py3nvml實(shí)現(xiàn)GPU相關(guān)信息讀取的案例分析

更新時(shí)間：2022年01月16日 12:08:04 作者：DECHIN

這篇文章主要介紹了py3nvml實(shí)現(xiàn)GPU相關(guān)信息讀取,此時(shí)就可以考慮使用py3nvml這樣的工具，針對(duì)于GPU任務(wù)執(zhí)行的過程進(jìn)行細(xì)化的分析，有助于提升GPU的利用率和程序執(zhí)行的性能，需要的朋友可以參考下

在深度學(xué)習(xí)或者其他類型的GPU運(yùn)算過程中，對(duì)于GPU信息的監(jiān)測也是一個(gè)非常常用的功能。如果僅僅是使用系統(tǒng)級(jí)的GPU監(jiān)測工具，就沒辦法非常細(xì)致的去跟蹤每一步的顯存和使用率的變化。如果是用profiler，又顯得過于細(xì)致，而且環(huán)境配置、信息輸出和篩選并不是很方便。此時(shí)就可以考慮使用py3nvml這樣的工具，針對(duì)于GPU任務(wù)執(zhí)行的過程進(jìn)行細(xì)化的分析，有助于提升GPU的利用率和程序執(zhí)行的性能。

技術(shù)背景

隨著模型運(yùn)算量的增長和硬件技術(shù)的發(fā)展，使用GPU來完成各種任務(wù)的計(jì)算已經(jīng)漸漸成為算法實(shí)現(xiàn)的主流手段。而對(duì)于運(yùn)行期間的一些GPU的占用，比如每一步的顯存使用率等諸如此類的信息，就需要一些比較細(xì)致的GPU信息讀取的工具，這里我們重點(diǎn)推薦使用py3nvml來對(duì)python代碼運(yùn)行的一個(gè)過程進(jìn)行監(jiān)控。

常規(guī)信息讀取

一般大家比較常用的就是nvidia-smi這個(gè)指令，來讀取GPU的使用率和顯存占用、驅(qū)動(dòng)版本等信息：

$ nvidia-smi
Wed Jan 12 15:52:04 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.42.01    Driver Version: 470.42.01    CUDA Version: 11.4     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Quadro RTX 4000     On   | 00000000:03:00.0  On |                  N/A |
| 30%   39C    P8    20W / 125W |    538MiB /  7979MiB |     16%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   1  Quadro RTX 4000     On   | 00000000:A6:00.0 Off |                  N/A |
| 30%   32C    P8     7W / 125W |      6MiB /  7982MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A      1643      G   /usr/lib/xorg/Xorg                412MiB |
|    0   N/A  N/A      2940      G   /usr/bin/gnome-shell               76MiB |
|    0   N/A  N/A     47102      G   ...AAAAAAAAA= --shared-files       35MiB |
|    0   N/A  N/A    172424      G   ...AAAAAAAAA= --shared-files       11MiB |
|    1   N/A  N/A      1643      G   /usr/lib/xorg/Xorg                  4MiB |
+-----------------------------------------------------------------------------+

但是如果不使用profile僅僅使用nvidia-smi這個(gè)指令的輸出的話，是沒有辦法非常細(xì)致的分析程序運(yùn)行過程中的變化的。這里順便推薦一個(gè)比較精致的跟nvidia-smi用法非常類似的小工具：gpustat。這個(gè)工具可以直接使用pip進(jìn)行安裝和管理：

$ python3 -m pip install gpustat
Collecting gpustat
  Downloading gpustat-0.6.0.tar.gz (78 kB)
     |████████████████████████████████| 78 kB 686 kB/s
Requirement already satisfied: six>=1.7 in /home/dechin/.local/lib/python3.8/site-packages (from gpustat) (1.16.0)
Collecting nvidia-ml-py3>=7.352.0
  Downloading nvidia-ml-py3-7.352.0.tar.gz (19 kB)
Requirement already satisfied: psutil in /home/dechin/.local/lib/python3.8/site-packages (from gpustat) (5.8.0)
Collecting blessings>=1.6
  Downloading blessings-1.7-py3-none-any.whl (18 kB)
Building wheels for collected packages: gpustat, nvidia-ml-py3
  Building wheel for gpustat (setup.py) ... done
  Created wheel for gpustat: filename=gpustat-0.6.0-py3-none-any.whl size=12617 sha256=4158e741b609c7a1bc6db07d76224db51cd7656a6f2e146e0b81185ce4e960ba
  Stored in directory: /home/dechin/.cache/pip/wheels/0d/d9/80/b6cbcdc9946c7b50ce35441cc9e7d8c5a9d066469ba99bae44
  Building wheel for nvidia-ml-py3 (setup.py) ... done
  Created wheel for nvidia-ml-py3: filename=nvidia_ml_py3-7.352.0-py3-none-any.whl size=19191 sha256=70cd8ffc92286944ad9f5dc4053709af76fc0e79928dc61b98a9819a719f1e31
  Stored in directory: /home/dechin/.cache/pip/wheels/b9/b1/68/cb4feab29709d4155310d29a421389665dcab9eb3b679b527b
Successfully built gpustat nvidia-ml-py3
Installing collected packages: nvidia-ml-py3, blessings, gpustat
Successfully installed blessings-1.7 gpustat-0.6.0 nvidia-ml-py3-7.352.0

使用的時(shí)候也是跟nvidia-smi非常類似的操作：

$ watch --color -n1 gpustat -cpu

返回結(jié)果如下所示：

Every 1.0s: gpustat -cpu ubuntu2004: Wed Jan 12 15:58:59 2022

ubuntu2004 Wed Jan 12 15:58:59 2022 470.42.01
[0] Quadro RTX 4000 | 39'C, 3 % | 537 / 7979 MB | root:Xorg/1643(412M) de
chin:gnome-shell/2940(75M) dechin:slack/47102(35M) dechin:chrome/172424(11M)
[1] Quadro RTX 4000 | 32'C, 0 % | 6 / 7982 MB | root:Xorg/1643(4M)

通過gpustat返回的結(jié)果，包含了GPU的型號(hào)、使用率和顯存使用大小和GPU當(dāng)前的溫度等常規(guī)信息。

py3nvml的安裝與使用

接下來正式看下py3nvml的安裝和使用方法，這是一個(gè)可以在python中實(shí)時(shí)查看和監(jiān)測GPU信息的一個(gè)庫，可以通過pip來安裝和管理：

$ python3 -m pip install py3nvml
Collecting py3nvml
  Downloading py3nvml-0.2.7-py3-none-any.whl (55 kB)
     |████████████████████████████████| 55 kB 650 kB/s
Requirement already satisfied: xmltodict in /home/dechin/anaconda3/lib/python3.8/site-packages (from py3nvml) (0.12.0)
Installing collected packages: py3nvml
Successfully installed py3nvml-0.2.7

py3nvml綁定GPU卡

有一些框架為了性能的最大化，在初始化的時(shí)候就會(huì)默認(rèn)去使用到整個(gè)資源池里面的所有GPU卡，比如如下使用Jax來演示的一個(gè)案例：

In [1]: import py3nvml

In [2]: from jax import numpy as jnp

In [3]: x = jnp.ones(1000000000)

In [4]: !nvidia-smi
Wed Jan 12 16:08:32 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.42.01    Driver Version: 470.42.01    CUDA Version: 11.4     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Quadro RTX 4000     On   | 00000000:03:00.0  On |                  N/A |
| 30%   41C    P0    38W / 125W |   7245MiB /  7979MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   1  Quadro RTX 4000     On   | 00000000:A6:00.0 Off |                  N/A |
| 30%   35C    P0    35W / 125W |    101MiB /  7982MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A      1643      G   /usr/lib/xorg/Xorg                412MiB |
|    0   N/A  N/A      2940      G   /usr/bin/gnome-shell               75MiB |
|    0   N/A  N/A     47102      G   ...AAAAAAAAA= --shared-files       35MiB |
|    0   N/A  N/A    172424      G   ...AAAAAAAAA= --shared-files       11MiB |
|    0   N/A  N/A    812125      C   /usr/local/bin/python            6705MiB |
|    1   N/A  N/A      1643      G   /usr/lib/xorg/Xorg                  4MiB |
|    1   N/A  N/A    812125      C   /usr/local/bin/python              93MiB |
+-----------------------------------------------------------------------------+

在這個(gè)案例中我們只是在顯存中分配了一塊空間用于存儲(chǔ)一個(gè)向量，但是Jax在初始化之后，自動(dòng)占據(jù)了本地的2張GPU卡。根據(jù)Jax官方提供的方法，我們可以使用如下的操作配置環(huán)境變量，使得Jax只能看到其中的1張卡，這樣就不會(huì)擴(kuò)張：

In [1]: import os

In [2]: os.environ["CUDA_VISIBLE_DEVICES"] = "1"

In [3]: from jax import numpy as jnp

In [4]: x = jnp.ones(1000000000)

In [5]: !nvidia-smi
Wed Jan 12 16:10:36 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.42.01    Driver Version: 470.42.01    CUDA Version: 11.4     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Quadro RTX 4000     On   | 00000000:03:00.0  On |                  N/A |
| 30%   40C    P8    19W / 125W |    537MiB /  7979MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   1  Quadro RTX 4000     On   | 00000000:A6:00.0 Off |                  N/A |
| 30%   35C    P0    35W / 125W |   7195MiB /  7982MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A      1643      G   /usr/lib/xorg/Xorg                412MiB |
|    0   N/A  N/A      2940      G   /usr/bin/gnome-shell               75MiB |
|    0   N/A  N/A     47102      G   ...AAAAAAAAA= --shared-files       35MiB |
|    0   N/A  N/A    172424      G   ...AAAAAAAAA= --shared-files       11MiB |
|    1   N/A  N/A      1643      G   /usr/lib/xorg/Xorg                  4MiB |
|    1   N/A  N/A    813030      C   /usr/local/bin/python            7187MiB |
+-----------------------------------------------------------------------------+

可以看到結(jié)果中已經(jīng)是只使用了1張GPU卡，達(dá)到了我們的目的，但是這種通過配置環(huán)境變量來實(shí)現(xiàn)的功能還是著實(shí)不夠pythonic，因此py3nvml中也提供了這樣的功能，可以指定某一系列的GPU卡用于執(zhí)行任務(wù)：

In [1]: import py3nvml

In [2]: from jax import numpy as jnp

In [3]: py3nvml.grab_gpus(num_gpus=1,gpu_select=[1])
Out[3]: 1

In [4]: x = jnp.ones(1000000000)

In [5]: !nvidia-smi
Wed Jan 12 16:12:37 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.42.01    Driver Version: 470.42.01    CUDA Version: 11.4     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Quadro RTX 4000     On   | 00000000:03:00.0  On |                  N/A |
| 30%   40C    P8    20W / 125W |    537MiB /  7979MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   1  Quadro RTX 4000     On   | 00000000:A6:00.0 Off |                  N/A |
| 30%   36C    P0    35W / 125W |   7195MiB /  7982MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A      1643      G   /usr/lib/xorg/Xorg                412MiB |
|    0   N/A  N/A      2940      G   /usr/bin/gnome-shell               75MiB |
|    0   N/A  N/A     47102      G   ...AAAAAAAAA= --shared-files       35MiB |
|    0   N/A  N/A    172424      G   ...AAAAAAAAA= --shared-files       11MiB |
|    1   N/A  N/A      1643      G   /usr/lib/xorg/Xorg                  4MiB |
|    1   N/A  N/A    814673      C   /usr/local/bin/python            7187MiB |
+-----------------------------------------------------------------------------+

可以看到結(jié)果中也是只使用了1張GPU卡，達(dá)到了跟上一步的操作一樣的效果。

查看空閑GPU

對(duì)于環(huán)境中可用的GPU，py3nvml的判斷標(biāo)準(zhǔn)就是在這個(gè)GPU上已經(jīng)沒有任何的進(jìn)程，那么這個(gè)就是一張可用的GPU卡：

In [1]: import py3nvml

In [2]: free_gpus = py3nvml.get_free_gpus()

In [3]: free_gpus
Out[3]: [True, True]

當(dāng)然這里需要說明的是，系統(tǒng)應(yīng)用在這里不會(huì)被識(shí)別，應(yīng)該是會(huì)判斷守護(hù)進(jìn)程。

命令行信息獲取

跟nvidia-smi非常類似的，py3nvml也可以在命令行中通過調(diào)用py3smi來使用。值得一提的是，如果需要用nvidia-smi來實(shí)時(shí)的監(jiān)測GPU的使用信息，往往是需要配合watch -n來使用的，但是如果是py3smi則不需要，直接用py3smi -l就可以實(shí)現(xiàn)類似的功能。

$ py3smi -l 5
Wed Jan 12 16:17:37 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI                        Driver Version: 470.42.01                 |
+---------------------------------+---------------------+---------------------+
| GPU Fan  Temp Perf Pwr:Usage/Cap|        Memory-Usage | GPU-Util Compute M. |
+=================================+=====================+=====================+
|   0 30%   39C    8   19W / 125W |   537MiB /  7979MiB |       0%    Default |
|   1 30%   33C    8    7W / 125W |     6MiB /  7982MiB |       0%    Default |
+---------------------------------+---------------------+---------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
| GPU        Owner      PID      Uptime  Process Name                   Usage |
+=============================================================================+
+-----------------------------------------------------------------------------+

可以看到略有區(qū)別的是，這里并不像nvidia-smi列出來的進(jìn)程那么多，應(yīng)該是自動(dòng)忽略了系統(tǒng)進(jìn)程。

單獨(dú)查看驅(qū)動(dòng)版本和顯卡型號(hào)

在py3nvml中把查看驅(qū)動(dòng)和型號(hào)的功能單獨(dú)列了出來：

In [1]: from py3nvml.py3nvml import *
In [2]: nvmlInit()
Out[2]: <CDLL 'libnvidia-ml.so.1', handle 560ad4d07a60 at 0x7fd13aa52340>
In [3]: print("Driver Version: {}".format(nvmlSystemGetDriverVersion()))
Driver Version: 470.42.01
In [4]: deviceCount = nvmlDeviceGetCount()
   ...: for i in range(deviceCount):
   ...:     handle = nvmlDeviceGetHandleByIndex(i)
   ...:     print("Device {}: {}".format(i, nvmlDeviceGetName(handle)))
   ...:
Device 0: Quadro RTX 4000
Device 1: Quadro RTX 4000
In [5]: nvmlShutdown()

這樣也不需要我們自己再去逐個(gè)的篩選，從靈活性和可擴(kuò)展性上來說還是比較方便的。

單獨(dú)查看顯存信息

這里同樣的也是把顯存的使用信息單獨(dú)列了出來，不需要用戶再去單獨(dú)篩選這個(gè)信息，相對(duì)而言比較細(xì)致：

In [1]: from py3nvml.py3nvml import *
In [2]: nvmlInit()
Out[2]: <CDLL 'libnvidia-ml.so.1', handle 55ae42aadd90 at 0x7f39c700e040>
In [3]: handle = nvmlDeviceGetHandleByIndex(0)
In [4]: info = nvmlDeviceGetMemoryInfo(handle)
In [5]: print("Total memory: {}MiB".format(info.total >> 20))
Total memory: 7979MiB
In [6]: print("Free memory: {}MiB".format(info.free >> 20))
Free memory: 7441MiB
In [7]: print("Used memory: {}MiB".format(info.used >> 20))
Used memory: 537MiB

如果把這些代碼插入到程序中，就可以獲悉每一步所占用的顯存的變化。