Question

我正在尝试计算用Python编写的tensorRT脚本的功耗。有了CUDA，nvprof cuda_script就可以了，但是我的Python脚本找不到类似的东西。

Python是否有类似的东西？人们如何计算p100 / v100的功耗？

适用于Nvidia GPU的东西-How to profile CPU usage of a Python script?

Answer 1

事实证明，以线程方式在一个python文件中运行此过程非常复杂。解决方案是与bash同时运行它们，然后使power脚本检查其他进程是否正在运行。示例代码如下：

def measurePower():

    tmp = os.popen("ps -Af").read()
    process_name = "PROCESSTOBEMEASURED.PY"
    power_measurement = []

    '''One checks if the other process is still running, if yes it measures power again.'''
    while(process_name in tmp[:]):
        measurementPower = os.popen("nvidia-smi -i 0 -q").read()
        tmp = os.popen("ps -Af").read()
        power_measurement.append(nvidiaSmiParser(measurementPower, ["Power Draw"],num))

    return power_measurement

TensorFlow：使用Python确定tensorRT功耗

1 个答案: