python性能分析工具

1）cProfile

cProfile可以嵌入到python代码中执行，比如：

import cProfile
cProfile.run('foo()', 'foo.out')

查看结果需要pstats模块，比如：

import pstats
p = pstats.Stats('foo.out')
p.print_stats()

pstats还可以排序， 以及打印排名靠前的记录。比如：

p.sort_stats('cumulative').print_stats(10)

也可以命令行方式独立运行，比如：

python -m cProfile -o profile.log my_test.py

通过pstats查看并将结果写入stats.out:

>>> import pstats
>>> stream = open('stats.out','w')
>>> p=pstats.Stats('profile.log',stream=stream)
>>> p.sort_stats('cumulative').print_stats()

查看stats.out:

Tue Mar  5 15:41:17 2019    profile.log

         349093 function calls (345925 primitive calls) in 0.989 seconds

   Ordered by: cumulative time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.005    0.005    1.028    1.028 my_test.py:1()
        1    0.001    0.001    0.609    0.609 build/bdist.linux-x86_64/egg/my_algo/utils.py:11()
        1    0.004    0.004    0.480    0.480 /usr/lib64/python2.7/site-packages/scipy/stats/__init__.py:364()
        1    0.002    0.002    0.454    0.454 /usr/lib64/python2.7/site-packages/scipy/stats/stats.py:158()
        1    0.009    0.009    0.346    0.346 /usr/lib64/python2.7/site-packages/scipy/stats/distributions.py:8()
        1    0.000    0.000    0.305    0.305 preprocess.py:4()

各列含义：
 

ncalls	调用次数
tottime	函数（代码块）总共执行时间， 不包含其调用的子函数执行时间
percall	单次调用占用时间，等于tottime除以ncalls
cumtime	函数（代码块）包含子调用的所有执行时间
percall	cumtime除以ncalls
filename:lineno(function)	函数（代码块）位置

pstats的sort_stats排序参数：

Valid Arg	Meaning
'calls'	call count
'cumulative'	cumulative time
'cumtime'	cumulative time
'file'	file name
'filename'	file name
'module'	file name
'ncalls'	call count
'pcalls'	primitive call count
'line'	line number
'name'	function name
'nfl'	name/file/line
'stdname'	standard name
'time'	internal time
'tottime'	internal time

pstats打印caller:  Stats.print_callers(*restrictions),  打印callee: Stats.print_callees(*restrictions)

例如：

>>> p.sort_stats('cumtime').print_callers(20)
   Ordered by: cumulative time
   List reduced from 3506 to 20 due to restriction <20>

Function                                                                                     was called by...
                                                                                                 ncalls  tottime  cumtime
my_test.py:1()                                                                        <-
build/bdist.linux-x86_64/egg/my_algo/utils.py:11()                                   <-       1    0.001    0.609  my_test.py:1()
/usr/lib64/python2.7/site-packages/scipy/stats/__init__.py:364()                     <-       1    0.004    0.480  build/bdist.linux-x86_64/egg/my_algo/utils.py:11()
/usr/lib64/python2.7/site-packages/scipy/stats/stats.py:158()                        <-       1    0.002    0.454  /usr/lib64/python2.7/site-packages/scipy/stats/__init__.py:364()
/usr/lib64/python2.7/site-packages/scipy/stats/distributions.py:8()                  <-       1    0.009    0.346  /usr/lib64/python2.7/site-packages/scipy/stats/stats.py:158()
pm_preprocess.py:4()                                                                 <-       1    0.000    0.305  my_test.py:1()
/usr/lib64/python2.7/site-packages/pandas/__init__.py:5()                            <-       1    0.004    0.302  preprocess.py:4()

 

2） line_profiler

cProfile可以定位到某个function, line_profiler可以定位到函数内部的每一行。可以通过@profile声明要分析的函数以及执行kernprof的方式来分析，比如：

@profile
def slow_function(a, b, c):
    ...

 然后执行：$ kernprof -l script_to_profile.py

分析结果保存在： script_to_profile.py.lprof，可以通过以下方式查看结果：

$ python -m line_profiler script_to_profile.py.lprof

也可以在执行kernprol的时候加入-v参数，直接在命令行显示：

$ kernprof -l -v script_to_profile.py

 

总结：性能分析的时候可以先用cProfile定位热点function, 再通过line_profile定位具体的热点行。

3）结果可视化

可以通过gprof2dot，pyprof2calltree，pyinstrument 等实现，以如下python程序为例：

num = 1000

def calculateSum(n):
    sum = 0
    counter = 1
    while counter <= n:
        sum = sum + counter
        counter += 1
    return sum

sumN = calculateSum(num)
print("The sum from 1 to %d is: %d" % (num,sumN))

gprof2dot结果可视化：

python -m cProfile -o output.pstats test.py

gprof2dot output.pstats -f pstats | dot -Tpng -o output.png

结果格式：

A node in the output graph represents a function and has the following layout:

+------------------------------+
|        function name         |
| total time % ( self time % ) |
|         total calls          |
+------------------------------+
where:

total time % is the percentage of the running time spent in this function and all its children;
self time % is the percentage of the running time spent in this function alone;
total calls is the total number of times this function was called (including recursive calls).
An edge represents the calls between two functions and has the following layout:

           total time %
              calls
parent --------------------> children
Where:

total time % is the percentage of the running time transfered from the children to this parent (if available);
calls is the number of calls the parent function called the children.

 

参考：

https://docs.python.org/3.2/library/profile.html

https://github.com/rkern/line_profiler

https://github.com/jrfonseca/gprof2dot

https://thirld.com/blog/2014/11/30/visualizing-the-results-of-profiling-python-code/