FlatBuffers与protobuf性能比較

FlatBuffers发布时。顺便也发布了它的性能数据，详细数据请见Benchmark。

它的測试用例由下面数据构成"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."

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

測试用例为：序列化一个通讯录personal_info_list（table），通讯录能够觉得是有每一个人的信息(personal_info)的集合。每一个人信息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)，可是struct不同意有数组或string成员，无奈我用table描写叙述它了。对应的idl文件例如以下：

//////////////////////////////////////////////////////
//// FILE     : tellist.fbs
//// DESC     : basic message for msg-center
//// AUTHOR   : v 0.1 written by Alex Stocks on June 22, 2014
//// LICENCE  :
//// MOD      :
////////////////////////////////////////////////////////

namespace as.tellist;

enum GENDER_TYPE : byte
{
	MALE	= 0,
	FEMALE	= 1,
	OTHER	= 2
}

table personal_info
{
	id : uint;
	name : string;
	age : byte;
	gender : GENDER_TYPE;
	phone_num : ulong;
}

table personal_info_list
{
	info : [personal_info];
}

root_type personal_info_list;

由于要以protobuf做性能參考。列出protobuf的idl文件例如以下：

//////////////////////////////////////////////////////
//// FILE     : tellist.proto
//// DESC     : basic message for msg-center
//// AUTHOR   : v 0.1 written by Alex Stocks on June 22, 2014
//// LICENCE  :
//// MOD      :
////////////////////////////////////////////////////////

package as.tellist;

enum gender_type
{
	MALE	= 0;
	FEMALE	= 1;
	OTHER	= 2;
}

message personal_info
{
	optional uint32			id = 1;
	optional string			name = 2;
	optional uint32			age = 3;
	optional gender_type	gender = 4;
	optional uint64			phone_num = 5;
}

message personal_info_list
{
	repeated personal_info	info = 1;
}

若用C的struct描写叙述相应的头文件(其相应的程序称之为“二进制”)。例如以下：

/**
 * FILE		: tellist.h
 * DESC		: to test tellist
 * AUTHOR	: v1.0 written by Alex Stocks
 * DATE		: on June 28, 2014
 * LICENCE	: GPL 2.0
 * MOD		:
 **/

#ifndef __TELLIST_H__
#define __TELLIST_H__

enum
{
 GENDER_TYPE_MALE = 0,
 GENDER_TYPE_FEMALE = 1,
 GENDER_TYPE_OTHER = 2,
};

inline const char **EnumNamesGENDER_TYPE()
{
 static const char *names[] = { "MALE", "FEMALE", "OTHER"};
 return names;
}

inline const char *EnumNameGENDER_TYPE(int e)
{
 return EnumNamesGENDER_TYPE()[e];
}

typedef struct personal_info_tag
{
 unsigned			id;
 unsigned char		age;
 char				gender;
 unsigned long long	phone_num;
 char				name[32];
} personal_info;

typedef struct personal_info_list_tag
{
 int				size;
 personal_info	info[0];
} personal_info_list;

#endif

// the end of the header file tellist.h

測试时，在内存中构造37个personal_info对象。并序列化之，反复这个过程100万次。然后再进行反序列化，再反复100万次。

測试结果例如以下(补充：tellist_pb是protobuf測试程序，tellist_fb是FlatBuffers測试程序，tellist_fb是二进制測试程序。)：

測试环境：12Core Intel(R) Xeon(R) CPU E5-2620 0 @ 2.00GHz
free
             total       used       free     shared    buffers     cached
Mem:      66081944   62222028    3859916          0     196448   43690828
-/+ buffers/cache:   18334752   47747192
Swap:       975864     855380     120484

protobuf三次測试结果：
bin/tellist_pb 
encode: loop = 1000000, time diff = 14210ms
decode: loop = 1000000, time diff = 11185ms
buf size:841

bin/tellist_pb 
encode: loop = 1000000, time diff = 14100ms
decode: loop = 1000000, time diff = 11234ms
buf size:841

bin/tellist_pb 
encode: loop = 1000000, time diff = 14145ms
decode: loop = 1000000, time diff = 11237ms
buf size:841
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

flatbuffers三次測试结果：
 bin/tellist_fb 
encode: loop = 1000000, time diff = 11666ms
decode: loop = 1000000, time diff = 1141ms
buf size:1712

bin/tellist_fb 
encode: loop = 1000000, time diff = 11539ms
decode: loop = 1000000, time diff = 1200ms
buf size:1712

bin/tellist_fb 
encode: loop = 1000000, time diff = 11737ms
decode: loop = 1000000, time diff = 1141ms
buf size:1712
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

二进制三次測试结果：
bin/tellist 
encode: loop = 1000000, time diff = 4967ms
decode: loop = 1000000, time diff = 688ms
buf size:304

 bin/tellist 
encode: loop = 1000000, time diff = 4971ms
decode: loop = 1000000, time diff = 687ms
buf size:304

bin/tellist 
encode: loop = 1000000, time diff = 4966ms
decode: loop = 1000000, time diff = 686ms
buf size:304
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

測试环境：1 Core Intel(R) Core(TM) i5-3210M CPU @ 2.50GHz
free
             total       used       free     shared    buffers     cached
Mem:        753932     356036     397896          0      50484     224848
-/+ buffers/cache:      80704     673228
Swap:      1324028        344    1323684
protobuf三次測试结果：
./bin/tellist_pb 
encode: loop = 1000000, time diff = 12451ms
decode: loop = 1000000, time diff = 9662ms
buf size:841

./bin/tellist_pb 
encode: loop = 1000000, time diff = 12545ms
decode: loop = 1000000, time diff = 9840ms
buf size:841

./bin/tellist_pb 
encode: loop = 1000000, time diff = 12554ms
decode: loop = 1000000, time diff = 10460ms
buf size:841
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

flatbuffers三次測试结果：
bin/tellist_fb 
encode: loop = 1000000, time diff = 9640ms
decode: loop = 1000000, time diff = 1164ms
buf size:1712

bin/tellist_fb 
encode: loop = 1000000, time diff = 9595ms
decode: loop = 1000000, time diff = 1170ms
buf size:1712

bin/tellist_fb 
encode: loop = 1000000, time diff = 9570ms
decode: loop = 1000000, time diff = 1172ms
buf size:1712
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

二进制三次測试结果：
bin/tellist 
encode: loop = 1000000, time diff = 4194ms
decode: loop = 1000000, time diff = 538ms
buf size:304

bin/tellist 
encode: loop = 1000000, time diff = 4387ms
decode: loop = 1000000, time diff = 544ms
buf size:304

bin/tellist 
encode: loop = 1000000, time diff = 4181ms
decode: loop = 1000000, time diff = 533ms
buf size:304
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不管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。但本文仅仅讨论FlatBuffers和protobuf。所以不让它的结果參与比較。

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較慢的原因）。

上面的測试环境是在公司的linux server端和我自己的mac pro分别进行的。请手机端开发人员自己也在手机端进行下測试， 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，假设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。只是相比使用protobuf时的一堆头文件和占18M之多的lib库，FlatBuffers只一个"flatbuffers/flatbuffers.h"就足够了。

測试程序已经上传到百度网盘，点击这个链接就可以下载。

欢迎各位的批评意见。