protobuf vs. msgpack in Python

benchmark.py

# -*- coding: utf-8 -*-
import time

import msgpack

from person_pb2 import Person


TIMES = 100000


def make_person(size):
    p = Person()
    p.id = 1
    p.name = 'Heungsub Lee'
    p.email = '[email protected]'
    for x in xrange(size):
        p.lucky_numbers.append(x)
    return p


def make_dict(size):
    return {'id': 1, 'name': 'Heungsub Lee', 'email': '[email protected]',
            'lucky_numbers': range(size)}


def make_tuple(size):
    return (1, 'Heungsub Lee', '[email protected]', range(size))


def protobuf_pack(times, size):
    p = make_person(size)
    t = time.time()
    for x in xrange(times):
        p.SerializeToString()
    return time.time() - t, len(p.SerializeToString())


def protobuf_unpack(times, size):
    s = make_person(size).SerializeToString()
    p = Person()
    t = time.time()
    for x in xrange(times):
        p.ParseFromString(s)
    return time.time() - t


def msgpack_pack(times, size, make=make_dict):
    d = make(size)
    t = time.time()
    for x in xrange(times):
        msgpack.packb(d)
    return time.time() - t, len(msgpack.packb(d))


def msgpack_unpack(times, size, make=make_dict):
    b = msgpack.packb(make(size))
    t = time.time()
    for x in xrange(times):
        msgpack.unpackb(b)
    return time.time() - t


if __name__ == '__main__':
    fmt = '{0:,}ns/pack {1:,}ns/unpack ({2:,} bytes)'
    ns = lambda d: d / TIMES * 10 ** 9
    for size in [0, 10, 100, 1000]:
        print 'array size: {}'.format(size)
        pack_duration, length = protobuf_pack(TIMES, size)
        unpack_duration = protobuf_unpack(TIMES, size)
        print 'protobuf',
        print fmt.format(ns(pack_duration), ns(unpack_duration), length)
        pack_duration, length = msgpack_pack(TIMES, size)
        unpack_duration = msgpack_unpack(TIMES, size)
        print 'msgpack with dict',
        print fmt.format(ns(pack_duration), ns(unpack_duration), length)
        pack_duration, length = msgpack_pack(TIMES, size, make=make_tuple)
        unpack_duration = msgpack_unpack(TIMES, size, make=make_tuple)
        print 'msgpack with tuple',
        print fmt.format(ns(pack_duration), ns(unpack_duration), length)


# array size: 0
# protobuf 12,794ns/pack 7,402ns/unpack (33 bytes)
# msgpack with dict 902ns/pack 517ns/unpack (60 bytes)
# msgpack with tuple 679ns/pack 286ns/unpack (32 bytes)
# array size: 10
# protobuf 23,731ns/pack 31,589ns/unpack (53 bytes)
# msgpack with dict 1,107ns/pack 620ns/unpack (70 bytes)
# msgpack with tuple 870ns/pack 436ns/unpack (42 bytes)
# array size: 100
# protobuf 112,659ns/pack 189,907ns/unpack (233 bytes)
# msgpack with dict 2,598ns/pack 1,533ns/unpack (162 bytes)
# msgpack with tuple 2,200ns/pack 1,291ns/unpack (134 bytes)
# array size: 1000
# protobuf 1,335,493ns/pack 2,117,658ns/unpack (2,905 bytes)
# msgpack with dict 18,968ns/pack 18,040ns/unpack (2,678 bytes)
# msgpack with tuple 16,929ns/pack 18,499ns/unpack (2,650 bytes)

person.proto

message Person {
  required int32 id = 1;
  required string name = 2;
  optional string email = 3;
  repeated int32 lucky_numbers = 4;
}

결론: "msgpack이 훨씬 빠르고 바이트도 적게 쓴다"가 맞나요?