Python内置marshal, cPickle等序列化库，但cPickle效率不太理想，marshal文档也说不保证版本兼容性。今天在列表中看到几个第三方库，故自己测试下：

测试脚本：

#!/usr/bin/env python

import sys,os,time
import cPickle
import marshal
import shelve
import tnetstring
import msgpack

def get_dict():
    d = {}
    for i in xrange(500000):
        d[i] = "abcd" * 10

    return d

import time, functools

def timeit(func):
    @functools.wraps(func)
    def __do__(*args,**kwargs):
        start = time.time()
        result= func(*args,**kwargs)
        print '%s used time %ss'%(func.__name__,time.time()-start)
        return result
    return __do__

@timeit
def marshal_dump(d):
    return marshal.dumps(d)

@timeit
def marshal_load(s):
    return marshal.loads(s)

def marshal_test(d):
    s = marshal_dump(d)
    marshal_load(s)

@timeit
def cPickle_dump(d):
    return cPickle.dumps(d, cPickle.HIGHEST_PROTOCOL)
@timeit
def cPickle_load(s):
    return cPickle.loads(s)

def cPickle_test(d):
    s = cPickle_dump(d)
    cPickle_load(s)

@timeit
def tnetstring_dump(d):
    return tnetstring.dumps(d)

@timeit
def tnetstring_load(s):
    return tnetstring.loads(s)

def tnetstring_test(d):
    s = tnetstring_dump(d)
    tnetstring_load(s)

@timeit
def msgpack_dump(d):
    return msgpack.packb(d)

@timeit
def msgpack_load(s):
    return msgpack.unpackb(s)

def msgpack_test(d):
    s= msgpack_dump(d)
    msgpack_load(s)

def main():
    d = get_dict()
    marshal_test(d)
    cPickle_test(d)
    tnetstring_test(d)
    msgpack_test(d)

if __name__ == "__main__":
    main()

在一台Ubuntu 10.04 32bit，双核Intel(R) Xeon(TM) CPU 3.06GHz机器上，有如下成绩：

marshal_dump used time 0.18213891983s
marshal_load used time 0.335217952728s
cPickle_dump used time 2.73869895935s
cPickle_load used time 0.558264017105s
tnetstring_dump used time 0.622045040131s
tnetstring_load used time 0.311910152435s
msgpack_dump used time 1.07127404213s
msgpack_load used time 0.27454996109s

结论

• cPickle最慢，基本上可以抛弃
• marshal虽然有不兼容隐患，但如果能保证相同的Python版本，还是可以一用
• tnetstring无论序列化或反序列化速度都比较中庸
• msgpack反序列化最快，如果读多写少的应用场景是最佳选择。而且msgpack支持多语言。