0x1 摘要

平时开发中经常听到序列与反序列化，特别是在分布式系统与RPC应用中，今天突然心血来潮对几种常用的序列化框架做个性能测试对比，测试对象：

• Java 原生序列
• Avro
• Thrift
• Protobuf

0x2 测试环境及工具

测试环境：
系统类型：64 位操作系统
CPU：Intel(R) Core(TM) i3-4130 CPU @ 3.40 GHz 3.40 GHz
内存：8 GB
开发工具：IDEA

测试工具：
JMH

0x3 测试实体

public class User implements Serializable{
    private static final long serialVersionUID = 5149128310592716591L;

    private int age;
    private String username;
    private String address;

    public int getAge() {
        return age;
    }

    public User setAge(int age) {
        this.age = age;
        return this;
    }

    public String getUsername() {
        return username;
    }

    public User setUsername(String username) {
        this.username = username;
        return this;
    }

    public String getAddress() {
        return address;
    }

    public User setAddress(String address) {
        this.address = address;
        return this;
    }
}

0x4 JMH参数设置

@BenchmarkMode(Mode.Throughput)
@Warmup(iterations = 3)
@Measurement(iterations = 10, time = 5, timeUnit = TimeUnit.SECONDS)
@Threads(4)
@OutputTimeUnit(TimeUnit.MILLISECONDS)

参数意义不在此文做介绍，大家可以自行网上搜索或者看官网。

0x5 测试结果

Encode：

Benchmark	Mode	Cnt	Score	Erorr	Units
SerializableEncodeTest.testAvroSerializableEncode	thrpt	20	2881.398 ±	27.619	ops/ms
SerializableEncodeTest.testJavaSerializableEncode	thrpt	20	1708.157 ±	122.516	ops/ms
SerializableEncodeTest.testProtobufSerializableEncode	thrpt	20	10349.962 ±	215.994	ops/ms
SerializableEncodeTest.testThriftSerializableEncode	thrpt	20	5292.191 ±	49.890	ops/ms

Decode：

Benchmark	Mode	Cnt	Score	Erorr	Units
SerializableEncodeTest.testAvroSerializableEncode	thrpt	20	931.167 ±	13.185	ops/ms
SerializableEncodeTest.testJavaSerializableEncode	thrpt	20	521.145 ±	6.993	ops/ms
SerializableEncodeTest.testProtobufSerializableEncode	thrpt	20	25335.295 ±	564.496	ops/ms
SerializableEncodeTest.testThriftSerializableEncode	thrpt	20	5239.726 ±	75.048	ops/ms

0x6 结论

从结果中不难看出，Protobuf不管是序列化还是反序列化都有绝对优势，Java原生都是最弱的。