Motivation:
A degradation in performance has been observed from the 4.0 branch as documented in https://github.com/netty/netty/issues/3962.
Modifications:
- Simplify Headers class hierarchy.
- Restore the DefaultHeaders to be based upon DefaultHttpHeaders from 4.0.
- Make various other modifications that are causing hot spots.
Result:
Performance is now on par with 4.0.
Motivation:
We noticed that the headers implementation in Netty for HTTP/2 uses quite a lot of memory
and that also at least the performance of randomly accessing a header is quite poor. The main
concern however was memory usage, as profiling has shown that a DefaultHttp2Headers
not only use a lot of memory it also wastes a lot due to the underlying hashmaps having
to be resized potentially several times as new headers are being inserted.
This is tracked as issue #3600.
Modifications:
We redesigned the DefaultHeaders to simply take a Map object in its constructor and
reimplemented the class using only the Map primitives. That way the implementation
is very concise and hopefully easy to understand and it allows each concrete headers
implementation to provide its own map or to even use a different headers implementation
for processing requests and writing responses i.e. incoming headers need to provide
fast random access while outgoing headers need fast insertion and fast iteration. The
new implementation can support this with hardly any code changes. It also comes
with the advantage that if the Netty project decides to add a third party collections library
as a dependency, one can simply plug in one of those very fast and memory efficient map
implementations and get faster and smaller headers for free.
For now, we are using the JDK's TreeMap for HTTP and HTTP/2 default headers.
Result:
- Significantly fewer lines of code in the implementation. While the total commit is still
roughly 400 lines less, the actual implementation is a lot less. I just added some more
tests and microbenchmarks.
- Overall performance is up. The current implementation should be significantly faster
for insertion and retrieval. However, it is slower when it comes to iteration. There is simply
no way a TreeMap can have the same iteration performance as a linked list (as used in the
current headers implementation). That's totally fine though, because when looking at the
benchmark results @ejona86 pointed out that the performance of the headers is completely
dominated by insertion, that is insertion is so significantly faster in the new implementation
that it does make up for several times the iteration speed. You can't iterate what you haven't
inserted. I am demonstrating that in this spreadsheet [1]. (Actually, iteration performance is
only down for HTTP, it's significantly improved for HTTP/2).
- Memory is down. The implementation with TreeMap uses on avg ~30% less memory. It also does not
produce any garbage while being resized. In load tests for GRPC we have seen a memory reduction
of up to 1.2KB per RPC. I summarized the memory improvements in this spreadsheet [1]. The data
was generated by [2] using JOL.
- While it was my original intend to only improve the memory usage for HTTP/2, it should be similarly
improved for HTTP, SPDY and STOMP as they all share a common implementation.
[1] https://docs.google.com/spreadsheets/d/1ck3RQklyzEcCLlyJoqDXPCWRGVUuS-ArZf0etSXLVDQ/edit#gid=0
[2] https://gist.github.com/buchgr/4458a8bdb51dd58c82b4
Motivation:
The HttpObjectDecoder is on the hot code path for the http codec. There are a few hot methods which can be modified to improve performance.
Modifications:
- Modify AppendableCharSequence to provide unsafe methods which don't need to re-check bounds for every call.
- Update HttpObjectDecoder methods to take advantage of new AppendableCharSequence methods.
Result:
Peformance boost for decoding http objects.
Motivation:
See #3783
Modifications:
- The DefaultHttp2RemoteFlowController should use Channel.isWritable() before attempting to do any write operations.
- The Flow controller methods should no longer take ChannelHandlerContext. The concept of flow control is tied to a connection and we do not support 1 flow controller keeping track of multiple ChannelHandlerContext.
Result:
Writes are delayed until isWritable() is true. Flow controller interface methods are more clear as to ChannelHandlerContext restrictions.
Motivation:
Coalescing many small writes into a larger DATA frame reduces framing overheads on the wire and reduces the number of calls to Http2FrameListeners on the remote side.
Delaying the write of WINDOW_UPDATE until flush allows for more consumed bytes to be returned as the aggregate of consumed bytes is returned and not the amount consumed when the threshold was crossed.
Modifications:
- Remote flow controller no longer immediately writes bytes when a flow-controlled payload is enqueued. Sequential data payloads are now merged into a single CompositeByteBuf which are written when 'writePendingBytes' is called.
- Listener added to remote flow-controller which observes written bytes per stream.
- Local flow-controller no longer immediately writes WINDOW_UPDATE when the ratio threshold is crossed. Now an explicit call to 'writeWindowUpdates' triggers the WINDOW_UPDATE for all streams who's ratio is exceeded at that time. This results in
fewer window updates being sent and more bytes being returned.
- Http2ConnectionHandler.flush triggers 'writeWindowUpdates' on the local flow-controller followed by 'writePendingBytes' on the remote flow-controller so WINDOW_UPDATES preceed DATA frames on the wire.
Result:
- Better throughput for writing many small DATA chunks followed by a flush, saving 9-bytes per coalesced frame.
- Fewer WINDOW_UPDATES being written and more flow-control bytes returned to remote side more quickly, thereby improving throughput.
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:
Streams currently maintain a hash map of user-defined properties, which has been shown to add significant memory overhead as well as being a performance bottleneck for lookup of frequently used properties.
Modifications:
Modifying the connection/stream to use an array as the storage of user-defined properties, indexed by the class that identifies the index into the array where the property is stored.
Result:
Stream processing performance should be improved.
Motivation:
Flow control is a required part of the HTTP/2 specification but it is currently structured more like an optional item. It must be accessed through the property map which is time consuming and does not represent its required nature. This access pattern does not give any insight into flow control outside of the codec (or flow controller implementation).
Modifications:
1. Create a read only public interface for LocalFlowState and RemoteFlowState.
2. Add a LocalFlowState localFlowState(); and RemoteFlowState remoteFlowState(); to Http2Stream.
Result:
Flow control is not part of the Http2Stream interface. This clarifies its responsibility and logical relationship to other interfaces. The flow controller no longer must be acquired though a map lookup.
Motivation:
There is no benchmark to measure the priority tree implementation performance.
Modifications:
Introduce a new benchmark which will populate the priority tree, and then shuffle parent/child links around.
Result:
A simple benchmark to get a baseline for the HTTP/2 codec's priority tree implementation.
Motivation:
Allows for running benchmarks from built jars which is useful in development environments that only take released artifacts.
Modifications:
Move benchmarks into 'main' from 'test'
Add @State annotations to benchmarks that are missing them
Fix timing issue grabbing context during channel initialization
Result:
Users can run benchmarks more easily.
Motivation:
The current implementation does byte by byte comparison, which we have seen
can be a performance bottleneck when the AsciiString is used as the key in
a Map.
Modifications:
Use sun.misc.Unsafe (on supporting platforms) to compare up to eight bytes at a time
and get closer to the performance of String.equals(Object).
Result:
Significant improvement (2x - 6x) in performance over the current implementation.
Benchmark (size) Mode Samples Score Score error Units
i.n.m.i.PlatformDependentBenchmark.arraysBytesEqual 10 thrpt 10 118843477.518 2347259.347 ops/s
i.n.m.i.PlatformDependentBenchmark.arraysBytesEqual 50 thrpt 10 43910319.773 198376.996 ops/s
i.n.m.i.PlatformDependentBenchmark.arraysBytesEqual 100 thrpt 10 26339969.001 159599.252 ops/s
i.n.m.i.PlatformDependentBenchmark.arraysBytesEqual 1000 thrpt 10 2873119.030 20779.056 ops/s
i.n.m.i.PlatformDependentBenchmark.arraysBytesEqual 10000 thrpt 10 306370.450 1933.303 ops/s
i.n.m.i.PlatformDependentBenchmark.arraysBytesEqual 100000 thrpt 10 25750.415 108.391 ops/s
i.n.m.i.PlatformDependentBenchmark.unsafeBytesEqual 10 thrpt 10 248077563.510 635320.093 ops/s
i.n.m.i.PlatformDependentBenchmark.unsafeBytesEqual 50 thrpt 10 128198943.138 614827.548 ops/s
i.n.m.i.PlatformDependentBenchmark.unsafeBytesEqual 100 thrpt 10 86195621.349 1063959.307 ops/s
i.n.m.i.PlatformDependentBenchmark.unsafeBytesEqual 1000 thrpt 10 16920264.598 61615.365 ops/s
i.n.m.i.PlatformDependentBenchmark.unsafeBytesEqual 10000 thrpt 10 1687454.747 6367.602 ops/s
i.n.m.i.PlatformDependentBenchmark.unsafeBytesEqual 100000 thrpt 10 153717.851 586.916 ops/s
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:
The Http2FrameWriterBenchmark JMH harness class name was not updated for the JVM arguments. The number of forks is 0 which means the JHM will share a JVM with the benchmarks. Sharing the JVM may lead to less reliable benchmarking results and as doesn't allow for the command line arguments to be applied for each benchmark.
Modifications:
- Update the JMH version from 0.9 to 1.7.1. Benchmarks wouldn't run on old version.
- Increase the number of forks from 0 to 1.
- Remove allocation of environment from static and cleanup AfterClass to using the Setup and Teardown methods. The forked JVM would not shut down correctly otherwise (and wait for 30+ seconds before timeing out).
Result:
Benchmarks that run as intended.
Motivation:
It currently takes a builder for the encoder and decoder, which makes it difficult to decorate them.
Modifications:
Removed the builders from the interfaces entirely. Left the builder for the decoder impl but removed it from the encoder since it's constructor only takes 2 parameters. Also added decorator base classes for the encoder and decoder and made the CompressorHttp2ConnectionEncoder extend the decorator.
Result:
Fixes#3530
Motivation:
The backport of a6c729bdf8 failed.
Modifications:
- Make sure the interfaces are correctly implemented when backporting.
Result:
Microbenchmark compiles and runs on 4.1 branch.
Motivation:
A microbenchmark will be useful to get a baseline for performance.
Modifications:
- Introduce a new microbenchmark which tests the Http2DefaultFrameWriter.
- Allow benchmarks to run without thread context switching between JMH and Netty.
Result:
Microbenchmark exists to test performance.
Motivation
----------
The performance tests for utf8 also used the getBytes on ASCII,
which is incorrect and also provides different performance numbers.
Modifications
-------------
Use CharsetUtil.UTF_8 instead of US_ASCII for the getBytes calls.
Result
------
Accurate and semantically correct benchmarking results on utf8
comparisons.
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:
The java implementations for Inet6Address.getHostName() do not follow the RFC 5952 (http://tools.ietf.org/html/rfc5952#section-4) for recommended string representation. This introduces inconsistencies when integrating with other technologies that do follow the RFC.
Modifications:
-NetUtil.java to have another public static method to convert InetAddress to string. Inet4Address will use the java InetAddress.getHostAddress() implementation and there will be new code to implement the RFC 5952 IPV6 string conversion.
-New unit tests to test the new method
Result:
Netty provides a RFC 5952 compliant string conversion method for IPV6 addresses
Motivation:
default*() tests are performing a test in a different way, and they must be same with other tests.
Modification:
Make sure default*() tests are same with the others
Result:
Easier to compare default and non-default allocators
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:
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:
Allocating a single buffer and releasing it repetitively for a benchmark will not involve the realistic execution path of the allocators.
Modifications:
Keep the last 8192 allocations and release them randomly.
Result:
We are now getting the result close to what we got with caliper.
The API changes made so far turned out to increase the memory footprint
and consumption while our intention was actually decreasing them.
Memory consumption issue:
When there are many connections which does not exchange data frequently,
the old Netty 4 API spent a lot more memory than 3 because it always
allocates per-handler buffer for each connection unless otherwise
explicitly stated by a user. In a usual real world load, a client
doesn't always send requests without pausing, so the idea of having a
buffer whose life cycle if bound to the life cycle of a connection
didn't work as expected.
Memory footprint issue:
The old Netty 4 API decreased overall memory footprint by a great deal
in many cases. It was mainly because the old Netty 4 API did not
allocate a new buffer and event object for each read. Instead, it
created a new buffer for each handler in a pipeline. This works pretty
well as long as the number of handlers in a pipeline is only a few.
However, for a highly modular application with many handlers which
handles connections which lasts for relatively short period, it actually
makes the memory footprint issue much worse.
Changes:
All in all, this is about retaining all the good changes we made in 4 so
far such as better thread model and going back to the way how we dealt
with message events in 3.
To fix the memory consumption/footprint issue mentioned above, we made a
hard decision to break the backward compatibility again with the
following changes:
- Remove MessageBuf
- Merge Buf into ByteBuf
- Merge ChannelInboundByte/MessageHandler and ChannelStateHandler into ChannelInboundHandler
- Similar changes were made to the adapter classes
- Merge ChannelOutboundByte/MessageHandler and ChannelOperationHandler into ChannelOutboundHandler
- Similar changes were made to the adapter classes
- Introduce MessageList which is similar to `MessageEvent` in Netty 3
- Replace inboundBufferUpdated(ctx) with messageReceived(ctx, MessageList)
- Replace flush(ctx, promise) with write(ctx, MessageList, promise)
- Remove ByteToByteEncoder/Decoder/Codec
- Replaced by MessageToByteEncoder<ByteBuf>, ByteToMessageDecoder<ByteBuf>, and ByteMessageCodec<ByteBuf>
- Merge EmbeddedByteChannel and EmbeddedMessageChannel into EmbeddedChannel
- Add SimpleChannelInboundHandler which is sometimes more useful than
ChannelInboundHandlerAdapter
- Bring back Channel.isWritable() from Netty 3
- Add ChannelInboundHandler.channelWritabilityChanges() event
- Add RecvByteBufAllocator configuration property
- Similar to ReceiveBufferSizePredictor in Netty 3
- Some existing configuration properties such as
DatagramChannelConfig.receivePacketSize is gone now.
- Remove suspend/resumeIntermediaryDeallocation() in ByteBuf
This change would have been impossible without @normanmaurer's help. He
fixed, ported, and improved many parts of the changes.
- Rename directbyDefault to preferDirect
- Add a system property 'io.netty.prederDirect' to allow a user from changing the preference on launch-time
- Merge UnpooledByteBufAllocator.DEFAULT_BY_* to DEFAULT
- Fixes#826
Unsafe.isFreed(), free(), suspend/resumeIntermediaryAllocations() are not that dangerous. internalNioBuffer() and internalNioBuffers() are dangerous but it seems like nobody is using it even inside Netty. Removing those two methods also removes the necessity to keep Unsafe interface at all.
- Add common optimization options when launching a new JVM to run a benchmark
- Fix a bug where a benchmark report is uploaded twice
- Simplify pom.xml and move the build instruction messages to DefaultBenchmark
- Print an empty line to prettify the output
This pull request introduces the new default ByteBufAllocator implementation based on jemalloc, with a some differences:
* Minimum possible buffer capacity is 16 (jemalloc: 2)
* Uses binary heap with random branching (jemalloc: red-black tree)
* No thread-local cache yet (jemalloc has thread-local cache)
* Default page size is 8 KiB (jemalloc: 4 KiB)
* Default chunk size is 16 MiB (jemalloc: 2 MiB)
* Cannot allocate a buffer bigger than the chunk size (jemalloc: possible) because we don't have control over memory layout in Java. A user can work around this issue by creating a composite buffer, but it's not always a feasible option. Although 16 MiB is a pretty big default, a user's handler might need to deal with the bounded buffers when the user wants to deal with a large message.
Also, to ensure the new allocator performs good enough, I wrote a microbenchmark for it and made it a dedicated Maven module. It uses Google's Caliper framework to run and publish the test result (example)
Miscellaneous changes:
* Made some ByteBuf implementations public so that those who implements a new allocator can make use of them.
* Added ByteBufAllocator.compositeBuffer() and its variants.
* ByteBufAllocator.ioBuffer() creates a buffer with 0 capacity.
