测试环境:4 x Xeon E5520 2.27GHz, 7.3GB 虚拟机
测试结果:
一、采用java内置序列化：
| tps(w) | thread | cpu(%) | load |
| 8.9 | 1 | 22.3 | 0.8 |
| 15.5 | 2 | 46.3 | 1.64 |
| 21 | 3 | 69.7 | 2.88 |
| 23 | 4 | 88.5 | 3.5 |
| 20.5 | 5 | 70 | 4.2 |
| 19.5 | 6 | 71 | 5 |
| 20 | 7 | 67 | 6.28 |
| 20.5 | 8 | 64 | 6.5 |
| 20.5 | 40 | 70 | 7.5 |
JAVA内置反序列化：
| tpc(w) | thread | cpu(%) | load |
| 2 | 1 | 24.6 | 1.1 |
| 3.85 | 2 | 49 | 1.9 |
| 5.8 | 3 | 72 | 2.8 |
| 7.65 | 4 | 96.6 | 4 |
| 7.6 | 5 | 96.8 | 4.5 |
| 7.6 | 6 | 96.9 | 6.3 |
| 7.6 | 7 | 97.1 | 7.2 |
| 7.6 | 8 | 98 | 9 |
| 7.6 | 40 | 97 | 14.5 |
二 protostuff序列化
| tpc(w) | thread | cpu(%) | load |
| 54 | 1 | 9.8 | 1 |
| 105 | 2 | 22 | 0.8 |
| 150 | 3 | 41 | 1.8 |
| 190 | 4 | 52 | 2.44 |
| 180 | 5 | 66 | 3.6 |
| 200 | 6 | 76 | 5 |
| 240 | 7 | 85.7 | 5.1 |
| 250 | 8 | 94.3 | 5.3 |
| 270 | 40 | 99 | 17 |
protostuff反序列化
| tpc(w) | thread | cpu(%) | load |
| 38 | 1 | 11 | 0.9 |
| 71 | 2 | 28.7 | 1.6 |
| 97 | 3 | 49 | 1.9 |
| 142 | 4 | 57.7 | 2.5 |
| 150 | 5 | 65 | 3.7 |
| 160 | 6 | 73 | 4.6 |
| 170 | 7 | 81 | 6.8 |
| 177 | 8 | 89 | 6.5 |
| 175 | 40 | 93 | 10.5 |
小结论：
1、线程数接近CPU核数能达到最好性能
2、java/protostuff序列化比反序列化性能都要好2倍左右
3、java序列化性能不理想，cpu跑满，也就10w左右，protostuff跟它完全不是一个量级，可以达到200w以上
相关推荐：
https://github.com/eishay/jvm-serializers/wiki （各种序列化对比测试，生成报表）
本文来源于"阿里中间件团队播客",原文发表时间" 2012-05-01"