kubernetes集群GPU支持方案
一.kubernetes对GPU的支持版本
kubernetes提供对分布式节点上的AMD GPU和NVIDIA GPU管理的实验性的支持。在V1.6中已经添加了对NVIDIA GPU的支持，并且经历了多次
向后不兼容的迭代。通过设备插件在v1.9中添加了对AMD GPU的支持。
从1.8版本开始，使用GPU的推荐方法是使用驱动插件。要是在1.10版本之前通过设备插件启用GPU支持，必须在整个系统中将DevicePlugins功能
设置为true: --feature-gates="DevicePlugins=true 。1.10之后版本不需要这么做了。
然后，必须在节点上安装相应供应商GPU驱动程序，并从GPU供应商(AMD/NVIDIA)运行相应的设备插件。
二.kubernetes集群部署GPU
kubernetes集群版本: 1.13.5
docker版本: 18.06.1-ce
os系统是版本: centos7.5
内核版本: 4.20.13-1.el7.elrepo.x86_64
Nvidia GPU型号: P4000
2.1 安装nvidia驱动
2.1.1 安装gcc
[root@k8s-01 ~]# yum install -y gcc
2.1.2 下载nvidia的驱动
下载链接: https://www.nvidia.cn/Download/driverResults.aspx/141795/cn
这里我们下载的是如下版本:
[root@k8s-01 ~]# ls NVIDIA-Linux-x86_64-410.93.run -alh
-rw-r--r-- 1 root root 103M Jul 25 17:22 NVIDIA-Linux-x86_64-410.93.run
2.1.3 修改/etc/modprobe.d/blacklist.conf文件，阻止nouveau模块的加载
[root@k8s-01 ~]# echo -e "blacklist nouveau\noptions nouveau modeset=0" > /etc/modprobe.d/blacklist.conf
2.1.4 重新建立initramfs image
[root@k8s-01 ~]# mv /boot/initramfs-$(uname -r).img /boot/initramfs-$(uname -r).img.bak
[root@k8s-01 ~]# dracut /boot/initramfs-$(uname -r).img $(uname -r)
2.1.5 执行驱动安装
[root@k8s-01 ~]# sh NVIDIA-Linux-x86_64-410.93.run -a -q -s
2.1.6 安装工具包
只有驱动是不够的，我们需要一些工具包便于我们使用，其中cuda、cudnn是相关工具包。
[root@k8s-01 ~]# cat /etc/yum.repos.d/cuda.repo
[cuda]
name=cuda
baseurl=http://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64
enabled=1
gpgcheck=1
gpgkey=http://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/7fa2af80.pub
[root@k8s-01 ~]#
2.2 安装nvidia-docker2
nvidia-docker是一个可以使用GPU的docker，在docker的基础上做了一层封装。目前基本被弃用。
nvidia-docker2是一个runtime，能更好的和docker兼容。
- 获取nvidia-docker2的yum源
[root@k8s-01 ~]# distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
[root@k8s-01 ~]# curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.repo | sudo tee /etc/yum.repos.d/nvidia-docker.repo
- 查看nvidia-docker2的列表
这里我们需要安装支持docker-18.06.1-ce版本的nvidia-docker2版本，否则会不支持。
[root@k8s-01 ~]# yum list nvidia-docker2 --showduplicate
Loaded plugins: fastestmirror
Loading mirror speeds from cached hostfile
* base: mirrors.aliyun.com
* epel: mirror01.idc.hinet.net
* extras: mirrors.aliyun.com
* updates: mirrors.163.com
Installed Packages
nvidia-docker2.noarch 2.0.3-1.docker18.06.1.ce @nvidia-docker
Available Packages
nvidia-docker2.noarch 2.0.0-1.docker1.12.6 nvidia-docker
nvidia-docker2.noarch 2.0.0-1.docker17.03.2.ce nvidia-docker
nvidia-docker2.noarch 2.0.0-1.docker17.06.1.ce nvidia-docker
nvidia-docker2.noarch 2.0.0-1.docker17.06.2.ce nvidia-docker
nvidia-docker2.noarch 2.0.0-1.docker17.09.0.ce nvidia-docker
nvidia-docker2.noarch 2.0.1-1.docker1.12.6 nvidia-docker
nvidia-docker2.noarch 2.0.1-1.docker1.13.1 nvidia-docker
nvidia-docker2.noarch 2.0.1-1.docker17.03.2.ce nvidia-docker
nvidia-docker2.noarch 2.0.1-1.docker17.06.2.ce nvidia-docker
nvidia-docker2.noarch 2.0.1-1.docker17.09.0.ce nvidia-docker
nvidia-docker2.noarch 2.0.1-1.docker17.09.1.ce nvidia-docker
nvidia-docker2.noarch 2.0.2-1.docker1.12.6 nvidia-docker
nvidia-docker2.noarch 2.0.2-1.docker1.13.1 nvidia-docker
nvidia-docker2.noarch 2.0.2-1.docker17.03.2.ce nvidia-docker
nvidia-docker2.noarch 2.0.2-1.docker17.06.2.ce nvidia-docker
nvidia-docker2.noarch 2.0.2-1.docker17.09.0.ce nvidia-docker
nvidia-docker2.noarch 2.0.2-1.docker17.09.1.ce nvidia-docker
nvidia-docker2.noarch 2.0.2-1.docker17.12.0.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker1.12.6 nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker1.13.1 nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker17.03.2.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker17.06.2.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker17.09.0.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker17.09.1.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker17.12.0.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker17.12.1.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker18.03.0.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker18.03.1.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker18.06.0.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker18.06.1.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker18.06.2 nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker18.09.0.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker18.09.1.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker18.09.2 nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker18.09.2.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker18.09.3.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-1.docker18.09.4.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-2.docker18.06.2.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-2.docker18.09.5.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-3.docker18.06.3.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-3.docker18.09.5.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-3.docker18.09.6.ce nvidia-docker
nvidia-docker2.noarch 2.0.3-3.docker18.09.7.ce nvidia-docker
nvidia-docker2.noarch 2.1.0-1 nvidia-docker
nvidia-docker2.noarch 2.1.1-1 nvidia-docker
nvidia-docker2.noarch 2.2.0-1 nvidia-docker
这里我们安装2.0.3-1.docker18.06.1.ce版本即可。
- 安装nvidia-docker2
[root@k8s-01 ~]# yum install -y nvidia-docker2-2.0.3-1.docker18.06.1.ce
- 配置默认的docker runtime为nvidia
[root@k8s-01 ~]# cat /etc/docker/daemon.json
{
"default-runtime": "nvidia",
"runtimes": {
"nvidia": {
"path": "/usr/bin/nvidia-container-runtime",
"runtimeArgs": []
}
}
}
- 重启docker
[root@k8s-01 ~]# systemctl restart docker
- 查看docker信息
[root@k8s-01 wf-deploy]# docker info
Containers: 63
Running: 0
Paused: 0
Stopped: 63
Images: 51
Server Version: 18.06.1-ce
Storage Driver: overlay2
Backing Filesystem: xfs
Supports d_type: true
Native Overlay Diff: true
Logging Driver: json-file
Cgroup Driver: cgroupfs
Plugins:
Volume: local
Network: bridge host macvlan null overlay
Log: awslogs fluentd gcplogs gelf journald json-file logentries splunk syslog
Swarm: inactive
Runtimes: runc nvidia
Default Runtime: nvidia
Init Binary: docker-init
containerd version: 468a545b9edcd5932818eb9de8e72413e616e86e
runc version: 69663f0bd4b60df09991c08812a60108003fa340-dirty (expected: 69663f0bd4b60df09991c08812a60108003fa340)
init version: fec3683
Security Options:
seccomp
Profile: default
Kernel Version: 4.20.13-1.el7.elrepo.x86_64
Operating System: CentOS Linux 7 (Core)
OSType: linux
Architecture: x86_64
CPUs: 2
Total Memory: 7.79GiB
Name: k8s-01
ID: DWPY:P2I4:NWL4:3U3O:UTGC:PLJC:IGTO:7ZXJ:A7CD:SJGT:7WT5:WNGX
Docker Root Dir: /var/lib/docker
Debug Mode (client): false
Debug Mode (server): false
Registry: https://index.docker.io/v1/
Labels:
Experimental: false
Insecure Registries:
192.168.50.2
127.0.0.0/8
Live Restore Enabled: false
可以看出docker的默认runtime为nvidia
2.3 安装驱动插件
- 获取插件的最新yaml文件
[root@k8s-01 ~]# kubectl apply -f https://raw.githubusercontent.com/NVIDIA/k8s-device-plugin/v1.11/nvidia-device-plugin.yml
2.4 查看有GPU的节点
[root@wf-229 ~]# kubectl get node 192.18.1.26 -ojson | jq '.status.allocatable'
{
"cpu": "48",
"ephemeral-storage": "258961942919",
"hugepages-1Gi": "0",
"hugepages-2Mi": "0",
"memory": "131471388Ki",
"nvidia.com/gpu": "1",
"pods": "200"
}
2.5 创建包含GPU资源的POD
[root@wf-229 gpu]# cat test.yaml
apiVersion: v1
kind: Pod
metadata:
labels:
k8s-app: nginx-pod
name: nginx-pod
spec:
containers:
- image: nginx:latest
imagePullPolicy: Always
name: nginx
ports:
- containerPort: 80
name: nginx
protocol: TCP
resources:
limits:
nvidia.com/gpu: "1"
2.6 查看Pod中分配的GPU资源
[root@wf-229 gpu]# kubectl exec -it nginx-pod bash
root@nginx-pod:/# nvidia-smi
Mon Aug 12 11:39:05 2019
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 410.93 Driver Version: 410.93 CUDA Version: N/A |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 Quadro P4000 Off | 00000000:3B:00.0 Off | N/A |
| 46% 34C P8 5W / 105W | 0MiB / 8119MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+
三.CLI介绍
- nvidia-container-cli
nvidia-container-cli 是一个命令行工具，用于配置Linux容器对GPU 硬件的使用。支持:
1）list: 打印nvidia驱动库及路径
2）info: 打印所有Nvidia GPU设备
3）configure: 进入给定进程的命名空间，执行必要操作保证容器内可以使用被指定的GPU以及对应能力（指定 Nvidia 驱动库）。 configure是我们使用到的主要命令，它将Nvidia 驱动库的so文件 和 GPU设备信息， 通过文件挂载的方式映射到容器中。
- 查看NODE节点GPU
kubectl get node 192.18.1.26 -ojson | jq '.status.allocatable'
