一文详解如何分析python程序cpu占用率问题

import cProfile

def main():
    ...  # 程序入口

if __name__ == "__main__":
    cProfile.run('main()', sort='cumtime')

你是一位资深的 python 专家和系统运维专家

你的任务是使用 cProfile 工具排查 python 程序 cpu 利用率过高的问题，并提出对应的优化和解决建议

cProfile 的输出如下

按照 cumtime 排序

---
ncalls	tottime  pe	rcall	cumtime  pe	rcall	filename:lineno(function)
91313	0.516	0	56.86	0.001	conv.py:53(forward_fuse)
27123	0.685	0	35.403	0.001	block.py:346(forward)
10435/10350	0.086	0	34.571	0	03 {method 'extend' of 'list' objects}
19892	0.086	0	32.822	0.002	block.py:239(<genexpr>)
97642	0.352	0	32.226	0	conv.py:553(forward)
3616	0.097	0	26.887	0.007	block.py:236(forward)
97642	0.259	0	21.341	0	conv.py:536(_conv_forward)
3616	0.106	0	20.761	0.006	block.py:459(forward)
97642	10.577	0	17.058	0	{built-in method torch.conv2d}
2372/2090	0.051	0	15.157	0.007	pooled_db.py:360(cache)
3671/3328	0.115	0	14.816	0.004	serving.py:259(write)
14464	0.043	0	13.573	0.001	block.py:462(<genexpr>)
7713/7070	0.046	0	12.46	0.002	socket.py:500(close)
3669/2219	0.1	0	12.376	0.006	selectors.py:451(select)
3798/3596	0.047	0	12.338	0.003	connections.py:431(_force_close)
2372/2091	0.042	0	12.221	0.006	steady_db.py:326(_reset)
904	0.107	0	11.436	0.013	head.py:317(forward)
87697	0.187	0	11.194	0	activation.py:431(forward)
2372/2091	0.024	0	9.936	0.005	steady_db.py:436(rollback)
3671/3585	0.056	0	9.425	0.003	server.py:493(send_response)
...(省略)
---

按照 tottime 如下

---
ncalls	tottime  pe	rcall	cumtime  pe	rcall	filename:lineno(function)
97642	10.577	0	17.058	0	{built-in method torch.conv2d}
1492/89	10.445	0.007	0.751	0.008	{method 'read' of 'cv2.VideoCapture' objects}
3660/317	9.277	0.003	2.85	0.009	{method 'poll' of 'select.poll' objects}
741/0	6.681	0.009	0		{built-in method time.sleep}
3669/2219	3.78	0.001	7.225	0.003	{method 'poll' of 'select.epoll' objects}
1522/185	2.446	0.002	0.4	0.002	{flip}
6328	1.948	0	3.359	0.001	{built-in method torch.einsum}
345354/12	1.613	0	1.368	0.114	module.py:1740(_call_impl)
903/1	1.581	0.002	0.003	0.003	{resize}
87697	1.573	0	4.584	0	{built-in method torch._C._nn.silu_}
16399/10112	1.556	0	3.308	0	00 {method 'recv_into' of '_socket.socket' objects}
23770/139	1.544	0	0.19	0.001	{method 'sendall' of '_socket.socket' objects}
345353/12	1.184	0	1.368	0.114	module.py:1732(_wrapped_call_impl)
3616	1.117	0	1.921	0.001	{built-in method torch._C._nn.linear}
439192	1.1	0	1.347	0	client.py:75(__setitem__)
9352/153	1.043	0	0.016	0	{method 'recv' of '_socket.socket' objects}
53147/50648	0.798	0	2.345	0	00 {method 'settimeout' of '_socket.socket' objects}
516359	0.783	0	0.783	0	module.py:1918(__getattr__)
18081	0.722	0	1.736	0	{built-in method torch.cat}
3616	0.689	0	7.775	0.002	block.py:426(forward)
...(省略)
---

你的分析结果应该包含每个问题的影响程度，以及你分析的依据（即从 cProfile 哪个数据得到的这个结论）

pip install py-spy

# -o 输出文件名
# -d 采样多久
py-spy record -o profile.svg -d 60 -p <pid>

pip install yappi

yappi.set_clock_type("cpu")  # cpu or wall
yappi.start()

my_function()

yappi.stop()
stats = yappi.get_func_stats()
# stats.print_all()
stats.save('yappi_stats.callgrind', type='callgrind')

import yappi
import signal

def start_all():
    app.run(debug=False, host='0.0.0.0', port=5000)


def handle_sigusr1(signum, frame):
    print("接收到 SIGUSR1 信号，正在停止 yappi 分析并保存结果...")
    yappi.stop()
    stats = yappi.get_func_stats()
    stats.save('yappi_stats.callgrind', type='callgrind')
    print("yappi 分析结果已保存为 yappi_stats.callgrind")

if __name__ == '__main__':
    yappi.set_clock_type("cpu")
    yappi.start()

    # 注册 SIGUSR1 信号处理函数
    signal.signal(signal.SIGUSR1, handle_sigusr1)

    start_all()