Python 性能分析工具简介

ChlZMF 4年前
   <p><strong>性能分析和调优工具简介</strong></p>    <p>总会遇到一个时候你会想提高程序执行效率,想看看哪部分耗时长成为瓶颈,想知道程序运行时内存和CPU使用情况。这时候你会需要一些方法对程序进行性能分析和调优。</p>    <h2><strong>By Context Manager</strong></h2>    <p>可以上下文管理器自己实现一个计时器, 参见之前的介绍 timeit 文章里做的那样,通过定义类的 __enter__ 和 __exit__ 方法来实现对管理的函数计时, 类似如:</p>    <pre>  <code class="language-python"># timer.py  importtime     class Timer(object):      def__init__(self, verbose=False):          self.verbose = verbose         def__enter__(self):          self.start = time.time()          return self         def__exit__(self, *args):          self.end = time.time()          self.secs = self.end - self.start          self.msecs = self.secs * 1000            # 毫秒          if self.verbose:              print 'elapsed time: %f ms' % self.msecs     </code></pre>    <p>使用方式如下:</p>    <pre>  <code class="language-python">fromtimerimportTimer     withTimer() as t:      foo()  print "=> foo() spends %s s" % t.secs     </code></pre>    <h2><strong>By Decorator</strong></h2>    <p>然而我认为装饰器的方式更加优雅</p>    <pre>  <code class="language-python">importtime  fromfunctoolsimportwraps     deftimer(function):      @wraps(function)      deffunction_timer(*args, **kwargs):          t0 = time.time()          result = function(*args, **kwargs)          t1 = time.time()          print ("Total time running %s: %s seconds" %                  (function.func_name, str(t1-t0))                  )          return result      return function_timer     </code></pre>    <p>使用就很简单了:</p>    <pre>  <code class="language-python"><a href="http://www.jobbole.com/members/Timer">@timer</a>  defmy_sum(n):      return sum([i for i in range(n)])     if __name__ == "__main__":      my_sum(10000000)     </code></pre>    <p>运行结果:</p>    <pre>  <code class="language-python">➜  pythonprofile.py  Totaltimerunningmy_sum: 0.817697048187 seconds     </code></pre>    <h2><strong>系统自带的time命令</strong></h2>    <p>使用示例如下:</p>    <pre>  <code class="language-python">➜ timepythonprofile.py  Totaltimerunningmy_sum: 0.854454040527 seconds  pythonprofile.py  0.79s user 0.18s system 98% cpu 0.977 total     </code></pre>    <p>上面的结果说明: 执行脚本消耗0.79sCPU时间, 0.18秒执行内核函数消耗的时间,总共0.977s时间。</p>    <p>其中, total时间 - (user时间 + system时间) = 消耗在输入输出和系统执行其它任务消耗的时间</p>    <h2><strong>python timeit 模块</strong></h2>    <p>可以用来做benchmark, 可以方便的重复一个程序执行的次数,来查看程序可以运行多块。</p>    <h2><strong>cProfile</strong></h2>    <p>直接看带注释的使用示例吧。</p>    <pre>  <code class="language-python">#coding=utf8     defsum_num(max_num):      total = 0      for i in range(max_num):          total += i      return total        deftest():      total = 0      for i in range(40000):          total += i         t1 = sum_num(100000)      t2 = sum_num(200000)      t3 = sum_num(300000)      t4 = sum_num(400000)      t5 = sum_num(500000)      test2()         return total     deftest2():      total = 0      for i in range(40000):          total += i         t6 = sum_num(600000)      t7 = sum_num(700000)         return total        if __name__ == "__main__":      importcProfile         # # 直接把分析结果打印到控制台      # cProfile.run("test()")      # # 把分析结果保存到文件中      # cProfile.run("test()", filename="result.out")      # 增加排序方式      cProfile.run("test()", filename="result.out", sort="cumulative")     </code></pre>    <p>cProfile将分析的结果保存到result.out文件中,但是以二进制形式存储的,想直接查看的话用提供的 pstats 来查看。</p>    <pre>  <code class="language-python">importpstats     # 创建Stats对象  p = pstats.Stats("result.out")     # strip_dirs(): 去掉无关的路径信息  # sort_stats(): 排序,支持的方式和上述的一致  # print_stats(): 打印分析结果,可以指定打印前几行     # 和直接运行cProfile.run("test()")的结果是一样的  p.strip_dirs().sort_stats(-1).print_stats()     # 按照函数名排序,只打印前3行函数的信息, 参数还可为小数,表示前百分之几的函数信息  p.strip_dirs().sort_stats("name").print_stats(3)     # 按照运行时间和函数名进行排序  p.strip_dirs().sort_stats("cumulative", "name").print_stats(0.5)     # 如果想知道有哪些函数调用了sum_num  p.print_callers(0.5, "sum_num")     # 查看test()函数中调用了哪些函数  p.print_callees("test")     </code></pre>    <p>截取一个查看test()调用了哪些函数的输出示例:</p>    <pre>  <code class="language-python">➜  pythonpythonprofile.py    Randomlistingorderwasused    Listreducedfrom 6 to 2 dueto restriction     Function              called...                            ncalls  tottime  cumtime  profile.py:24(test2)  ->      2    0.061    0.077  profile.py:3(sum_num)                                1    0.000    0.000  {range}  profile.py:10(test)  ->      5    0.073    0.094  profile.py:3(sum_num)                                1    0.002    0.079  profile.py:24(test2)                                1    0.001    0.001  {range}     </code></pre>    <h3><strong>profile.Profile</strong></h3>    <p>cProfile还提供了可以自定义的类,可以更精细的分析。</p>    <p>格式如: class profile.Profile(timer=None, timeunit=0.0, subcalls=True, builtins=True)</p>    <p>下面这个例子来自官方文档:</p>    <pre>  <code class="language-python">importcProfile, pstats, StringIO  pr = cProfile.Profile()  pr.enable()  # ... do something ...  pr.disable()  s = StringIO.StringIO()  sortby = 'cumulative'  ps = pstats.Stats(pr, stream=s).sort_stats(sortby)  ps.print_stats()  print s.getvalue()     </code></pre>    <h2><strong><sub>line profiler</sub></strong></h2>    <p><sub>line profiler是一个对函数进行逐行性能分析的工具</sub></p>    <h3><strong>示例</strong></h3>    <pre>  <code class="language-python">#coding=utf8     defsum_num(max_num):      total = 0      for i in range(max_num):          total += i      return total        @profile                    # 添加@profile 来标注分析哪个函数  deftest():      total = 0      for i in range(40000):          total += i         t1 = sum_num(10000000)      t2 = sum_num(200000)      t3 = sum_num(300000)      t4 = sum_num(400000)      t5 = sum_num(500000)      test2()         return total     deftest2():      total = 0      for i in range(40000):          total += i         t6 = sum_num(600000)      t7 = sum_num(700000)         return total     test()     </code></pre>    <p>通过 kernprof 命令来注入分析,运行结果如下:</p>    <pre>  <code class="language-python">➜ kernprof -l -v profile.py  Wroteprofileresultsto profile.py.lprof  Timerunit: 1e-06 s     Totaltime: 3.80125 s  File: profile.py  Function: testatline 10     Line #      Hits         Time  Per Hit   % Time  Line Contents  ==============================================================      10                                          @profile      11                                          deftest():      12        1            5      5.0      0.0      total = 0      13    40001        19511      0.5      0.5      for i in range(40000):      14    40000        19066      0.5      0.5          total += i      15      16        1      2974373 2974373.0    78.2      t1 = sum_num(10000000)      17        1        58702  58702.0      1.5      t2 = sum_num(200000)      18        1        81170  81170.0      2.1      t3 = sum_num(300000)      19        1      114901 114901.0      3.0      t4 = sum_num(400000)      20        1      155261 155261.0      4.1      t5 = sum_num(500000)      21        1      378257 378257.0    10.0      test2()      22      23        1            2      2.0      0.0      return total     </code></pre>    <p>hits(执行次数) 和 time(耗时) 值高的地方是有比较大优化空间的地方。</p>    <h2><strong>memory <sub>profiler</sub></strong></h2>    <p>类似于”line <sub>profiler</sub> “对基于行分析程序内存使用情况的模块。</p>    <p>同样是上面”line <sub>profiler</sub> “中的代码,运行 python -m memory_profiler profile.py 命令生成结果如下:</p>    <pre>  <code class="language-python">➜ python -m memory_profilerprofile.py  Filename: profile.py     Line #    Mem usage    Increment   Line Contents  ================================================      10  24.473 MiB    0.000 MiB  @profile      11                            deftest():      12  24.473 MiB    0.000 MiB      total = 0      13  25.719 MiB    1.246 MiB      for i in range(40000):      14  25.719 MiB    0.000 MiB          total += i      15      16  335.594 MiB  309.875 MiB      t1 = sum_num(10000000)      17  337.121 MiB    1.527 MiB      t2 = sum_num(200000)      18  339.410 MiB    2.289 MiB      t3 = sum_num(300000)      19  342.465 MiB    3.055 MiB      t4 = sum_num(400000)      20  346.281 MiB    3.816 MiB      t5 = sum_num(500000)      21  356.203 MiB    9.922 MiB      test2()      22      23  356.203 MiB    0.000 MiB      return total     </code></pre>    <h2><strong>TODO objgraph</strong></h2>    <h2><strong>参考资料:</strong></h2>    <ul>     <li><a href="/misc/goto?guid=4959725694060343510" rel="nofollow,noindex">https://docs.python.org/2/lib…</a></li>     <li><a href="/misc/goto?guid=4959725694220354892" rel="nofollow,noindex">http://xianglong.me/article/a…</a></li>     <li><a href="/misc/goto?guid=4959725694308924738" rel="nofollow,noindex">http://www.cnblogs.com/btchen…</a></li>     <li><a href="/misc/goto?guid=4959725694393053413" rel="nofollow,noindex">https://www.huyng.com/posts/p…</a></li>     <li><a href="/misc/goto?guid=4959725694479783973" rel="nofollow,noindex">http://www.marinamele.com/7-t…</a></li>    </ul>    <p> </p>    <p>来自:http://python.jobbole.com/86833/</p>    <p> </p>