目录
前言
GPU功耗检测方法
CPU功耗检测方法
sudo的困扰与解决
完整功耗分析示例代码
转载请注明出处:小锋学长生活大爆炸[xfxuezhang.cn]
相关一些检测工具挺多的,比如powertop、powerstat、s-tui等。但如何通过代码的方式来实时检测,是个麻烦的问题。通过许久的搜索和自己的摸索,发现了可以检测CPU和GPU功耗的方法。如果有什么不对,或有更好的方法,欢迎评论留言!
文末附完整功耗分析的示例代码!
如果是常规的工具,可以使用官方的NVML。但这里需要Python控制,所以使用了对应的封装:pynvml。
先安装:
pip install pynvml
关于这个库,网上的使用教程挺多的。这里直接给出简单的示例代码:
import pynvml
pynvml.nvmlInit()
handle = pynvml.nvmlDeviceGetHandleByIndex(0)
powerusage = pynvml.nvmlDeviceGetPowerUsage(handle) / 1000
这个方法获取的值,跟使用“nvidia-smi”指令得到的是一样的。
附赠一个来自网上的获取更详细信息的函数:
def get_sensor_values():
"""
get Sensor values
:return:
"""
values = list()
# get gpu driver version
version = pynvml.nvmlSystemGetDriverVersion()
values.append("GPU_device_driver_version:" + version.decode())
gpucount = pynvml.nvmlDeviceGetCount() # 显示有几块GPU
for gpu_id in range(gpucount):
handle = pynvml.nvmlDeviceGetHandleByIndex(gpu_id)
name = pynvml.nvmlDeviceGetName(handle).decode()
meminfo = pynvml.nvmlDeviceGetMemoryInfo(handle)
# print(meminfo.total) # 显卡总的显存大小
gpu_id = str(gpu_id)
values.append("GPU " + gpu_id + " " + name + " 总共显存大小:" + str(common.bytes2human(meminfo.total)))
# print(meminfo.used) # 显存使用大小
values.append("GPU " + gpu_id + " " + name + " 显存使用大小:" + str(common.bytes2human(meminfo.used)))
# print(meminfo.free) # 显卡剩余显存大小
values.append("GPU " + gpu_id + " " + name + " 剩余显存大小:" + str(common.bytes2human(meminfo.free)))
values.append("GPU " + gpu_id + " " + name + " 剩余显存比例:" + str(int((meminfo.free / meminfo.total) * 100)))
utilization = pynvml.nvmlDeviceGetUtilizationRates(handle)
# print(utilization.gpu) # gpu利用率
values.append("GPU " + gpu_id + " " + name + " GPU利用率:" + str(utilization.gpu))
powerusage = pynvml.nvmlDeviceGetPowerUsage(handle)
# print(powerusage / 1000) # 当前功耗, 原始单位是mWa
values.append("GPU " + gpu_id + " " + name + " 当前功耗(W):" + str(powerusage / 1000))
# 当前gpu power capacity
# pynvml.nvmlDeviceGetEnforcedPowerLimit(handle)
# 通过以下方法可以获取到gpu的温度,暂时采用ipmi sdr获取gpu的温度,此处暂不处理
# temp = pynvml.nvmlDeviceGetTemperature(handle,0)
print('\n'.join(values))
return values
这个没有找到开源可以直接用的库。但经过搜索,发现大家都在用的s-tui工具是开源的!通过查看源码,发现他是有获取CPU功耗部分的代码,所以就参考他的源码写了一下。
先安装:
sudo apt install s-tui
pip install s-tui
先直接运行工具看一下效果(不使用sudo是不会出来Power的):
sudo s-tui
说明这个工具确实能获取到CPU的功耗。其中package就是2个CPU,dram是内存条功耗(一般不准,可以不用)。
直接给出简单的示例代码:
from s_tui.sources.rapl_power_source import RaplPowerSource
source.update()
summary = dict(source.get_sensors_summary())
cpu_power_total = str(sum(list(map(float, [summary[key] for key in summary.keys() if key.startswith('package')]))))
不过注意!由于需要sudo权限,所以运行这个py文件时候,也需要sudo方式,比如:
sudo python demo.py
上面提到,由于必须要sudo方式,但sudo python就换了运行脚本的环境了呀,这个比较棘手。后来想了个方法,曲线救国一下。通过sudo运行一个脚本,并开启socket监听;而我们自己真正的脚本,在需要获取CPU功耗时候,连接一下socket就行。
为什么这里使用socket而不是http呢?因为socket更高效一点!
我们写一个“power_listener.py”来监听:
from s_tui.sources.rapl_power_source import RaplPowerSource
import socket
import json
def output_to_terminal(source):
results = {}
if source.get_is_available():
source.update()
source_name = source.get_source_name()
results[source_name] = source.get_sensors_summary()
for key, value in results.items():
print(str(key) + ": ")
for skey, svalue in value.items():
print(str(skey) + ": " + str(svalue) + ", ")
source = RaplPowerSource()
# output_to_terminal(source)
s = socket.socket()
host = socket.gethostname()
port = 8888
s.bind((host, port))
s.listen(5)
print("等待客户端连接...")
while True:
c, addr = s.accept()
source.update()
summary = dict(source.get_sensors_summary())
#msg = json.dumps(summary)
# package表示CPU,dram表示内存(一般不准)
power_total = str(sum(list(map(float, [summary[key] for key in summary.keys() if key.startswith('package')]))))
print(f'发送给{addr}:{power_total}')
c.send(power_total.encode('utf-8'))
c.close() # 关闭连接
因此,在需要获取CPU功耗时候,只需要:
s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
host = socket.gethostname()
port = 8888
s.connect((host, port))
msg = s.recv(1024)
s.close()
power_usage_cpu = float(msg.decode('utf-8'))
提供一个我自己编写和使用的功耗分析代码,仅供参考。(注意上面的power_listener.py需要运行着)
import cv2
import socket
import sys
import threading
import json
import statistics
from psutil import _common as common
import pynvml
pynvml.nvmlInit()
class Timer:
def __init__(self, name = '', is_verbose = False):
self._name = name
self._is_verbose = is_verbose
self._is_paused = False
self._start_time = None
self._accumulated = 0
self._elapsed = 0
self.start()
def start(self):
self._accumulated = 0
self._start_time = cv2.getTickCount()
def pause(self):
now_time = cv2.getTickCount()
self._accumulated += (now_time - self._start_time)/cv2.getTickFrequency()
self._is_paused = True
def resume(self):
if self._is_paused: # considered only if paused
self._start_time = cv2.getTickCount()
self._is_paused = False
def elapsed(self):
if self._is_paused:
self._elapsed = self._accumulated
else:
now = cv2.getTickCount()
self._elapsed = self._accumulated + (now - self._start_time)/cv2.getTickFrequency()
if self._is_verbose is True:
name = self._name
if self._is_paused:
name += ' [paused]'
message = 'Timer::' + name + ' - elapsed: ' + str(self._elapsed)
timer_print(message)
return self._elapsed
class PowerUsage:
'''
demo:
power_usage = PowerUsage()
power_usage.analyze_start()
time.sleep(2)
time_used, power_usage_gpu, power_usage_cpu = power_usage.analyze_end()
print(time_used)
print(power_usage_gpu)
print(power_usage_cpu)
'''
def __init__(self):
self.start_analyze = False
self.power_usage_gpu_values = list()
self.power_usage_cpu_values = list()
self.thread = None
self.timer = Timer(name='GpuPowerUsage', is_verbose=False)
def analyze_start(self, gpu_id=0, delay=0.1):
handle = pynvml.nvmlDeviceGetHandleByIndex(gpu_id)
def start():
self.power_usage_gpu_values.clear()
self.power_usage_cpu_values.clear()
self.start_analyze = True
self.timer.start()
while self.start_analyze:
powerusage = pynvml.nvmlDeviceGetPowerUsage(handle)
self.power_usage_gpu_values.append(powerusage/1000)
s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
host = socket.gethostname()
port = 8888
s.connect((host, port))
msg = s.recv(1024)
s.close()
self.power_usage_cpu_values.append(float(msg.decode('utf-8')))
time.sleep(delay)
self.thread = threading.Thread(target=start, daemon=True)
self.thread.start()
def analyze_end(self, mean=True):
self.start_analyze = False
while self.thread and self.thread.isAlive():
time.sleep(0.01)
time_used = self.timer.elapsed()
self.thread = None
power_usage_gpu = statistics.mean(self.power_usage_gpu_values) if mean else self.power_usage_gpu_values
power_usage_cpu = statistics.mean(self.power_usage_cpu_values) if mean else self.power_usage_cpu_values
return time_used, power_usage_gpu, power_usage_cpu
power_usage = PowerUsage()
def power_usage_api(func, note=''):
@wraps(func)
def wrapper(*args, **kwargs):
power_usage.analyze_start()
result = func(*args, **kwargs)
print(f'{note}{power_usage.analyze_end()}')
return result
return wrapper
def power_usage_api2(note=''):
def decorator(func):
@wraps(func)
def wrapper(*args, **kwargs):
power_usage.analyze_start()
result = func(*args, **kwargs)
print(f'{note}{power_usage.analyze_end()}')
return result
return wrapper
return decorator
用法示例:
power_usage = PowerUsage()
power_usage.analyze_start()
# ----------------------
# xxx 某一段待分析的代码
# 这里以sleep表示运行时长
time.sleep(2)
# ----------------------
time_used, power_usage_gpu, power_usage_cpu = power_usage.analyze_end()
print(f'time_used: {time_used}')
print(f'power_usage_gpu: {power_usage_gpu}')
print(f'power_usage_cpu: {power_usage_cpu}')