背景
在人工智能领域，GPU 是必不可少的。在本文中，我们将介绍如何在服务器上安装和部署 NVIDIA GPU 的 docker
升级你的CUDA
选择你的系统对应版本进行安装
安装
- 确保你的系统已经安装了NVIDIA驱动和Docker引擎。确保驱动版本与Docker引擎兼容。你还需要安装nvidia-docker2软件包，它是NVIDIA Docker的一个插件。可以在https://github.com/NVIDIA/nvidia-docker上找到安装说明
1.1 卸载旧版本的nvidia-docker(如果已安装)
1sudo apt-get remove nvidia-docker
1.2 添加nvidia-docker2仓库
1distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
2curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add -
3curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list
1.3 更新软件包列表和安装nvidia-docker2
1sudo apt-get update
2sudo apt-get install nvidia-docker2
1.4 重启Docker守护进程
1sudo systemctl restart docker
- 通过以下命令测试是否安装成功
1docker run --gpus all nvidia/cuda:10.0-base nvidia-smi
- 如果安装成功，你将看到类似下面的输出
1+-----------------------------------------------------------------------------+
2| NVIDIA-SMI 440.33.01 Driver Version: 440.33.01 CUDA Version: 10.2 |
3|-------------------------------+----------------------+----------------------+
4| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
5| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
6|===============================+======================+======================|
7| 0 Tesla P100-PCIE... On | 00000000:00:1E.0 Off | 0 |
8| N/A 32C P0 32W / 250W | 0MiB / 16280MiB | 0% Default |
9+-------------------------------+----------------------+----------------------+
10| 1 Tesla P100-PCIE... On | 00000000:00:1F.0 Off | 0 |
11| N/A 32C P0 32W / 250W | 0MiB / 16280MiB | 0% Default |
12+-------------------------------+----------------------+----------------------+
13| 2 Tesla P100-PCIE... On | 00000000:00:20.0 Off | 0 |
14| N/A 32C P0 32W / 250W | 0MiB / 16280MiB | 0% Default |
15+-------------------------------+----------------------+----------------------+
16| 3 Tesla P100-PCIE... On | 00000000:00:21.0 Off | 0 |
17| N/A 32C P0 32W / 250W | 0MiB / 16280MiB | 0% Default |
18+-------------------------------+----------------------+----------------------+
19| 4 Tesla P100-PCIE... On | 00000000:00:22.0 Off | 0 |
20| N/A 32C P0 32W / 250W | 0MiB / 16280MiB | 0% Default |
21+-------------------------------+----------------------+----------------------+
22| 5 Tesla P100-PCIE... On | 00000000:00:23.0 Off | 0 |
23| N/A 32C P0 32W / 250W | 0MiB / 16280MiB | 0% Default |
24+-------------------------------+----------------------+----------------------+
25| 6 Tesla P100-PCIE... On | 00000000:00:24.0 Off | 0 |
26| N/A 32C P0 32W / 250W | 0MiB / 16280MiB | 0% Default |
27+-------------------------------+----------------------+----------------------+
28| 7 Tesla P100-PCIE... On | 00000000:00:25.0 Off | 0 |
29| N/A 32C P0 32W / 250W | 0MiB / 16280MiB | 0% Default |
30+-------------------------------+----------------------+----------------------+
部署
- 创建一个新的dockerfile
1FROM nvidia/cuda:11.7.0-cudnn8-runtime-ubuntu18.04
2
3# 安装软件包
4RUN apt-get update && apt-get install -y \
5 build-essential \
6 git \
7 wget \
8 vim \
9 openssh-server
10
11RUN mkdir -p /run/sshd
12RUN echo "root:password" | chpasswd
13RUN sed -i 's/#PermitRootLogin prohibit-password/PermitRootLogin yes/' /etc/ssh/sshd_config
14EXPOSE 22
15CMD ["/usr/sbin/sshd", "-D"]
- 创建一个新的docker-compose.yml
1version: "3"
2services:
3 cuda-container:
4 build: .
5 runtime: nvidia
6 container_name: cuda-container
7 ports:
8 - 2201:22
9 volumes:
10 - ./app:/app
- 启动docker
1docker-compose up -d
- 验证
1docker exec -ti cuda-container nvidia-smi
