Python 性能调试工具Line-profiler极简使用笔记
2021-01-08 本文已影响0人
井底蛙蛙呱呱呱
虽然Python屡屡被人诟病速度问题,但是该用的还得用,速度问题只能靠代码优化来解决了。Line-Profiler是一个代码优化工具,利用line—profiler我们可以得到我们每一行代码的运行总时间以及单次平均运行时间,以便我们对耗时最长的地方进行优化。
安装:
pip install line_profiler
1、极简模式
下面我们使用line-profiler查看一个简单实例各行代码时间都花在哪。
import random
def do_stuff():
numbers = []
for i in range(1000):
numbers.append(random.randint(0,1000))
s = sum(numbers)
l = [numbers[i]/43 for i in range(len(numbers))]
m = ['hello'+str(numbers[i]) for i in range(len(numbers))]
return None
if __name__ == '__main__':
from line_profiler import LineProfiler
lp = LineProfiler()
lp_wrapper = lp(do_stuff)
lp_wrapper()
lp.print_stats()
下面我们再命令行运行一下看看时间都去哪了:
$ root@cpu-k8ss-0 # python loader_test.py
Timer unit: 1e-06 s
Total time: 0.0075 s
File: loader_test.py
Function: do_stuff at line 88
Line # Hits Time Per Hit % Time Line Contents
==============================================================
88 def do_stuff():
89 1 1.0 1.0 0.0 numbers = []
90 1001 697.0 0.7 9.3 for i in range(1000):
91 1000 6093.0 6.1 81.2 numbers.append(random.randint(0,1000))
92 1 10.0 10.0 0.1 s = sum(numbers)
93 1 240.0 240.0 3.2 l = [numbers[i]/43 for i in range(len(numbers))]
94 1 458.0 458.0 6.1 m = ['hello'+str(numbers[i]) for i in range(len(numbers))]
95 1 1.0 1.0 0.0 return None
上面输出中,可以看到我们测试的函数为do_stuff
, 起始于位于脚本的第88行。运行程序共花费时间0.0075 s。
紧接着的各列含义如下:
-
Line #
, 对应脚本中的代码行数; -
Hits
,对应行共计运行次数; -
Time
,对应行运行总时间; -
Per Hit
,对应行平均每次运行时间; -
% Time
, 对应行运行时间占程序运行总时间的比例; -
Line Contents
,对应的每一行的代码内容。
从上面可以看到我们的时间主要花在了第一个生成数据的for循环中,共计占了90.5%,我们可以使用列表推导式来对其进行优化。
$ root@cpu-k8ss-0 # python loader_test.py
Timer unit: 1e-06 s
Total time: 0.00503 s
File: loader_test.py
Function: do_stuff at line 88
Line # Hits Time Per Hit % Time Line Contents
==============================================================
88 def do_stuff():
89 # numbers = []
90 # for i in range(1000):
91 # numbers.append(random.randint(0,1000))
92 1 4434.0 4434.0 88.2 numbers = [random.randint(1,100) for i in range(1000)]
93 1 8.0 8.0 0.2 s = sum(numbers)
94 1 194.0 194.0 3.9 l = [numbers[i]/43 for i in range(len(numbers))]
95 1 393.0 393.0 7.8 m = ['hello'+str(numbers[i]) for i in range(len(numbers))]
96 1 1.0 1.0 0.0 return None
可以看到使用列表推导式后,我们的总时间减少了,生成数据这一步占程序总运行时间比例也降低了。
给函数传入参数
下面我们将numbers数据生成过程放到外面,以参数形式传入到函数中:
def do_stuff(numbers):
# numbers = []
# for i in range(1000):
# numbers.append(random.randint(0,1000))
#numbers = [random.randint(1,100) for i in range(1000)]
s = sum(numbers)
l = [numbers[i]/43 for i in range(len(numbers))]
m = ['hello'+str(numbers[i]) for i in range(len(numbers))]
return None
if __name__ == '__main__':
from line_profiler import LineProfiler
lp = LineProfiler()
lp_wrapper = lp(do_stuff)
# 生成参数并传入
numbers = [random.randint(1,100) for i in range(100000)]
lp_wrapper(numbers)
lp.print_stats()
命令行运行:
$ root@cpu-k8ss-0 # python loader_test.py
Timer unit: 1e-06 s
Total time: 0.084887 s
File: loader_test.py
Function: do_stuff at line 88
Line # Hits Time Per Hit % Time Line Contents
==============================================================
88 def do_stuff(numbers):
89 # numbers = []
90 # for i in range(1000):
91 # numbers.append(random.randint(0,1000))
92 #numbers = [random.randint(1,100) for i in range(1000)]
93 1 788.0 788.0 0.9 s = sum(numbers)
94 1 26504.0 26504.0 31.2 l = [numbers[i]/43 for i in range(len(numbers))]
95 1 57593.0 57593.0 67.8 m = ['hello'+str(numbers[i]) for i in range(len(numbers))]
96 1 2.0 2.0 0.0 return None
可以看到,现在程序运行的时间主要花在计算除数和连接字符串了。
显示内层调用函数的运行时间
在上面的例子中,我们将所有代码放在了一个函数中,只查看这个函数的代码的运行时间。line-profiler支持查看调用的函数内部运行时间。
def generate_numbers(n):
numbers = [random.randint(1,100) for i in range(1000)]
return numbers
def do_stuff(n):
numbers = generate_numbers(n)
s = sum(numbers)
l = [numbers[i]/43 for i in range(len(numbers))]
m = ['hello'+str(numbers[i]) for i in range(len(numbers))]
return None
if __name__ == '__main__':
from line_profiler import LineProfiler
lp = LineProfiler()
lp_wrapper = lp(do_stuff)
# 加入并显示调用函数各行代码用时
lp.add_function(generate_numbers)
# 生成参数并传入
n = 100000
lp_wrapper(n)
lp.print_stats()
命令行运行:
$ root@cpu-k8ss-0 # python loader_test.py
Timer unit: 1e-06 s
Total time: 0.005112 s
File: loader_test.py
Function: generate_numbers at line 87
Line # Hits Time Per Hit % Time Line Contents
==============================================================
87 def generate_numbers(n):
88 1 5111.0 5111.0 100.0 numbers = [random.randint(1,100) for i in range(1000)]
89 1 1.0 1.0 0.0 return numbers
Total time: 0.005709 s
File: loader_test.py
Function: do_stuff at line 92
Line # Hits Time Per Hit % Time Line Contents
==============================================================
92 def do_stuff(n):
93 1 5125.0 5125.0 89.8 numbers = generate_numbers(n)
94 1 8.0 8.0 0.1 s = sum(numbers)
95 1 189.0 189.0 3.3 l = [numbers[i]/43 for i in range(len(numbers))]
96 1 387.0 387.0 6.8 m = ['hello'+str(numbers[i]) for i in range(len(numbers))]
97 1 0.0 0.0 0.0 return None
在上面我们可以看到两个函数各行代码的运行时间,非常方便。
当然,line-profile还可以通过装饰器@profiler方式统计我们想要优化的函数。然后在命令行中通过命令$ kernprof -l script_to_profile.py
来进行运行。
参考:
官方说明
使用line_profiler查看api接口函数每行代码执行时间
python 性能调试工具(line_profiler)使用