PyPy比CPython慢很多。

14

我一直在测试自己制作的缓存系统,它的目的是加速Django Web应用程序。它将所有内容存储在内存中。根据cProfile,在我的测试中大部分时间都花费在QuerySet._clone()内部,这证明效率极低(考虑到实现方式,其实也不奇怪)。

我非常期望使用PyPy来提高速度。我有一个64位机器。然而,在安装了所有必需的库之后,结果显示,与普通Python代码相比,PyPy编译的代码运行速度约慢2.5倍,我不知道该如何解释。代码绑定CPU(完全没有数据库查询,因此IO绑定不是选项)。单个测试运行约10秒,所以我想应该足够触发JIT即时编译了。我正在使用PyPy 1.5。请注意,我没有自己编译源代码,只是下载了一个64位的Linux版本。

我想知道,CPU密集型代码实际上在PyPy下运行更慢的频率有多高。希望我能做错了什么,阻止PyPy的最佳运行。

编辑:

确切的cPython输出:

PyPy 1.5:

    3439146 function calls (3218654 primitive calls) in 19.094 seconds

    Ordered by: cumulative time

    ncalls  tottime  percall  cumtime  percall filename:lineno(function)
       2/1    0.000    0.000   18.956   18.956 <string>:1(<module>)
       2/1    0.000    0.000   18.956   18.956 /path/to/my/project/common/integrity/models/transactions.py:200(newfn)
       2/1    0.000    0.000   18.956   18.956 /path/to/my/project/common/integrity/models/transactions.py:134(recur)
       2/1    0.000    0.000   18.956   18.956 /usr/local/pypy/site-packages/django/db/transaction.py:210(inner)
       2/1    0.172    0.086   18.899   18.899 /path/to/my/project/common/integrity/tests/optimization.py:369(func_cached)
      9990    0.122    0.000   18.632    0.002 /usr/local/pypy/site-packages/django/db/models/manager.py:131(get)
      9990    0.127    0.000   16.638    0.002 /path/to/my/project/common/integrity/models/cache.py:1068(get)
      9990    0.073    0.000   12.478    0.001 /usr/local/pypy/site-packages/django/db/models/query.py:547(filter)
      9990    0.263    0.000   12.405    0.001 /path/to/my/project/common/integrity/models/cache.py:1047(_filter_or_exclude)
      9990    0.226    0.000   12.096    0.001 /usr/local/pypy/site-packages/django/db/models/query.py:561(_filter_or_exclude)
      9990    0.187    0.000    8.383    0.001 /path/to/my/project/common/integrity/models/cache.py:765(_clone)
      9990    0.212    0.000    7.662    0.001 /usr/local/pypy/site-packages/django/db/models/query.py:772(_clone)
      9990    1.025    0.000    7.125    0.001 /usr/local/pypy/site-packages/django/db/models/sql/query.py:226(clone)
129942/49972  1.674    0.000    6.021    0.000 /usr/local/pypy/lib-python/2.7/copy.py:145(deepcopy)
140575/110605 0.120    0.000    4.066    0.000 {len}
      9990    0.182    0.000    3.972    0.000 /usr/local/pypy/site-packages/django/db/models/query.py:74(__len__)
     19980    0.260    0.000    3.777    0.000 /path/to/my/project/common/integrity/models/cache.py:1062(iterator)
      9990    0.255    0.000    3.154    0.000 /usr/local/pypy/site-packages/django/db/models/sql/query.py:1149(add_q)
      9990    0.210    0.000    3.073    0.000 /path/to/my/project/common/integrity/models/cache.py:973(_query)
      9990    0.371    0.000    2.316    0.000 /usr/local/pypy/site-packages/django/db/models/sql/query.py:997(add_filter)
      9990    0.364    0.000    2.168    0.000 /path/to/my/project/common/integrity/models/cache.py:892(_deduct)
29974/9994    0.448    0.000    2.078    0.000 /usr/local/pypy/lib-python/2.7/copy.py:234(_deepcopy_tuple)
     19990    0.362    0.000    2.065    0.000 /path/to/my/project/common/integrity/models/cache.py:566(__init__)
     10000    0.086    0.000    1.874    0.000 /path/to/my/project/common/integrity/models/cache.py:1090(get_query_set)
     19990    0.269    0.000    1.703    0.000 /usr/local/pypy/site-packages/django/db/models/query.py:31(__init__)
      9990    0.122    0.000    1.643    0.000 /path/to/my/project/common/integrity/models/cache.py:836(_deduct_recur)
     19980    0.274    0.000    1.636    0.000 /usr/local/pypy/site-packages/django/utils/tree.py:55(__deepcopy__)
      9990    0.607    0.000    1.458    0.000 /path/to/my/project/common/integrity/models/cache.py:789(_deduct_local)
     10020    0.633    0.000    1.437    0.000 /usr/local/pypy/site-packages/django/db/models/sql/query.py:99(__init__)
    129942    0.841    0.000    1.191    0.000 /usr/local/pypy/lib-python/2.7/copy.py:267(_keep_alive)
 9994/9992    0.201    0.000    1.019    0.000 /usr/local/pypy/lib-python/2.7/copy.py:306(_reconstruct)

Python 2.7:

   3326403 function calls (3206359 primitive calls) in 12.430 CPU seconds

   Ordered by: cumulative time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.000    0.000   12.457   12.457 <string>:1(<module>)
        1    0.000    0.000   12.457   12.457 /path/to/my/project/common/integrity/models/transactions.py:200(newfn)
        1    0.000    0.000   12.457   12.457 /path/to/my/project/common/integrity/models/transactions.py:134(recur)
        1    0.000    0.000   12.457   12.457 /usr/local/lib/python2.7/dist-packages/django/db/transaction.py:210(inner)
        1    0.000    0.000   12.457   12.457 /path/to/my/project/common/integrity/models/transactions.py:165(recur2)
        1    0.089    0.089   12.450   12.450 /path/to/my/project/common/integrity/tests/optimization.py:369(func_cached)
     9990    0.198    0.000   12.269    0.001 /usr/local/lib/python2.7/dist-packages/django/db/models/manager.py:131(get)
     9990    0.087    0.000   11.281    0.001 /path/to/my/project/common/integrity/models/cache.py:1068(get)
     9990    0.040    0.000    8.161    0.001 /usr/local/lib/python2.7/dist-packages/django/db/models/query.py:547(filter)
     9990    0.110    0.000    8.121    0.001 /path/to/my/project/common/integrity/models/cache.py:1047(_filter_or_exclude)
     9990    0.127    0.000    7.983    0.001 /usr/local/lib/python2.7/dist-packages/django/db/models/query.py:561(_filter_or_exclude)
     9990    0.100    0.000    5.593    0.001 /path/to/my/project/common/integrity/models/cache.py:765(_clone)
     9990    0.122    0.000    5.125    0.001 /usr/local/lib/python2.7/dist-packages/django/db/models/query.py:772(_clone)
     9990    0.405    0.000    4.899    0.000 /usr/local/lib/python2.7/dist-packages/django/db/models/sql/query.py:226(clone)
129942/49972 1.456    0.000    4.505    0.000 /usr/lib/python2.7/copy.py:145(deepcopy)
129899/99929 0.191    0.000    3.117    0.000 {len}
     9990    0.111    0.000    2.968    0.000 /usr/local/lib/python2.7/dist-packages/django/db/models/query.py:74(__len__)
    19980    0.070    0.000    2.843    0.000 /path/to/my/project/common/integrity/models/cache.py:1062(iterator)
     9990    0.208    0.000    2.190    0.000 /path/to/my/project/common/integrity/models/cache.py:973(_query)
     9990    0.182    0.000    2.114    0.000 /usr/local/lib/python2.7/dist-packages/django/db/models/sql/query.py:1149(add_q)
19984/9994   0.291    0.000    1.644    0.000 /usr/lib/python2.7/copy.py:234(_deepcopy_tuple)
     9990    0.288    0.000    1.599    0.000 /usr/local/lib/python2.7/dist-packages/django/db/models/sql/query.py:997(add_filter)
     9990    0.171    0.000    1.454    0.000 /path/to/my/project/common/integrity/models/cache.py:892(_deduct)
    19980    0.177    0.000    1.208    0.000 /usr/local/lib/python2.7/dist-packages/django/utils/tree.py:55(__deepcopy__)
     9990    0.099    0.000    1.199    0.000 /path/to/my/project/common/integrity/models/cache.py:836(_deduct_recur)
     9990    0.349    0.000    1.040    0.000 /path/to/my/project/common/integrity/models/cache.py:789(_deduct_local)

3
你有对 PyPy 代码进行分析,看看它在哪些地方花费了过多的时间吗? - agf
2
我们很想看到这段代码。 - Jakob Bowyer
2
好吧,从长远来看,它似乎是一样的。在40秒的差异几乎没有什么区别。这令人失望。 - julx
3
您可以在此处看到:http://speed.pypy.org/,有时pypy会更慢。但是没有代码,就没什么更多的可说的了... - xubuntix
5
没有进行分析(并且没有启用调试模式/工具栏),运行速度会变慢吗?这被认为会大大降低PyPy的运行速度。 - TryPyPy
显示剩余4条评论
1个回答

20

抛开PyPy在您的情况下可能本质上更慢这一事实,有些因素可能会使它不必要地变慢:

  • 已知性能分析比CPython更慢。
  • 一些调试/日志记录代码可能会禁用优化(例如:强制帧)。
  • 您使用的服务器可能是性能的主导因素(想想如果JIT与经典CGI一起使用会多么糟糕:它永远无法预热)。 它还可以简单地影响结果(不同的WSGI服务器显示了各种加速度)。
  • 旧式类比新式类更慢。
  • 即使一切都在内存中,您也可能会遇到例如PyPy的SQLite中的缓慢路径。

您还可以查看JIT Friendliness维基页面,了解更多可能使PyPy变慢的提示。 每夜构建版本也可能会更快,因为相对于1.5有许多改进。

有关您的堆栈(服务器、操作系统、数据库)和设置的更详细说明(您如何进行基准测试?有多少查询?)将使我们能够给出更好的答案。


2
此外,由于当前卡片引用垃圾回收机制,Pypy中生成非常大的字典会变慢。如果所有内容都被缓存,那么就意味着所有东西都被放入了一个大字典中。希望我们能尽快找到一些加速的方法。 - Justin Peel
更新的链接:https://foss.heptapod.net/pypy/pypy/-/wikis/JitFriendliness 和 https://www.pypy.org/performance.html - Phani Rithvij

网页内容由stack overflow 提供, 点击上面的
可以查看英文原文,
原文链接