:
Always check if the buffer is released before all to return the byte array and memoryaddress.
Result:
No more JVM crashes because of released buffers when doing bulk operations on Unsafe*ByteBuf implementations.
Motivation:
Found performance issues via FindBugs and PMD.
Modifications:
- Removed unnecessary boxing/unboxing operations in DefaultTextHeaders.convertToInt(CharSequence) and DefaultTextHeaders.convertToLong(CharSequence). A boxed primitive is created from a string, just to extract the unboxed primitive value.
- Added a static modifier for DefaultHttp2Connection.ParentChangedEvent class. 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.
- Added a static compiled Pattern to avoid compile it each time it is used when we need to replace some part of authority.
- Improved using of StringBuilders.
Result:
Performance improvements.
Motivation:
We introduced a PoolThreadCache which is used in our PooledByteBufAllocator to reduce the synchronization overhead on PoolArenas when allocate / deallocate PooledByteBuf instances. This cache is used for both the allocation path and deallocation path by:
- Look for cached memory in the PoolThreadCache for the Thread that tries to allocate a new PooledByteBuf and if one is found return it.
- Add the memory that is used by a PooledByteBuf to the PoolThreadCache of the Thread that release the PooledByteBuf
This works out very well when all allocation / deallocation is done in the EventLoop as the EventLoop will be used for read and write. On the otherside this can lead to surprising side-effects if the user allocate from outside the EventLoop and and pass the ByteBuf over for writing. The problem here is that the memory will be added to the PoolThreadCache that did the actual write on the underlying transport and not on the Thread that previously allocated the buffer.
Modifications:
Don't cache if different Threads are used for allocating/deallocating
Result:
Less confusing behavior for users that allocate PooledByteBufs from outside the EventLoop.
Motivation:
When MemoryRegionCache.trim() is called, some unused cache entries will be freed (started from head). However, in MeoryRegionCache.trim() the head is not updated, which make entry list's head point to an entry whose chunk is null now and following allocate of MeoryRegionCache will return false immediately.
In other word, cache is no longer usable once trim happen.
Modifications:
Update head to correct idx after free entries in trim().
Result:
MemoryRegionCache behaves correctly even after calling trim().
Motivation:
We received a bug-report that the ByteBuf.refCnt() does sometimes not show the correct value when release() and refCnt() is called from different Threads.
Modifications:
Add test-case which shows that all is working like expected
Result:
Test-case added which shows everything is ok.
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.
