FlatBuffers与protobuf性能比较

浅浅的花香味﹌ 2022-07-21 02:11 160阅读 0赞

FlatBuffers发布时，顺便也公布了它的性能数据，具体数据请见[Benchmark][]。

它的测试用例由以下数据构成"a set of about 10 objects containing an array, 4 strings, and a large variety of int/float scalar values of all sizes, meant to be representative of game data, e.g. a scene format."

我感觉这样测试如同儿戏，便自己设计了一个测试用例，主要关注CPU计算时间和内存空间占用两个指标，参考对象是protobuf。

测试用例为：序列化一个通讯录personal\_info\_list（table），通讯录可以认为是有每个人的信息(personal\_info)的集合。每个人信息personal\_info(table)有：个人id(uint)、名字(string)、年龄(byte)、性别(enum, byte)和电话号码(ulong)。本来我想用struct表示personal\_info(table)，但是struct不允许有数组或string成员，无奈我用table描述它了。相应的idl文件如下：

**\[cpp\]** [view plain][] [copy][view plain]

1.  //
2.   FILE : tellist.fbs
3.   DESC : basic message for msg-center
4.   AUTHOR : v 0.1 written by Alex Stocks on June 22, 2014
5.   LICENCE :
6.   MOD :
7.  
8.  
9.  **namespace** as.tellist; 
10. 
11. **enum** GENDER\_TYPE : byte 
12. \{ 
13.  MALE = 0, 
14.  FEMALE = 1, 
15.  OTHER = 2 
16. \} 
17. 
18. table personal\_info 
19. \{ 
20.  id : uint; 
21.  name : string; 
22.  age : byte; 
23.  gender : GENDER\_TYPE; 
24.  phone\_num : ulong; 
25. \} 
26. 
27. table personal\_info\_list 
28. \{ 
29.  info : \[personal\_info\]; 
30. \} 
31. 
32. root\_type personal\_info\_list;

因为要以protobuf做性能参考，列出protobuf的idl文件如下：

**\[cpp\]** [view plain][] [copy][view plain]

1.  //
2.   FILE : tellist.proto
3.   DESC : basic message for msg-center
4.   AUTHOR : v 0.1 written by Alex Stocks on June 22, 2014
5.   LICENCE :
6.   MOD :
7.  
8.  
9.  package as.tellist; 
10. 
11. **enum** gender\_type 
12. \{ 
13.  MALE = 0; 
14.  FEMALE = 1; 
15.  OTHER = 2; 
16. \} 
17. 
18. message personal\_info 
19. \{ 
20.  optional uint32 id = 1; 
21.  optional string name = 2; 
22.  optional uint32 age = 3; 
23.  optional gender\_type gender = 4; 
24.  optional uint64 phone\_num = 5; 
25. \} 
26. 
27. message personal\_info\_list 
28. \{ 
29.  repeated personal\_info info = 1; 
30. \}

若用C的struct描述对应的头文件(其对应的程序称之为“二进制”)，如下：

**\[cpp\]** [view plain][] [copy][view plain]

1.  /\*\*
2.   \* FILE : tellist.h
3.   \* DESC : to test tellist
4.   \* AUTHOR : v1.0 written by Alex Stocks
5.   \* DATE : on June 28, 2014
6.   \* LICENCE : GPL 2.0
7.   \* MOD :
8.   \*\*/
9.  
10. \#ifndef \_\_TELLIST\_H\_\_
11. \#define \_\_TELLIST\_H\_\_
12. 
13. **enum**
14. \{ 
15.  GENDER\_TYPE\_MALE = 0, 
16.  GENDER\_TYPE\_FEMALE = 1, 
17.  GENDER\_TYPE\_OTHER = 2, 
18. \}; 
19. 
20. 
21. **inline****const****char** \*\*EnumNamesGENDER\_TYPE() 
22. \{ 
23. **static****const****char** \*names\[\] = \{ "MALE", "FEMALE", "OTHER"\}; 
24. **return** names; 
25. \} 
26. 
27. 
28. **inline****const****char** \*EnumNameGENDER\_TYPE(**int** e) 
29. \{ 
30. **return** EnumNamesGENDER\_TYPE()\[e\]; 
31. \} 
32. 
33. **typedef****struct** personal\_info\_tag 
34. \{ 
35.  unsigned id; 
36.  unsigned **char** age; 
37. **char** gender; 
38.  unsigned **long****long** phone\_num; 
39. **char** name\[32\]; 
40. \} personal\_info; 
41. 
42. **typedef****struct** personal\_info\_list\_tag 
43. \{ 
44. **int** size; 
45.  personal\_info info\[0\]; 
46. \} personal\_info\_list; 
47. 
48. \#endif
49. 
50. // the end of the header file tellist.h

测试时，在内存中构造37个personal\_info对象，并序列化之，重复这个过程100万次，然后再进行反序列化，再重复100万次。

测试结果如下(补充：tellist\_pb是protobuf测试程序，tellist\_fb是FlatBuffers测试程序，tellist\_fb是二进制测试程序，)：

**\[html\]** [view plain][] [copy][view plain]

1.  测试环境：12Core Intel(R) Xeon(R) CPU E5-2620 0 @ 2.00GHz 
2.  free 
3.   total used free shared buffers cached 
4.  Mem: 66081944 62222028 3859916 0 196448 43690828 
5.  \-/+ buffers/cache: 18334752 47747192 
6.  Swap: 975864 855380 120484 
7.  
8.  protobuf三次测试结果： 
9.  bin/tellist\_pb 
10. encode: loop = 1000000, time diff = 14210ms
11. decode: loop = 1000000, time diff = 11185ms
12. buf size:841 
13. 
14. bin/tellist\_pb 
15. encode: loop = 1000000, time diff = 14100ms
16. decode: loop = 1000000, time diff = 11234ms
17. buf size:841 
18. 
19. bin/tellist\_pb 
20. encode: loop = 1000000, time diff = 14145ms
21. decode: loop = 1000000, time diff = 11237ms
22. buf size:841 
23. 序列化后占用内存空间841Byte，encode平均运算时间42455ms / 3 = 14151.7ms，decode平均计算时间33656ms / 3 = 11218.7ms 
24. 
25. flatbuffers三次测试结果： 
26.  bin/tellist\_fb 
27. encode: loop = 1000000, time diff = 11666ms
28. decode: loop = 1000000, time diff = 1141ms
29. buf size:1712 
30. 
31. bin/tellist\_fb 
32. encode: loop = 1000000, time diff = 11539ms
33. decode: loop = 1000000, time diff = 1200ms
34. buf size:1712 
35. 
36. bin/tellist\_fb 
37. encode: loop = 1000000, time diff = 11737ms
38. decode: loop = 1000000, time diff = 1141ms
39. buf size:1712 
40. 序列化后占用内存空间1712Byte，encode平均运算时间34942ms / 3 = 11647.3ms，decode平均计算时间3482ms / 3 = 1160.7ms 
41. 
42. 二进制三次测试结果： 
43. bin/tellist 
44. encode: loop = 1000000, time diff = 4967ms
45. decode: loop = 1000000, time diff = 688ms
46. buf size:304 
47. 
48.  bin/tellist 
49. encode: loop = 1000000, time diff = 4971ms
50. decode: loop = 1000000, time diff = 687ms
51. buf size:304 
52. 
53. bin/tellist 
54. encode: loop = 1000000, time diff = 4966ms
55. decode: loop = 1000000, time diff = 686ms
56. buf size:304 
57. 序列化后占用内存空间304Byte，encode平均运算时间14904ms / 3 = 4968ms，decode平均计算时间2061ms / 3 = 687ms
58. 
59. 测试环境：1 Core Intel(R) Core(TM) i5-3210M CPU @ 2.50GHz 
60. free 
61.  total used free shared buffers cached 
62. Mem: 753932 356036 397896 0 50484 224848 
63. \-/+ buffers/cache: 80704 673228 
64. Swap: 1324028 344 1323684 
65. protobuf三次测试结果： 
66. ./bin/tellist\_pb 
67. encode: loop = 1000000, time diff = 12451ms
68. decode: loop = 1000000, time diff = 9662ms
69. buf size:841 
70. 
71. ./bin/tellist\_pb 
72. encode: loop = 1000000, time diff = 12545ms
73. decode: loop = 1000000, time diff = 9840ms
74. buf size:841 
75. 
76. ./bin/tellist\_pb 
77. encode: loop = 1000000, time diff = 12554ms
78. decode: loop = 1000000, time diff = 10460ms
79. buf size:841 
80. 序列化后占用内存空间841Byte，encode平均运算时间37550ms / 3 = 12516.7ms，decode平均计算时间29962ms / 3 = 9987.3ms 
81. 
82. flatbuffers三次测试结果： 
83. bin/tellist\_fb 
84. encode: loop = 1000000, time diff = 9640ms
85. decode: loop = 1000000, time diff = 1164ms
86. buf size:1712 
87. 
88. bin/tellist\_fb 
89. encode: loop = 1000000, time diff = 9595ms
90. decode: loop = 1000000, time diff = 1170ms
91. buf size:1712 
92. 
93. bin/tellist\_fb 
94. encode: loop = 1000000, time diff = 9570ms
95. decode: loop = 1000000, time diff = 1172ms
96. buf size:1712 
97. 序列化后占用内存空间1712Byte，encode平均运算时间28805ms / 3 = 9345ms，decode平均计算时间3506ms / 3 = 1168.7ms 
98. 
99. 二进制三次测试结果： 
100. bin/tellist 
101. encode: loop = 1000000, time diff = 4194ms
102. decode: loop = 1000000, time diff = 538ms
103. buf size:304 
104. 
105. bin/tellist 
106. encode: loop = 1000000, time diff = 4387ms
107. decode: loop = 1000000, time diff = 544ms
108. buf size:304 
109. 
110. bin/tellist 
111. encode: loop = 1000000, time diff = 4181ms
112. decode: loop = 1000000, time diff = 533ms
113. buf size:304 
114. 序列化后占用内存空间304Byte，encode平均运算时间12762ms / 3 = 4254ms，decode平均计算时间1615ms / 3 = 538.3ms

上面的二进制程序的结果无论在内存空间占用还是cpu计算时间这两个指标上都是最快的。但本文只讨论FlatBuffers和protobuf，所以不让它的结果参与比较。

从以上数据看出，在内存空间占用这个指标上，FlatBuffers占用的内存空间比protobuf多了两倍。序列化时二者的cpu计算时间FB比PB快了3000ms左右，反序列化时二者的cpu计算时间FB比PB快了9000ms左右。FB在计算时间上占优势，而PB则在内存空间上占优（相比FB，这也正是它计算时间比较慢的原因）。

上面的测试环境是在公司的linux server端和我自己的mac pro分别进行的。请手机端开发者自己也在手机端进行下测试， 应该能得到类似的结果。Google宣称FB适合游戏开发是有道理的，如果在乎计算时间我想它也适用于后台开发。

另外，FB大量使用了C++11的语法，其从idl生成的代码接口不如protubuf友好。不过相比使用protobuf时的一堆头文件和占18M之多的lib库，FlatBuffers仅仅一个"flatbuffers/flatbuffers.h"就足够了。

测试程序已经上传到[百度网盘][Link 1]，点击这个链接即可下载。欢迎各位的批评意见。

[Benchmark]: http://google.github.io/flatbuffers/md__benchmarks.html
[view plain]: http://blog.csdn.net/menggucaoyuan/article/details/34409433#
[Link 1]: http://pan.baidu.com/s/1dD7sC1b