Motivation:
The JMH doc suggests to use BlackHoles to avoid dead code elimination hence would be better to follow this best practice.
Modifications:
Each benchmark method is returning the ByteBuf/ByteBuffer to avoid the JVM to perform any dead code elimination.
Result:
The results are more reliable and comparable to the others provided by other ByteBuf benchmarks (eg HeapByteBufBenchmark)
Motivation:
HttpMethod#valueOf shows up on profiler results in the top set of
results. Since it is a relatively simple operation it can be improved in
isolation.
Modifications:
- Introduce a special case map which assigns each HttpMethod to a unique
index in an array and provides constant time lookup from a hash code
algorithm. When the bucket is matched we can then directly do equality
comparison instead of potentially following a linked structure when
HashMap has hash collisions.
Result:
~10% improvement in benchmark results for HttpMethod#valueOf
Benchmark Mode Cnt Score Error Units
HttpMethodMapBenchmark.newMapKnownMethods thrpt 16 31.831 ± 0.928 ops/us
HttpMethodMapBenchmark.newMapMixMethods thrpt 16 25.568 ± 0.400 ops/us
HttpMethodMapBenchmark.newMapUnknownMethods thrpt 16 51.413 ± 1.824 ops/us
HttpMethodMapBenchmark.oldMapKnownMethods thrpt 16 29.226 ± 0.330 ops/us
HttpMethodMapBenchmark.oldMapMixMethods thrpt 16 21.073 ± 0.247 ops/us
HttpMethodMapBenchmark.oldMapUnknownMethods thrpt 16 49.081 ± 0.577 ops/us
Motivation:
DefaultHttp2FrameWriter#writeData allocates a DataFrameHeader for each write operation. DataFrameHeader maintains internal state and allocates multiple slices of a buffer which is a maximum of 30 bytes. This 30 byte buffer may not always be necessary and the additional slice operations can utilize retainedSlice to take advantage of pooled objects. We can also save computation and object allocations if there is no padding which is a common case in practice.
Modifications:
- Remove DataFrameHeader
- Add a fast path for padding == 0
Result:
Less object allocation in DefaultHttp2FrameWriter
Motivation:
`AbstractScheduledEventExecutor` uses a standard `java.util.PriorityQueue` to keep track of task deadlines. `ScheduledFuture.cancel` removes tasks from this `PriorityQueue`. Unfortunately, `PriorityQueue.remove` has `O(n)` performance since it must search for the item in the entire queue before removing it. This is fast when the future is at the front of the queue (e.g., already triggered) but not when it's randomly located in the queue.
Many servers will use `ScheduledFuture.cancel` on all requests, e.g., to manage a request timeout. As these cancellations will be happen in arbitrary order, when there are many scheduled futures, `PriorityQueue.remove` is a bottleneck and greatly hurts performance with many concurrent requests (>10K).
Modification:
Use netty's `DefaultPriorityQueue` for scheduling futures instead of the JDK. `DefaultPriorityQueue` is almost identical to the JDK version except it is able to remove futures without searching for them in the queue. This means `DefaultPriorityQueue.remove` has `O(log n)` performance.
Result:
Before - cancelling futures has varying performance, capped at `O(n)`
After - cancelling futures has stable performance, capped at `O(log n)`
Benchmark results
After - cancelling in order and in reverse order have similar performance within `O(log n)` bounds
```
Benchmark (num) Mode Cnt Score Error Units
ScheduledFutureTaskBenchmark.cancelInOrder 100 thrpt 20 137779.616 ± 7709.751 ops/s
ScheduledFutureTaskBenchmark.cancelInOrder 1000 thrpt 20 11049.448 ± 385.832 ops/s
ScheduledFutureTaskBenchmark.cancelInOrder 10000 thrpt 20 943.294 ± 12.391 ops/s
ScheduledFutureTaskBenchmark.cancelInOrder 100000 thrpt 20 64.210 ± 1.824 ops/s
ScheduledFutureTaskBenchmark.cancelInReverseOrder 100 thrpt 20 167531.096 ± 9187.865 ops/s
ScheduledFutureTaskBenchmark.cancelInReverseOrder 1000 thrpt 20 33019.786 ± 4737.770 ops/s
ScheduledFutureTaskBenchmark.cancelInReverseOrder 10000 thrpt 20 2976.955 ± 248.555 ops/s
ScheduledFutureTaskBenchmark.cancelInReverseOrder 100000 thrpt 20 362.654 ± 45.716 ops/s
```
Before - cancelling in order and in reverse order have significantly different performance at higher queue size, orders of magnitude worse than the new implementation.
```
Benchmark (num) Mode Cnt Score Error Units
ScheduledFutureTaskBenchmark.cancelInOrder 100 thrpt 20 139968.586 ± 12951.333 ops/s
ScheduledFutureTaskBenchmark.cancelInOrder 1000 thrpt 20 12274.420 ± 337.800 ops/s
ScheduledFutureTaskBenchmark.cancelInOrder 10000 thrpt 20 958.168 ± 15.350 ops/s
ScheduledFutureTaskBenchmark.cancelInOrder 100000 thrpt 20 53.381 ± 13.981 ops/s
ScheduledFutureTaskBenchmark.cancelInReverseOrder 100 thrpt 20 123918.829 ± 3642.517 ops/s
ScheduledFutureTaskBenchmark.cancelInReverseOrder 1000 thrpt 20 5099.810 ± 206.992 ops/s
ScheduledFutureTaskBenchmark.cancelInReverseOrder 10000 thrpt 20 72.335 ± 0.443 ops/s
ScheduledFutureTaskBenchmark.cancelInReverseOrder 100000 thrpt 20 0.743 ± 0.003 ops/s
```
Motivation:
Highly retained and released objects have contention on their ref
count. Currently, the ref count is updated using compareAndSet
with care to make sure the count doesn't overflow, double free, or
revive the object.
Profiling has shown that a non trivial (~1%) of CPU time on gRPC
latency benchmarks is from the ref count updating.
Modification:
Rather than pessimistically assuming the ref count will be invalid,
optimistically update it assuming it will be. If the update was
wrong, then use the slow path to revert the change and throw an
execption. Most of the time, the ref counts are correct.
This changes from using compareAndSet to getAndAdd, which emits a
different CPU instruction on x86 (CMPXCHG to XADD). Because the
CPU knows it will modifiy the memory, it can avoid contention.
On a highly contended machine, this can be about 2x faster.
There is a downside to the new approach. The ref counters can
temporarily enter invalid states if over retained or over released.
The code does handle these overflow and underflow scenarios, but it
is possible that another concurrent access may push the failure to
a different location. For example:
Time 1 Thread 1: obj.retain(INT_MAX - 1)
Time 2 Thread 1: obj.retain(2)
Time 2 Thread 2: obj.retain(1)
Previously Thread 2 would always succeed and Thread 1 would always
fail on the second access. Now, thread 2 could fail while thread 1
is rolling back its change.
====
There are a few reasons why I think this is okay:
1. Buggy code is going to have bugs. An exception _is_ going to be
thrown. This just causes the other threads to notice the state
is messed up and stop early.
2. If high retention counts are a use case, then ref count should
be a long rather than an int.
3. The critical section is greatly reduced compared to the previous
version, so the likelihood of this happening is lower
4. On error, the code always rollsback the change atomically, so
there is no possibility of corruption.
Result:
Faster refcounting
```
BEFORE:
Benchmark (delay) Mode Cnt Score Error Units
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 1 sample 2901361 804.579 ± 1.835 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 10 sample 3038729 785.376 ± 16.471 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 100 sample 2899401 817.392 ± 6.668 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 1000 sample 3650566 2077.700 ± 0.600 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 10000 sample 3005467 19949.334 ± 4.243 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 1 sample 456091 48.610 ± 1.162 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 10 sample 732051 62.599 ± 0.815 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 100 sample 778925 228.629 ± 1.205 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 1000 sample 633682 2002.987 ± 2.856 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 10000 sample 506442 19735.345 ± 12.312 ns/op
AFTER:
Benchmark (delay) Mode Cnt Score Error Units
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 1 sample 3761980 383.436 ± 1.315 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 10 sample 3667304 474.429 ± 1.101 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 100 sample 3039374 479.267 ± 0.435 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 1000 sample 3709210 2044.603 ± 0.989 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 10000 sample 3011591 19904.227 ± 18.025 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 1 sample 494975 52.269 ± 8.345 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 10 sample 771094 62.290 ± 0.795 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 100 sample 763230 235.044 ± 1.552 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 1000 sample 634037 2006.578 ± 3.574 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 10000 sample 506284 19742.605 ± 13.729 ns/op
```
Motiviation:
The ResourceLeakDetector helps to detect and troubleshoot resource leaks and is often used even in production enviroments with a low level. Because of this its import that we try to keep the overhead as low as overhead. Most of the times no leak is detected (as all is correctly handled) so we should keep the overhead for this case as low as possible.
Modifications:
- Only call getStackTrace() if a leak is reported as it is a very expensive native call. Also handle the filtering and creating of the String in a lazy fashion
- Remove the need to mantain a Queue to store the last access records
- Add benchmark
Result:
Huge decrease of performance overhead.
Before the patch:
Benchmark (recordTimes) Mode Cnt Score Error Units
ResourceLeakDetectorRecordBenchmark.record 8 thrpt 20 4358.367 ± 116.419 ops/s
ResourceLeakDetectorRecordBenchmark.record 16 thrpt 20 2306.027 ± 55.044 ops/s
ResourceLeakDetectorRecordBenchmark.recordWithHint 8 thrpt 20 4220.979 ± 114.046 ops/s
ResourceLeakDetectorRecordBenchmark.recordWithHint 16 thrpt 20 2250.734 ± 55.352 ops/s
With this patch:
Benchmark (recordTimes) Mode Cnt Score Error Units
ResourceLeakDetectorRecordBenchmark.record 8 thrpt 20 71398.957 ± 2695.925 ops/s
ResourceLeakDetectorRecordBenchmark.record 16 thrpt 20 38643.963 ± 1446.694 ops/s
ResourceLeakDetectorRecordBenchmark.recordWithHint 8 thrpt 20 71677.882 ± 2923.622 ops/s
ResourceLeakDetectorRecordBenchmark.recordWithHint 16 thrpt 20 38660.176 ± 1467.732 ops/s
Motivation:
1. Some encoders used a `ByteBuf#writeBytes` to write short constant byte array (2-3 bytes). This can be replaced with more faster `ByteBuf#writeShort` or `ByteBuf#writeMedium` which do not access the memory.
2. Two chained calls of the `ByteBuf#setByte` with constants can be replaced with one `ByteBuf#setShort` to reduce index checks.
3. The signature of method `HttpHeadersEncoder#encoderHeader` has an unnecessary `throws`.
Modifications:
1. Use `ByteBuf#writeShort` or `ByteBuf#writeMedium` instead of `ByteBuf#writeBytes` for the constants.
2. Use `ByteBuf#setShort` instead of chained call of the `ByteBuf#setByte` with constants.
3. Remove an unnecessary `throws` from `HttpHeadersEncoder#encoderHeader`.
Result:
A bit faster writes constants into buffers.
Motivation:
Right now HttpRequestEncoder does insertion of slash for url like http://localhost?pararm=1 before the question mark. It is done not effectively.
Modification:
Code:
new StringBuilder(len + 1)
.append(uri, 0, index)
.append(SLASH)
.append(uri, index, len)
.toString();
Replaced with:
new StringBuilder(uri)
.insert(index, SLASH)
.toString();
Result:
Faster HttpRequestEncoder. Additional small test. Attached benchmark in PR.
Benchmark Mode Cnt Score Error Units
HttpRequestEncoderInsertBenchmark.newEncoder thrpt 40 3704843.303 ± 98950.919 ops/s
HttpRequestEncoderInsertBenchmark.oldEncoder thrpt 40 3284236.960 ± 134433.217 ops/s
Motivation:
A `StringUtil#escapeCsv` creates new `StringBuilder` on each value even if the same string is returned in the end.
Modifications:
Create new `StringBuilder` only if it really needed. Otherwise, return the original string (or just trimmed substring).
Result:
Less GC load. Up to 4x faster work for not changed strings.
Motivation:
When I run Netty micro benchmarks I get many warnings like:
WARNING: -Dio.netty.noResourceLeakDetection is deprecated. Use '-Dio.netty.leakDetection.level=simple' instead.
Modification:
-Dio.netty.noResourceLeakDetection replaced with -Dio.netty.leakDetection.level=disabled.
Result:
No warnings.
Motivation:
IPv4/6 validation methods use allocations, which can be avoided.
IPv4 parse method use StringTokenizer.
Modifications:
Rewriting IPv4/6 validation methods to avoid allocations.
Rewriting IPv4 parse method without use StringTokenizer.
Result:
IPv4/6 validation and IPv4 parsing faster up to 2-10x.
Motivation:
We currently don't have a native transport which supports kqueue https://www.freebsd.org/cgi/man.cgi?query=kqueue&sektion=2. This can be useful for BSD systems such as MacOS to take advantage of native features, and provide feature parity with the Linux native transport.
Modifications:
- Make a new transport-native-unix-common module with all the java classes and JNI code for generic unix items. This module will build a static library for each unix platform, and included in the dynamic libraries used for JNI (e.g. transport-native-epoll, and eventually kqueue).
- Make a new transport-native-unix-common-tests module where the tests for the transport-native-unix-common module will live. This is so each unix platform can inherit from these test and ensure they pass.
- Add a new transport-native-kqueue module which uses JNI to directly interact with kqueue
Result:
JNI support for kqueue.
Fixes https://github.com/netty/netty/issues/2448
Fixes https://github.com/netty/netty/issues/4231
Motivation:
Java9 added a new method to Unsafe which allows to allocate a byte[] without memset it. This can have a massive impact in allocation times when the byte[] is big. This change allows to enable this when using Java9 with the io.netty.tryAllocateUninitializedArray property when running Java9+. Please note that you will need to open up the jdk.internal.misc package via '--add-opens java.base/jdk.internal.misc=ALL-UNNAMED' as well.
Modifications:
Allow to allocate byte[] without memset on Java9+
Result:
Better performance when allocate big heap buffers and using java9.
Motivation:
There are two files that still use `system.out.println` to log their status
Modification:
Replace `system.out.println` with a `debug` function inside an instance of `InternalLoggerFactory`
Result:
Introduce an instance of `InternalLoggerFactory` in class `AbstractMicrobenchmark.java` and `AbstractSharedExecutorMicrobenchmark.java`
Motivation:
We currently don't have a benchmark which includes SslHandler. The SslEngine benchmarks also always include a single TLS packet when encoding/decoding. In practice when reading data from the network there may be multiple TLS packets present and we should expand the benchmarks to understand this use case.
Modifications:
- SslEngine benchmarks should include wrapping/unwrapping of multiple TLS packets
- Introduce SslHandler benchmarks which can also account for wrapping/unwrapping of multiple TLS packets
Result:
SslHandler and SslEngine benchmarks are more comprehensive.
Motivation:
The internal.hpack classes are no longer exposed in our public APIs and can be made package private in the http2 package.
Modifications:
- Make the hpack classes package private in the http2 package
Result:
Less APIs exposed as public.
Motivation:
Often its useful for the user to be able to get some stats about the memory allocated via an allocator.
Modifications:
- Allow to obtain the used heap and direct memory for an allocator
- Add test case
Result:
Fixes [#6341]
Motivation:
Issue [#6349] brought up the idea to not use UnpooledUnsafeNoCleanerDirectByteBuf by default. To decide what to do a benchmark is needed.
Modifications:
Add benchmarks for UnpooledUnsafeNoCleanerDirectByteBuf vs UnpooledUnsafeDirectB
yteBuf
Result:
Better idea about impact of using UnpooledUnsafeNoCleanerDirectByteBuf.
Motivation:
As we provide our own SSLEngine implementation we should have benchmarks to compare it against JDK impl.
Modifications:
Add benchmarks for wrap / unwrap and handshake performance.
Result:
Benchmarks FTW.
Motivation:
Allmost all our benchmarks are in src/main/java but a few are in src/test/java. We should make it consistent.
Modifications:
Move everything to src/main/java
Result:
Consistent code base.
Motivation:
64-byte alignment is recommended by the Intel performance guide (https://software.intel.com/en-us/articles/practical-intel-avx-optimization-on-2nd-generation-intel-core-processors) for data-structures over 64 bytes.
Requiring padding to a multiple of 64 bytes allows for using SIMD instructions consistently in loops without additional conditional checks. This should allow for simpler and more efficient code.
Modification:
At the moment cache alignment must be setup manually. But probably it might be taken from the system. The original code was introduced by @normanmaurer https://github.com/netty/netty/pull/4726/files
Result:
Buffer alignment works better than miss-align cache.
Motivation:
codec-http2 couples the dependency tree state with the remainder of the stream state (Http2Stream). This makes implementing constraints where stream state and dependency tree state diverge in the RFC challenging. For example the RFC recommends retaining dependency tree state after a stream transitions to closed [1]. Dependency tree state can be exchanged on streams in IDLE. In practice clients may use stream IDs for the purpose of establishing QoS classes and therefore retaining this dependency tree state can be important to client perceived performance. It is difficult to limit the total amount of state we retain when stream state and dependency tree state is combined.
Modifications:
- Remove dependency tree, priority, and weight related items from public facing Http2Connection and Http2Stream APIs. This information is optional to track and depends on the flow controller implementation.
- Move all dependency tree, priority, and weight related code from DefaultHttp2Connection to WeightedFairQueueByteDistributor. This is currently the only place which cares about priority. We can pull out the dependency tree related code in the future if it is generally useful to expose for other implementations.
- DefaultHttp2Connection should explicitly limit the number of reserved streams now that IDLE streams are no longer created.
Result:
More compliant with the HTTP/2 RFC.
Fixes https://github.com/netty/netty/issues/6206.
[1] https://tools.ietf.org/html/rfc7540#section-5.3.4
