1. 安装驱动
1.1 查看系统是否识别显卡
1
2
3
4
| lspci | grep -i vga
03:00.0 VGA compatible controller: NVIDIA Corporation GP102 [TITAN X] (rev a1)
0a:00.0 VGA compatible controller: Matrox Electronics Systems Ltd. G200eR2 (rev 01)
|
识别出显卡为 NVIDIA 的 TITAN X。
1.2 禁用 nouveau
如果有输出,说明 nouveau 已经加载,需要禁用。如果没有输出,则可以跳过此操作。
- 关闭自动更新
1
| sed -i.bak 's/1/0/' /etc/apt/apt.conf.d/10periodic
|
编辑配置文件:
1
| vim /etc/apt/apt.conf.d/50unattended-upgrades
|
去掉以下内容的注释
1
2
3
4
| Unattended-Upgrade::Package-Blacklist {
"linux-image-*";
"linux-headers-*";
};
|
- 编辑系统 blacklist
1
| vim /etc/modprobe.d/blacklist-nouveau.conf
|
添加以下配置禁用 nouveau
1
2
| blacklist nouveau
options nouveau modeset=0
|
- 更新 initramfs
- 重启系统
- 编辑系统 blacklist
1
| vim /etc/modprobe.d/blacklist-nouveau.conf
|
添加配置禁用 nouveau
1
2
| blacklist nouveau
options nouveau modeset=0
|
- 更新 initramfs
1
2
| mv /boot/initramfs-$(uname -r).img /boot/initramfs-$(uname -r).img.bak
dracut /boot/initramfs-$(uname -r).img $(uname -r)
|
- 重启系统
此时不应该有输出。
1.3 安装驱动
1
2
| apt install build-essential gcc-9 g++-9
update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-9 90 --slave /usr/bin/g++ g++ /usr/bin/g++-9 --slave /usr/bin/gcov gcov /usr/bin/gcov-9
|
访问 https://www.nvidia.cn/Download/index.aspx 选择对应的驱动版本下载。这里以 Linux 64-bit 的 TITAN X 驱动为例:
1
| wget https://cn.download.nvidia.com/XFree86/Linux-x86_64/535.146.02/NVIDIA-Linux-x86_64-535.146.02.run
|
1
2
| chmod +x NVIDIA-Linux-x86_64-535.146.02.run
./NVIDIA-Linux-x86_64-535.146.02.run
|
2. 安装 nvidia-container-runtime
2.1 安装 Containerd
1
2
| apt-get update
apt-get install -y ca-certificates curl gnupg lsb-release
|
1
2
| mkdir -p /etc/apt/keyrings
curl -fsSL https://download.docker.com/linux/ubuntu/gpg | gpg --dearmor -o /etc/apt/keyrings/docker.gpg
|
1
2
3
| echo \
"deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/docker.gpg] https://download.docker.com/linux/ubuntu \
$(lsb_release -cs) stable" | tee /etc/apt/sources.list.d/docker.list > /dev/null
|
1
2
| apt update
apt install containerd.io=1.6.31-1
|
1
2
| mkdir -p /etc/containerd
containerd config default > /etc/containerd/config.toml
|
1
2
3
4
| sed -i 's#root = "/var/lib/containerd"#root = "/data/containerd"#g' /etc/containerd/config.toml
sed -i 's#state = "/run/containerd"#state = "/data/run/containerd"#g' /etc/containerd/config.toml
sed -i 's#sandbox_image = "registry.k8s.io/pause:3.6"#sandbox_image = "registry.aliyuncs.com/google_containers/pause:3.9"#g' /etc/containerd/config.toml
sed -i 's#SystemdCgroup = false#SystemdCgroup = true#g' /etc/containerd/config.toml
|
1
| systemctl restart containerd
|
2.2 安装 nvidia-container-runtime
1
2
3
4
| curl -s -L https://nvidia.github.io/nvidia-container-runtime/gpgkey | apt-key add -
distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
curl -s -L https://nvidia.github.io/nvidia-container-runtime/$distribution/nvidia-container-runtime.list | tee /etc/apt/sources.list.d/nvidia-container-runtime.list
apt-get update
|
1
| apt-get install -y nvidia-container-runtime
|
1
2
| distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
curl -s -L https://nvidia.github.io/nvidia-container-runtime/$distribution/nvidia-container-runtime.repo | tee /etc/yum.repos.d/nvidia-container-runtime.repo
|
1
| yum install -y nvidia-container-runtime
|
2.3 Docker 配置
配置 Docker 开启 GPU 支持
1
| vim /etc/docker/daemon.json
|
添加以下内容:
1
2
3
4
5
6
7
8
9
| {
"default-runtime": "nvidia",
"runtimes": {
"nvidia": {
"path": "/usr/bin/nvidia-container-runtime",
"runtimeArgs": []
}
}
}
|
1
2
| systemctl daemon-reload
systemctl restart docker
|
1
| docker run --rm --gpus all registry.cn-beijing.aliyuncs.com/opshub/ubuntu nvidia-smi
|
此时可以看到输出的 GPU 信息。
2.4 Containerd 配置
1
| vim /etc/containerd/config.toml
|
在与 plugins."io.containerd.grpc.v1.cri".containerd.runtimes
中添加:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
| [plugins."io.containerd.grpc.v1.cri".containerd.runtimes.nvidia]
privileged_without_host_devices = false
runtime_engine = ""
runtime_root = ""
runtime_type = "io.containerd.runc.v2"
[plugins."io.containerd.grpc.v1.cri".containerd.runtimes.nvidia.options]
BinaryName = "/usr/bin/nvidia-container-runtime"
CriuImagePath = ""
CriuPath = ""
CriuWorkPath = ""
IoGid = 0
IoUid = 0
NoNewKeyring = false
NoPivotRoot = false
Root = ""
ShimCgroup = ""
SystemdCgroup = true
|
将默认的 runtime 设置为 nvidia
1
2
| [plugins."io.containerd.grpc.v1.cri".containerd]
default_runtime_name = "nvidia"
|
1
2
| systemctl daemon-reload
systemctl restart containerd
|
1
2
| export IMAGE=registry.cn-beijing.aliyuncs.com/opshub/ubuntu
nerdctl run --rm --gpus all $IMAGE nvidia-smi
|
CUDA 是 NVIDIA 推出的通用并行计算架构,用于在 GPU 上进行通用计算。CUDA Toolkit 是 CUDA 的开发工具包,包含了编译器(NVCC)、库、调试器等工具。
3.1 检查系统是否支持
参考 https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#system-requirements 有最新的 CUDA 对 CPU 架构、操作系统、GCC 版本、GLIBC 版本的依赖要求。
1
| uname -m && cat /etc/os-release
|
3.2 兼容性说明
使用 nvidia-smi
命令可以看到一个 CUDA 的版本号,但这个版本号是 CUDA driver libcuda.so 的版本号,不是 CUDA Toolkit 的版本号。
如上图 CUDA driver 是向后兼容的,即支持之前的 CUDA Toolkit 版本。
如上图,CUDA driver 支持向前的次要版本兼容,即大版本号相同就支持。参考[2]。
3.3 安装 CUDA
前往 https://developer.nvidia.com/cuda-downloads 选择对应的版本下载。这里以 Ubuntu 20.04 的 runfile(local) 为例:
1
| wget https://developer.download.nvidia.com/compute/cuda/12.3.1/local_installers/cuda_12.3.1_545.23.08_linux.run
|
1
| sh cuda_12.3.1_545.23.08_linux.run
|
增加以下内容:
1
2
3
| export PATH=$PATH:$PATH:/usr/local/cuda/bin
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64
export CUDA_HOME=$CUDA_HOME:/usr/local/cuda
|
使环境变量立即生效:
4. 安装 cuDNN
cuDNN 是 NVIDIA 基于 CUDA 开发的深度神经网络加速库。
前往 https://docs.nvidia.com/deeplearning/cudnn/support-matrix/index.html 查看 cuDNN 与 CUDA、Driver、操作系统的兼容性是否满足要求。
前往 https://developer.nvidia.com/rdp/cudnn-archive 下载对应的版本,选择 Local Installer for Linux x86_64 (Tar) ,会得到一个 tar.xz 的压缩包。
1
| tar -xvf cudnn-linux-*-archive.tar.xz
|
1
2
3
| cp cudnn-*-archive/include/cudnn*.h /usr/local/cuda/include
cp -P cudnn-*-archive/lib/libcudnn* /usr/local/cuda/lib64
chmod a+r /usr/local/cuda/include/cudnn*.h /usr/local/cuda/lib64/libcudnn*
|
5. 开启持久模式
使用 nvidia-smi -pm 1
能够开启持久模式,但重启后会失效,同时使用 nvidia-smi
的方式已经被归档,推荐使用 nvidia-persistenced
常驻进程。
开启持久模式之后,驱动一直会被加载,会消耗更多能源,但能有效改善各种显卡故障。
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
| cat <<EOF > /lib/systemd/system/nvidia-persistenced.service
[Unit]
Description=NVIDIA Persistence Daemon
After=syslog.target
[Service]
Type=forking
PIDFile=/var/run/nvidia-persistenced/nvidia-persistenced.pid
Restart=always
ExecStart=/usr/bin/nvidia-persistenced --verbose
ExecStopPost=/bin/rm -rf /var/run/nvidia-persistenced/*
TimeoutSec=300
[Install]
WantedBy=multi-user.target
EOF
|
1
| systemctl start nvidia-persistenced
|
1
| systemctl status nvidia-persistenced
|
1
| systemctl enable nvidia-persistenced
|
6. 安装 NVLink 和 NVSwitch 驱动
如果装配了 NVLink 或者 NVSwitch ,还需要安装 nvidia-fabricmanager,否则无法正常工作。
在 https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/ 找到合适的版本。
1
| wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/nvidia-fabricmanager-535_535.129.03-1_amd64.deb
|
1
| apt install ./nvidia-fabricmanager-535_535.129.03-1_amd64.deb
|
- 启动 nvidia-fabricmanager 服务
1
| systemctl start nvidia-fabricmanager
|
- 查看 nvidia-fabricmanager 服务
1
| systemctl status nvidia-fabricmanager
|
1
| systemctl enable nvidia-fabricmanager
|
7. 安装 InfiniBand 驱动
1
| wget https://content.mellanox.com/ofed/MLNX_OFED-4.9-5.1.0.0/MLNX_OFED_LINUX-4.9-5.1.0.0-ubuntu20.04-x86_64.tgz
|
1
2
3
| tar zxf MLNX_OFED_LINUX-4.9-5.1.0.0-ubuntu20.04-x86_64.tgz
cd MLNX_OFED_LINUX-4.9-5.1.0.0-ubuntu20.04-x86_64
./mlnxofedinstall
|
然后重启机器,可以查看驱动状态
1
2
3
4
5
6
7
8
9
10
11
| systemctl status openibd
● openibd.service - openibd - configure Mellanox devices
Loaded: loaded (/lib/systemd/system/openibd.service; enabled; vendor preset: enabled)
Active: active (exited) since Mon 2024-03-11 15:30:58 CST; 1 weeks 0 days ago
Docs: file:/etc/infiniband/openib.conf
Process: 2261 ExecStart=/etc/init.d/openibd start bootid=65648015406c4b88b831c8b907ad4ec6 (code=exited, status=0/SUCCESS)
Main PID: 2261 (code=exited, status=0/SUCCESS)
Tasks: 0 (limit: 618654)
Memory: 24.6M
CGroup: /system.slice/openibd.service
|
通过 ibstat 可以查看设备信息
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
| ibstat
ibstat
CA 'mlx5_0'
CA type: MT4123
Number of ports: 1
Firmware version: 20.35.1012
Hardware version: 0
Node GUID: 0x946dae03008bcc68
System image GUID: 0x946dae03008bcc68
Port 1:
State: Active
Physical state: LinkUp
Rate: 200
Base lid: 124
LMC: 0
SM lid: 1
Capability mask: 0xa651e848
Port GUID: 0x946dae03008bcc68
Link layer: InfiniBand
CA 'mlx5_1'
CA type: MT4123
Number of ports: 1
Firmware version: 20.35.1012
Hardware version: 0
Node GUID: 0x946dae03008bcc3c
System image GUID: 0x946dae03008bcc3c
Port 1:
State: Active
Physical state: LinkUp
Rate: 200
Base lid: 126
LMC: 0
SM lid: 1
Capability mask: 0xa651e848
Port GUID: 0x946dae03008bcc3c
Link layer: InfiniBand
|
8. 加入 K8s 集群
8.1 修改 Hostname
1
| export HOSTNAME=k8s-worker-gpu-01
|
1
| hostnamectl set-hostname ${HOSTNAME}
|
8.2 初始化内核参数
1
| opscli task -f ~/.ops/tasks/set-host.yaml
|
8.3 安装 K8s 基础组件
https://developer.aliyun.com/mirror/kubernetes/ 1.28 以下版本添加
1
| curl https://mirrors.aliyun.com/kubernetes/apt/doc/apt-key.gpg | apt-key add -
|
1
2
3
| cat <<EOF >/etc/apt/sources.list.d/kubernetes.list
deb https://mirrors.aliyun.com/kubernetes/apt/ kubernetes-xenial main
EOF
|
1
| export K8S_VERSION=1.27.6
|
1
| apt-get install kubeadm=${K8S_VERSION}-00 kubelet=${K8S_VERSION}-00 kubectl=${K8S_VERSION}-00 -y
|
8.4 加入集群
在 master 节点生成 token
1
| kubeadm token create --print-join-command
|
1
2
| kubeadm join x.x.x.x:6443 --token xxx \
--discovery-token-ca-cert-hash sha256:xxx
|
8.5 创建测试的 Pod
国内
1
| export IMAGE=registry.cn-beijing.aliyuncs.com/opshub/nvidia-cuda:12.3.2-base-ubuntu22.04
|
海外
1
| export IMAGE=hubimage/nvidia-cuda:12.3.2-base-ubuntu22.04
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
| kubectl apply -f - <<EOF
apiVersion: v1
kind: Pod
metadata:
name: gpu-demo
spec:
nodeName: ${HOSTNAME}
containers:
- name: gpu-demo
image: ${IMAGE}
command: ["nvidia-smi"]
resources:
requests:
tencent.com/vcuda-core: 400
limits:
tencent.com/vcuda-core: 400
EOF
|
1
| kubectl delete pod gpu-demo
|
9. 部署 k8s-rdma-shared-dev-plugin
为了让 Kubernetes 能够发现 RDMA 设备,比如 IfiniBand ,并且被多个 Pod 使用,需要安装 k8s-rdma-shared-dev-plugin。
- 安装 k8s-rdma-shared-dev-plugin
1
| kubectl apply -f https://raw.githubusercontent.com/shaowenchen/hubimage/main/network/k8s-rdma-shared-dev-plugin.yaml
|
1
| kubectl -n kube-system edit cm rdma-devices
|
在 spec 中配置 rdma/ib
就可以使用了。
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
| spec:
containers:
- command:
- /bin/sh
- -c
- mkdir -p /var/run/sshd; /usr/sbin/sshd;bash llama_distributed_v3.0_check.sh
resources:
limits:
cpu: "64"
memory: 950Gi
rdma/ib: "8"
tencent.com/vcuda-core: "800"
requests:
cpu: "64"
memory: 950Gi
rdma/ib: "8"
tencent.com/vcuda-core: "800"
|
10. 参考
- https://nvidia.github.io/nvidia-container-runtime/
- https://tianzhipeng-git.github.io/2023/11/21/cuda-version.html