Motivation:
https://github.com/netty/netty/pull/10267 introduced a change that reduced the fragmentation. Unfortunally it also introduced a regression when it comes to caching of normal allocations. This can have a negative performance impact depending on the allocation sizes.
Modifications:
- Fix algorithm to calculate the array size for normal allocation caches
- Correctly calculate indeox for normal caches
- Add unit test
Result:
Fixes https://github.com/netty/netty/issues/10805
Fix issue #10508 where PARANOID mode slow down about 1000 times compared to ADVANCED.
Also fix a rare issue when internal buffer was growing over a limit, it was partially discarded
using `discardReadBytes()` which causes bad changes within previously discovered HttpData.
Reasons were:
Too many `readByte()` method calls while other ways exist (such as keep in memory the last scan position when trying to find a delimiter or using `bytesBefore(firstByte)` instead of looping externally).
Changes done:
- major change on way buffer are parsed: instead of read byte per byte until found delimiter, try to find the delimiter using `bytesBefore()` and keep the last unfound position to skeep already parsed parts (algorithms are the same but implementation of scan are different)
- Change the condition to discard read bytes when refCnt is at most 1.
Observations using Async-Profiler:
==================================
1) Without optimizations, most of the time (more than 95%) is through `readByte()` method within `loadDataMultipartStandard` method.
2) With using `bytesBefore(byte)` instead of `readByte()` to find various delimiter, the `loadDataMultipartStandard` method is going down to 19 to 33% depending on the test used. the `readByte()` method or equivalent `getByte(pos)` method are going down to 15% (from 95%).
Times are confirming those profiling:
- With optimizations, in SIMPLE mode about 82% better, in ADVANCED mode about 79% better and in PARANOID mode about 99% better (most of the duplicate read accesses are removed or make internally through `bytesBefore(byte)` method)
A benchmark is added to show the behavior of the various cases (one big item, such as File upload, and many items) and various level of detection (Disabled, Simple, Advanced, Paranoid). This benchmark is intend to alert if new implementations make too many differences (such as the previous version where about PARANOID gives about 1000 times slower than other levels, while it is now about at most 10 times).
Extract of Benchmark run:
=========================
Run complete. Total time: 00:13:27
Benchmark Mode Cnt Score Error Units
HttpPostMultipartRequestDecoderBenchmark.multipartRequestDecoderBigAdvancedLevel thrpt 6 2,248 ± 0,198 ops/ms
HttpPostMultipartRequestDecoderBenchmark.multipartRequestDecoderBigDisabledLevel thrpt 6 2,067 ± 1,219 ops/ms
HttpPostMultipartRequestDecoderBenchmark.multipartRequestDecoderBigParanoidLevel thrpt 6 1,109 ± 0,038 ops/ms
HttpPostMultipartRequestDecoderBenchmark.multipartRequestDecoderBigSimpleLevel thrpt 6 2,326 ± 0,314 ops/ms
HttpPostMultipartRequestDecoderBenchmark.multipartRequestDecoderHighAdvancedLevel thrpt 6 1,444 ± 0,226 ops/ms
HttpPostMultipartRequestDecoderBenchmark.multipartRequestDecoderHighDisabledLevel thrpt 6 1,462 ± 0,642 ops/ms
HttpPostMultipartRequestDecoderBenchmark.multipartRequestDecoderHighParanoidLevel thrpt 6 0,159 ± 0,003 ops/ms
HttpPostMultipartRequestDecoderBenchmark.multipartRequestDecoderHighSimpleLevel thrpt 6 1,522 ± 0,049 ops/ms
Motivation:
https in xmlns URIs does not work and will let the maven release plugin fail:
```
[INFO] ------------------------------------------------------------------------
[INFO] BUILD FAILURE
[INFO] ------------------------------------------------------------------------
[INFO] Total time: 1.779 s
[INFO] Finished at: 2020-11-10T07:45:21Z
[INFO] ------------------------------------------------------------------------
[ERROR] Failed to execute goal org.apache.maven.plugins:maven-release-plugin:2.5.3:prepare (default-cli) on project netty-parent: Execution default-cli of goal org.apache.maven.plugins:maven-release-plugin:2.5.3:prepare failed: The namespace xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" could not be added as a namespace to "project": The namespace prefix "xsi" collides with an additional namespace declared by the element -> [Help 1]
[ERROR]
```
See also https://issues.apache.org/jira/browse/HBASE-24014.
Modifications:
Use http for xmlns
Result:
Be able to use maven release plugin
Motivation:
JUnit 5 is the new hotness. It's more expressive, extensible, and composable in many ways, and it's better able to run tests in parallel. But most importantly, it's able to directly run JUnit 4 tests.
This means we can update and start using JUnit 5 without touching any of our existing tests.
I'm also introducing a dependency on assertj-core, which is like hamcrest, but arguably has a nicer and more discoverable API.
Modification:
Add the JUnit 5 and assertj-core dependencies, without converting any tests at time time.
Result:
All our tests are now executed through the JUnit 5 Vintage Engine.
Also, the JUnit 5 test APIs are available, and any JUnit 5 tests that are added from now on will also be executed.
Motivation:
HTTP is a plaintext protocol which means that someone may be able
to eavesdrop the data. To prevent this, HTTPS should be used whenever
possible. However, maintaining using https:// in all URLs may be
difficult. The nohttp tool can help here. The tool scans all the files
in a repository and reports where http:// is used.
Modifications:
- Added nohttp (via checkstyle) into the build process.
- Suppressed findings for the websites
that don't support HTTPS or that are not reachable
Result:
- Prevent using HTTP in the future.
- Encourage users to use HTTPS when they follow the links they found in
the code.
Motivation:
DefaultAttributeMap::attr has a blocking behaviour on lookup of an existing attribute:
it can be made non-blocking.
Modification:
Replace the existing fixed bucket table using a locked intrusive linked list
with an hand-rolled copy-on-write ordered single array
Result:
Non blocking behaviour for the lookup happy path
Motivation:
writeUtf8 can suffer from inlining issues and/or megamorphic call-sites on the hot path due to ByteBuf hierarchy
Modifications:
Duplicate and specialize the code paths to reduce the need of polymorphic calls
Result:
Performance are more stable in user code
Reduce garbage on MQTT encoding
Motivation:
MQTT encoding and decoding is doing unnecessary object allocation in a number of places:
- MqttEncoder create many byte[] to encode Strings into UTF-8 bytes
- MqttProperties uses Integer keys instead of int
- Some enums valueOf create unnecessary arrays on the hot paths
- MqttDecoder was using unecessary Result<T>
Modification:
- ByteBufUtil::utf8Bytes and ByteBufUtil::reserveAndWriteUtf8 allows to perform the same operation GC-free
- MqttProperties uses a primitive key map
- Implemented GC free const table lookup/switch valueOf
- Use some bit-tricks to pack 2 ints into a single primitive long to store both result and numberOfBytesConsumed and use byte[].length to compute numberOfByteConsumed on fly. These changes allowed to save creating Result<T>.
Result:
Significantly less garbage produced in MQTT encoding/decoding
Motivation:
We have found out that ByteBufUtil.indexOf can be inefficient for substring search on
ByteBuf, both in terms of algorithm complexity (worst case O(needle.readableBytes *
haystack.readableBytes)), and in constant factor (esp. on Composite buffers).
With implementation of more performant search algorithms we have seen improvements on
the order of magnitude.
Modifications:
This change introduces three search algorithms:
1. Knuth Morris Pratt - classical textbook algorithm, a good default choice.
2. Bit mask based algorithm - stable performance on any input, but limited to maximum
search substring (the needle) length of 64 bytes.
3. Aho–Corasick - worse performance and higher memory consumption than [1] and [2], but
it supports multiple substring (the needles) search simultaneously, by inspecting every
byte of the haystack only once.
Each algorithm processes every byte of underlying buffer only once, they are implemented
as ByteProcessor.
Result:
Efficient search algorithms with linear time complexity available in Netty (I will share
benchmark results in a comment on a PR).
Motivation:
PoolChunk.usage() method has non-trivial computations. It is used currently in hot path methods invoked when an allocation and de-allocation are happened.
The idea is to replace usage() output comparison against percent thresholds by Chunk.freeBytes plain comparison against absolute thresholds. In such way the majority of computations from the threshold conditions are moved to init logic.
Modifications:
Replace PoolChunk.usage() conditions in PoolChunkList with equivalent conditions for PoolChunk.freeBytes()
Result:
Improve performance of allocation and de-allocation of ByteBuf from normal size cache pool
Motivation:
decodeHexNibble can be a lot faster using a lookup table
Modifications:
decodeHexNibble is made faster by using a lookup table
Result:
decodeHexNibble is faster
Motivation:
Currently, characters are appended to the encoded string char-by-char even when no encoding is needed. We can instead separate out codepath that appends the entire string in one go for better `StringBuilder` allocation performance.
Modification:
Only go into char-by-char loop when finding a character that requires encoding.
Result:
The results aren't so clear with noise on my hot laptop - the biggest impact is on long strings, both to reduce resizes of the buffer and also to reduce complexity of the loop. I don't think there's a significant downside though for the cases that hit the slow path.
After
```
Benchmark Mode Cnt Score Error Units
QueryStringEncoderBenchmark.longAscii thrpt 6 1.406 ± 0.069 ops/us
QueryStringEncoderBenchmark.longAsciiFirst thrpt 6 0.046 ± 0.001 ops/us
QueryStringEncoderBenchmark.longUtf8 thrpt 6 0.046 ± 0.001 ops/us
QueryStringEncoderBenchmark.shortAscii thrpt 6 15.781 ± 0.949 ops/us
QueryStringEncoderBenchmark.shortAsciiFirst thrpt 6 3.171 ± 0.232 ops/us
QueryStringEncoderBenchmark.shortUtf8 thrpt 6 3.900 ± 0.667 ops/us
```
Before
```
Benchmark Mode Cnt Score Error Units
QueryStringEncoderBenchmark.longAscii thrpt 6 0.444 ± 0.072 ops/us
QueryStringEncoderBenchmark.longAsciiFirst thrpt 6 0.043 ± 0.002 ops/us
QueryStringEncoderBenchmark.longUtf8 thrpt 6 0.047 ± 0.001 ops/us
QueryStringEncoderBenchmark.shortAscii thrpt 6 16.503 ± 1.015 ops/us
QueryStringEncoderBenchmark.shortAsciiFirst thrpt 6 3.316 ± 0.154 ops/us
QueryStringEncoderBenchmark.shortUtf8 thrpt 6 3.776 ± 0.956 ops/us
```
Motivation:
In Java, it is almost always at least slower to use `ByteBuffer` than `byte[]` without pooling or I/O. `QueryStringDecoder` can use `byte[]` with arguably simpler code.
Modification:
Replace `ByteBuffer` / `CharsetDecoder` with `byte[]` and `new String`
Result:
After
```
Benchmark Mode Cnt Score Error Units
QueryStringDecoderBenchmark.noDecoding thrpt 6 5.612 ± 2.639 ops/us
QueryStringDecoderBenchmark.onlyDecoding thrpt 6 1.393 ± 0.067 ops/us
QueryStringDecoderBenchmark.mixedDecoding thrpt 6 1.223 ± 0.048 ops/us
```
Before
```
Benchmark Mode Cnt Score Error Units
QueryStringDecoderBenchmark.noDecoding thrpt 6 6.123 ± 0.250 ops/us
QueryStringDecoderBenchmark.onlyDecoding thrpt 6 0.922 ± 0.159 ops/us
QueryStringDecoderBenchmark.mixedDecoding thrpt 6 1.032 ± 0.178 ops/us
```
I notice #6781 switched from an array to `ByteBuffer` but I can't find any motivation for that in the PR. Unit tests pass fine with an array and we get a reasonable speed bump.
Motivation
JMH 1.22 was released recently, we might as well use the latest when
running benchmarks.
Summary of changes:
https://mail.openjdk.java.net/pipermail/jmh-dev/2019-November/002879.html
Modifications
Update jmh dependencies in microbench module from version 1.21 to 1.22.
Result
Benchmarks run using latest JMH
Motivation
Currently when future tasks are scheduled via EventExecutors from a
different thread, at least two allocations are performed - the
ScheduledFutureTask wrapping the to-be-run task, and a Runnable wrapping
the action to add to the scheduled task priority queue. The latter can
be avoided by incorporating this logic into the former.
Modification
- When scheduling or cancelling a future task from outside the event
loop, enqueue the task itself rather than wrapping in a Runnable
- Have ScheduledFutureTask#run first verify the task's deadline has
passed and if not add or remove it from the scheduledTaskQueue depending
on its cancellation state
- Add new outside-event-loop benchmarks to ScheduleFutureTaskBenchmark
Result
Fewer allocations when scheduling/cancelling future tasks
Motivation
Currently a static AtomicLong is used to allocate a unique id whenever a
task is scheduled to any event loop. This could be a source of
contention if delayed tasks are scheduled at a high frequency and can be
easily avoided by having a non-volatile id counter per queue.
Modifications
- Replace static AtomicLong ScheduledFutureTask#nextTaskId with a long
field in AbstractScheduledExecutorService
- Set ScheduledFutureTask#id based on this when adding the task to the
queue (in event loop) instead of at construction time
- Add simple benchmark
Result
Less contention / cache-miss possibility when scheduling future tasks
Before:
Benchmark (num) Mode Cnt Score Error Units
scheduleLots 100000 thrpt 20 346.008 ± 21.931 ops/s
Benchmark (num) Mode Cnt Score Error Units
scheduleLots 100000 thrpt 20 654.824 ± 22.064 ops/s
