[BUG] 对于JavaBean的序列化速度比Fastjson1要慢

[xinhochen]

问题描述

对于JavaBean的序列化速度比Fastjson1要慢，同时在去掉JSONWriter.Feature.ReferenceDetection之后，会出现OutOfMemoryError。于之对应的，Fastjson1只是偶尔出现OutOfMemoryError。

环境信息

请填写以下信息：

• OS信息： Win11 6Core 2.90GHz 16GB
• JDK信息： Temurin - 11.0.14.1
• 版本信息：Fastjson 2.0.10 & Fastjson 1.2.80
• JVM参数：-Xms128m -Xmx128m

重现步骤

https://gist.github.com/xinhochen/36f1d0cce97ae765aa9fa127833f2166

期待的正确结果

Fastjson2的序列化吞吐量和耗时优于Fastjson1

相关日志输出

Benchmark	Mode	Cnt	Score	Error	Units
BenchmarkJSON.fastJSON1ArrayDeThroughput	thrpt	5	16.857	± 0.118	ops/s
BenchmarkJSON.fastJSON2ArrayDeThroughput	thrpt	5	29.403	± 0.172	ops/s
BenchmarkJSON.fastJSON1ArraySeThroughput	thrpt	5	16.060	± 0.518	ops/s
BenchmarkJSON.fastJSON2ArraySeThroughput	thrpt	5	OOM		ops/s
BenchmarkJSON.fastJSON1ObjDeThroughput	thrpt	5	50.981	± 1.420	ops/s
BenchmarkJSON.fastJSON2ObjDeThroughput	thrpt	5	53.947	± 0.178	ops/s
BenchmarkJSON.fastJSON1ObjSeThroughput	thrpt	5	49.275	± 0.164	ops/s
BenchmarkJSON.fastJSON2ObjSeThroughput	thrpt	5	17.894	± 0.070	ops/s
BenchmarkJSON.fastJSON1ArrayDeTime	avgt	5	55.277	± 0.885	ms/op
BenchmarkJSON.fastJSON2ArrayDeTime	avgt	5	36.688	± 0.687	ms/op
BenchmarkJSON.fastJSON1ArraySeTime	avgt	5	61.726	± 2.049	ms/op
BenchmarkJSON.fastJSON2ArraySeTime	avgt	5	69.249	± 3.040	ms/op
BenchmarkJSON.fastJSON1ObjDeTime	avgt	5	19.844	± 0.372	ms/op
BenchmarkJSON.fastJSON2ObjDeTime	avgt	5	18.643	± 0.042	ms/op
BenchmarkJSON.fastJSON1ObjSeTime	avgt	5	22.432	± 0.037	ms/op
BenchmarkJSON.fastJSON2ObjSeTime	avgt	5	56.034	± 0.181	ms/op

附加信息

无

[wenshao]

https://oss.sonatype.org/content/repositories/snapshots/com/alibaba/fastjson2/fastjson2/2.0.11-SNAPSHOT/
问题已经修复，请用2.0.11-SNAPSHOT版本验证，2.0.11版本预计在8月7日发布

Benchmark                             Mode  Cnt    Score    Error  Units
Issue609.fastJSON1ArrayDeThroughput  thrpt    5   39.779 ±  5.517  ops/s
Issue609.fastJSON2ArrayDeThroughput  thrpt    5   68.184 ±  5.387  ops/s

Issue609.fastJSON1ArraySeThroughput  thrpt    5   42.854 ±  1.831  ops/s
Issue609.fastJSON2ArraySeThroughput  thrpt    5   56.745 ±  0.564  ops/s


Issue609.fastJSON1ObjDeThroughput    thrpt    5   68.175 ±  0.566  ops/s
Issue609.fastJSON2ObjDeThroughput    thrpt    5  100.525 ±  1.874  ops/s

Issue609.fastJSON1ObjSeThroughput    thrpt    5   62.280 ±  0.869  ops/s
Issue609.fastJSON2ObjSeThroughput    thrpt    5   92.590 ±  2.127  ops/s

Issue609.fastJSON1ArrayDeTime         avgt    5   29.628 ± 12.199  ms/op
Issue609.fastJSON2ArrayDeTime         avgt    5   15.280 ±  0.636  ms/op

Issue609.fastJSON2ArraySeTime         avgt    5   17.026 ±  0.227  ms/op
Issue609.fastJSON1ArraySeTime         avgt    5   23.259 ±  0.866  ms/op

Issue609.fastJSON1ObjDeTime           avgt    5   15.204 ±  0.225  ms/op
Issue609.fastJSON2ObjDeTime           avgt    5    9.907 ±  0.066  ms/op

Issue609.fastJSON1ObjSeTime           avgt    5   16.105 ±  0.277  ms/op
Issue609.fastJSON2ObjSeTime           avgt    5   10.677 ±  0.117  ms/op

[xinhochen]

刚刚用2.0.11-SNAPSHOT又跑了两次，还是fastjson1快一些
第一次是fastjson1出了OOM：

Benchmark                                  Mode  Cnt   Score    Error  Units
BenchmarkJSON.fastJSON1ArrayDeThroughput  thrpt    5  16.178 ±  0.091  ops/s
BenchmarkJSON.fastJSON2ArrayDeThroughput  thrpt    5  26.502 ±  0.181  ops/s

BenchmarkJSON.fastJSON1ArraySeThroughput  thrpt    5  12.312 ±  0.868  ops/s
BenchmarkJSON.fastJSON2ArraySeThroughput  thrpt    5  13.073 ±  0.464  ops/s

BenchmarkJSON.fastJSON1ObjDeThroughput    thrpt    5  38.096 ±  0.162  ops/s
BenchmarkJSON.fastJSON2ObjDeThroughput    thrpt    5  48.042 ±  0.140  ops/s

BenchmarkJSON.fastJSON1ObjSeThroughput    thrpt    5  28.130 ±  0.158  ops/s
BenchmarkJSON.fastJSON2ObjSeThroughput    thrpt    5  17.655 ±  0.046  ops/s <= 吞吐量小于fastjson1

BenchmarkJSON.fastJSON1ArrayDeTime         avgt    5  61.460 ±  0.222  ms/op
BenchmarkJSON.fastJSON2ArrayDeTime         avgt    5  36.497 ±  0.087  ms/op

BenchmarkJSON.fastJSON1ObjDeTime           avgt    5  26.347 ±  0.083  ms/op
BenchmarkJSON.fastJSON2ObjDeTime           avgt    5  21.208 ±  0.083  ms/op

BenchmarkJSON.fastJSON1ObjSeTime           avgt    5  37.991 ±  0.623  ms/op
BenchmarkJSON.fastJSON2ObjSeTime           avgt    5  56.020 ±  0.151  ms/op <= 耗时大于fastjson1

BenchmarkJSON.fastJSON2ArraySeTime         avgt    5  70.058 ± 11.770  ms/op

第二次：

Benchmark                                  Mode  Cnt   Score   Error  Units
BenchmarkJSON.fastJSON1ArrayDeThroughput  thrpt    5  15.587 ± 0.069  ops/s
BenchmarkJSON.fastJSON2ArrayDeThroughput  thrpt    5  26.932 ± 0.277  ops/s

BenchmarkJSON.fastJSON1ArraySeThroughput  thrpt    5  12.038 ± 1.162  ops/s
BenchmarkJSON.fastJSON2ArraySeThroughput  thrpt    5  14.216 ± 0.947  ops/s

BenchmarkJSON.fastJSON1ObjDeThroughput    thrpt    5  38.550 ± 0.139  ops/s
BenchmarkJSON.fastJSON2ObjDeThroughput    thrpt    5  48.215 ± 0.256  ops/s

BenchmarkJSON.fastJSON1ObjSeThroughput    thrpt    5  27.953 ± 0.060  ops/s
BenchmarkJSON.fastJSON2ObjSeThroughput    thrpt    5  17.526 ± 0.054  ops/s <= 吞吐量小于fastjson1

BenchmarkJSON.fastJSON1ArrayDeTime         avgt    5  64.037 ± 0.805  ms/op
BenchmarkJSON.fastJSON2ArrayDeTime         avgt    5  38.977 ± 0.184  ms/op

BenchmarkJSON.fastJSON1ArraySeTime         avgt    5  79.019 ± 6.010  ms/op
BenchmarkJSON.fastJSON2ArraySeTime         avgt    5  71.198 ± 7.717  ms/op

BenchmarkJSON.fastJSON1ObjDeTime           avgt    5  26.214 ± 0.132  ms/op
BenchmarkJSON.fastJSON2ObjDeTime           avgt    5  21.445 ± 0.118  ms/op

BenchmarkJSON.fastJSON1ObjSeTime           avgt    5  34.891 ± 0.135  ms/op
BenchmarkJSON.fastJSON2ObjSeTime           avgt    5  58.859 ± 0.091  ms/op <= 耗时大于fastjson1

[wenshao]

能提供下你跑benchmark的环境信息么？比如硬件、内存、JDK版本、JVM参数配置

[xinhochen]

已经在最开始提供了

[xinhochen]

我debug了序列化的逻辑，发现最终执行的是ObjectWriterAdpter的write方法

        jsonWriter.startObject();

        if (((features | this.features) & JSONWriter.Feature.WriteClassName.mask) != 0 || jsonWriter.isWriteTypeInfo(object, features)) {
            writeTypeInfo(jsonWriter);
        }

        for (int i = 0, size = fieldWriters.size(); i < size; ++i) {
            FieldWriter fieldWriter = fieldWriters.get(i);
            fieldWriter.write(jsonWriter, object);
        }

        jsonWriter.endObject();

其中的FieldWriter的write方法在本次benchmark中对应FieldWriterStringMethod和FieldWriterInt32，对应的write方法要考虑UTF8，我猜测可能这是慢的原因之一。
另外，FieldWriterInt32的writeInt32方法中，我发现两点：

1. utf8ValueCache是一个成员变量，对于每一个对象第一次序列化int32值时都要初始化。这里是否可以修改为静态变量，减少内存占用和重复初始化的开销
2. 对于nameWithColonUTF8，有一个Arrays.copyOf，做了一个深拷贝存储在original中。我觉得这个操作是多余的，因为后续又把这个orignal通过Arrays.copyOf深拷贝到bytes中了。后续的操作都是针对bytes的，这里的orginal我实在是找不到它存在的原因。

[wenshao]

https://oss.sonatype.org/content/repositories/snapshots/com/alibaba/fastjson2/fastjson2/2.0.12-SNAPSHOT/
2.0.12-SNAPSHOT版本修复了问题，序列化之前没有针对中文场景做优化，测试的都是ascii场景。

[xinhochen]

👍

[wenshao]

https://github.com/alibaba/fastjson2/releases/tag/2.0.12
问题已修复，请用新版本