文章/答案/技术大牛

发布

社区首页 >专栏 >nvidia-smi详解（二）- python

nvidia-smi详解（二）- python

原创

何其不顾四月天

发布于 2024-07-18 17:32:59

50800

代码可运行

文章被收录于专栏：四月天的专栏四月天的专栏

运行总次数：0

代码可运行

Nvidia-smi and Python

前言

在之前的nvidia-smi 详解（一)写的过程中，查资料查到了有对应的python支持方法，就计划写这个了，随后写加查资料就写好代码了，但是就是犯懒一直没写文章，墨迹到了现在。

也是做了一些简单的接口测试以及使用，主要还是查询的接口，没测试控制类接口。

简介

对应的py库介绍主要是来自： nvidia-ml-py。

Provides a Python interface to GPU management and monitoring functions.

This is a wrapper around the NVML library. For information about the NVML library, see the NVML developer page http://developer.nvidia.com/nvidia-management-library-nvml

Download the latest package from: http://pypi.python.org/pypi/nvidia-ml-py/

Note this file can be run with 'python -m doctest -v README.txt' although the results are system dependent

The nvml header file contains function documentation that is relevant to this wrapper. The header file is distributed with. https://developer.nvidia.com/gpu-deployment-kit

The main difference is this library handles allocating structs and passing pointers to the functions, before returning the desired value. Non-success return codes are raised as exceptions as described in the section below.

安装

pip安装

python3 -m pip install nvidia-ml-py

安装包安装

下载地址：nvidia-ml-py 12.555.43

wget https://files.pythonhosted.org/packages/ee/c6/2348fc1fb776ff41a34635fb1f18010a6d6fd7ba6e57184dabd9d98ba9cf/nvidia-ml-py-12.555.43.tar.gz
tar -xzf nvidia-ml-py-$major-$minor-$patch.tar.gz`
cd nvidia-ml-py-$major-$minor-$patch
sudo python setup.py install

注意：$major-$minor-$patch 这三个变量的替换

代码

导入

from pynvml import *

初始化

nvmlInit()

版本查询

print(f"Driver Version: {nvmlSystemGetDriverVersion()}")

输出结果: Driver Version: 552.22

GPU个数获取

deviceCount = nvmlDeviceGetCount()

获取对应索引GPU句柄

handle = nvmlDeviceGetHandleByIndex(i)
print(f"Device {i} : {nvmlDeviceGetName(handle)}")

输出结果：

Device 0 : NVIDIA GeForce RTX 4060 Ti

获取显存信息

info = nvmlDeviceGetMemoryInfo(handle)
print(f"Total memory: {info.total} MiB")
print(f"Free memory: {info.free} MiB")
print(f"Used memory: {info.used} MiB")

输出结果：

Total memory: 8585740288 MiB
Free memory: 6701080576 MiB
Used memory: 1884659712 MiB

获取显存利用率

info = nvmlDeviceGetUtilizationRates(handle)
print(f"UtilizationRates Gpu: {info.gpu}%")
print(f"UtilizationRates Memory: {info.memory}%")

输出结果：

UtilizationRates Gpu: 1%
UtilizationRates Memory: 17%

获取编码利用率信息

info = nvmlDeviceGetEncoderUtilization(handle)
print(f"EncoderUtilization {info}")

输出结果：

EncoderUtilization [0, 200000]

获取解码利用率信息

info = nvmlDeviceGetDecoderUtilization(handle)
print(f"DecoderUtilization {info}")

输出结果：

DecoderUtilization [0, 200000]

获取JPG利用信息

info = nvmlDeviceGetJpgUtilization(handle)
print(f"JpgUtilization {info}")

输出结果：

JpgUtilization [0, 200000]

通过Pid获取进程名字

def get_process_name_by_pid(pid):
    try:
        return psutil.Process(pid).name()
    except:
        return None

获取正在运行计算操作的进程列表

info = nvmlDeviceGetComputeRunningProcesses(handle)
for index,item in enumerate(info):
    print(f"{index} : {item} ,name: {get_process_name_by_pid(item.pid)}")

输出信息(只放部分进程)：

0 : {'pid': 1716, 'usedGpuMemory': None, 'gpuInstanceId': 4294967295, 'computeInstanceId': 4294967295} ,name: dwm.exe
1 : {'pid': 6384, 'usedGpuMemory': None, 'gpuInstanceId': 4294967295, 'computeInstanceId': 4294967295} ,name: explorer.exe
2 : {'pid': 6592, 'usedGpuMemory': None, 'gpuInstanceId': 4294967295, 'computeInstanceId': 4294967295} ,name: SearchHost.exe
3 : {'pid': 7064, 'usedGpuMemory': None, 'gpuInstanceId': 4294967295, 'computeInstanceId': 4294967295} ,name: StartMenuExperienceHost.exe

获取正在运行的图形进程列表

info = nvmlDeviceGetGraphicsRunningProcesses(handle)
for index,item in enumerate(info):
    print(f"{index} : {item} ,name: {get_process_name_by_pid(item.pid)}")

输出信息(只放部分输出结果);

0 : {'pid': 1716, 'usedGpuMemory': None, 'gpuInstanceId': 4294967295, 'computeInstanceId': 4294967295} ,name: dwm.exe
1 : {'pid': 6384, 'usedGpuMemory': None, 'gpuInstanceId': 4294967295, 'computeInstanceId': 4294967295} ,name: explorer.exe
2 : {'pid': 6592, 'usedGpuMemory': None, 'gpuInstanceId': 4294967295, 'computeInstanceId': 4294967295} ,name: SearchHost.exe
3 : {'pid': 7064, 'usedGpuMemory': None, 'gpuInstanceId': 4294967295, 'computeInstanceId': 4294967295} ,name: StartMenuExperienceHost.exe

获取GPU的当前温度

temperature_gpu = nvmlDeviceGetTemperature(handle, NVML_TEMPERATURE_GPU)
print(f"temperature_gpu {temperature_gpu}")

输出结果:

temperature_gpu 29

获取GPU的上限温度

temperature_threshold = nvmlDeviceGetTemperatureThreshold(handle, NVML_TEMPERATURE_THRESHOLD_GPU_MAX)
print(f"temperature_threshold {temperature_threshold}")

输出结果：

temperature_threshold 90

获取GPU设备的UUID

gpu_uuid = nvmlDeviceGetUUID(handle)
print(f"gpu_uuid {gpu_uuid}")

输出结果：

gpu_uuid GPU-3fd9292f-3024-fbdb-4596-5c5560b91654

获取GPU设备的PCI信息

pci_info = nvmlDeviceGetPciInfo(handle)
print(f"pci info {pci_info}")

获取的其他信息 CLOCK相关

print(f"GRAPHICS clock info {nvmlDeviceGetClockInfo(handle, NVML_CLOCK_GRAPHICS)}")
print(f"SM clock info {nvmlDeviceGetClockInfo(handle, NVML_CLOCK_SM)}")
print(f"MEM clock info {nvmlDeviceGetClockInfo(handle, NVML_CLOCK_MEM)}")
print(f"VIDEO clock info {nvmlDeviceGetClockInfo(handle, NVML_CLOCK_VIDEO)}")

输出结果:

GRAPHICS clock info 270
SM clock info 270
MEM clock info 405
VIDEO clock info 1185

获取电源相关信息

print(f"power state {nvmlDeviceGetPowerState(handle)}")
print(f"power usage watts {nvmlDeviceGetPowerUsage(handle) / 1000}")

输出结果:

power state 8
power usage percent 27.649

结束调用

nvmlShutdown()

全部代码

主要是遍历输出所有Nvidia状态信息，可以根据自己需要来进行编写设置。

from pynvml import *
import psutil

def get_process_name_by_pid(pid):
    try:
        return psutil.Process(pid).name()
    except:
        return None

nvmlInit()
print(f"Driver Version: {nvmlSystemGetDriverVersion()}")

deviceCount = nvmlDeviceGetCount()
try:
    for i in range(deviceCount):
        handle = nvmlDeviceGetHandleByIndex(i)
        print(f"Device {i} : {nvmlDeviceGetName(handle)}")
        print(" ")
        try:
            (current, pending) = nvmlDeviceGetEccMode(handle)
            print("current",current,"pending",pending)
        except NVMLError as error:
            print(error)
        
        print(" ")
        info = nvmlDeviceGetMemoryInfo(handle)
        print(f"Total memory: {info.total} MiB")
        print(f"Free memory: {info.free} MiB")
        print(f"Used memory: {info.used} MiB")

        print(" ")
        info = nvmlDeviceGetUtilizationRates(handle)
        print(f"UtilizationRates Gpu: {info.gpu}%")
        print(f"UtilizationRates Memory: {info.memory}%")

        print(" ")
        info = nvmlDeviceGetEncoderUtilization(handle)
        print(f"EncoderUtilization {info}")

        print(" ")
        info = nvmlDeviceGetDecoderUtilization(handle)
        print(f"DecoderUtilization {info}")

        print(" ")
        info = nvmlDeviceGetJpgUtilization(handle)
        print(f"JpgUtilization {info}")

        print(" ")
        info = nvmlDeviceGetOfaUtilization(handle)
        print(f"OfaUtilization {info}")

        print(" ")
        info = nvmlDeviceGetComputeRunningProcesses(handle)
        for index,item in enumerate(info):
            print(f"{index} : {item} ,name: {get_process_name_by_pid(item.pid)}")

        print(" ")
        info = nvmlDeviceGetGraphicsRunningProcesses(handle)
        for index,item in enumerate(info):
            print(f"{index} : {item} ,name: {get_process_name_by_pid(item.pid)}")

        print(" ")
        try:
            # NVML_PROCESS_MODE_COMPUTE | NVML_PROCESS_MODE_GRAPHICS | NVML_PROCESS_MODE_MPS
            info = nvmlDeviceGetRunningProcessDetailList(handle, nvmlProcessDetailList_v1, NVML_PROCESS_MODE_COMPUTE)
            for item in info:
                print(f"ProcessDetailList {item}")
        except NVMLError as error:
            print(f"nvmlDeviceGetRunningProcessDetailList {error}")

        temperature_gpu = nvmlDeviceGetTemperature(handle, NVML_TEMPERATURE_GPU)
        print(f"temperature_gpu {temperature_gpu}")

        temperature_threshold = nvmlDeviceGetTemperatureThreshold(handle, NVML_TEMPERATURE_THRESHOLD_GPU_MAX)
        print(f"temperature_threshold {temperature_threshold}")

        print("")
        gpu_uuid = nvmlDeviceGetUUID(handle)
        print(f"gpu_uuid {gpu_uuid}")

        print("")
        pci_info = nvmlDeviceGetPciInfo(handle)
        print(f"pci info {pci_info}")

        print("")
        print(f"GRAPHICS clock info {nvmlDeviceGetClockInfo(handle, NVML_CLOCK_GRAPHICS)}")
        print(f"SM clock info {nvmlDeviceGetClockInfo(handle, NVML_CLOCK_SM)}")
        print(f"MEM clock info {nvmlDeviceGetClockInfo(handle, NVML_CLOCK_MEM)}")
        print(f"VIDEO clock info {nvmlDeviceGetClockInfo(handle, NVML_CLOCK_VIDEO)}")

        print("")
        print(f"power state {nvmlDeviceGetPowerState(handle)}")
        print(f"power usage watts {nvmlDeviceGetPowerUsage(handle) / 1000}")
        
except NVMLError as error:
    print(error)

nvmlShutdown()

结束语

其实写的比较简单，只是摘取了我觉得日常会用的一些信息，还有很多其他函数，可以在pynvml.py文件里边查看获取。或者 NVIDIA GPU等网站查看自己需要的一些信息，样例等。这半年其实一直没怎写，日常工作也比较无聊，开始研究srs,以及一些pyton脚本的编写，希望把文章抓起来，多写一些。

原创声明：本文系作者授权腾讯云开发者社区发表，未经许可，不得转载。

如有侵权，请联系 cloudcommunity@tencent.com 删除。

python