+-
使用GPU-Operator简化深度学习训练与监控GPU

本文将从 GPU-Operator 概念介绍、安装部署、深度训练测试应用部署,以及在 KubeSphere 使用自定义监控面板对接 GPU 监控,从原理到实践,逐步浅析介绍与实践 GPU-Operator。

GPU-Operator简介

众所周知,Kubernetes 平台通过设备插件框架提供对特殊硬件资源的访问,如 nvidia GPU、网卡、Infiniband 适配器和其他设备。然而,使用这些硬件资源配置和管理节点需要配置多个软件组件,如驱动程序、容器运行时或其他依赖库,这是困难的和容易出错的。

NVIDIA GPU Operator 由 Nvidia 公司开源,利用了 Kubernetes 平台的 Operator 控制模式,方便地自动化集成管理 GPU 所需的 NVIDIA 设备组件,有效地解决了上述GPU设备集成的痛点。这些组件包括 NVIDIA 驱动程序(用于启用 CUDA )、用于 GPU 的 Kubernetes 设备插件、NVIDIA Container 运行时、自动节点标签、基于 DCGM 的监控等。

NVIDIA GPU Operator 的不仅实现了设备和组件一体化集成,而且它管理 GPU 节点就像管理 cpu 节点一样方便,无需单独为 GPU 节点提供特殊的操作系统。值得关注的是,它将GPU各组件容器化,提供 GPU 能力,非常适合快速扩展和管理规模 GPU 节点。当然,对于已经为GPU组件构建了特殊操作系统的应用场景来说,显得并不是那么合适了。

GPU-Operator 架构原理

前文提到,NVIDIA GPU Operator 管理 GPU 节点就像管理 CPU 节点一样方便,那么它是如何实现这一能力呢?

我们一起来看看 GPU-Operator 运行时的架构图:

...

通过图中的描述,我们可以知道, GPU-Operator 是通过实现了 Nvidia 容器运行时,以runC作为输入,在runC中preStart hook中注入了一个名叫nvidia-container-toolkit的脚本,该脚本调用libnvidia-container CLI设置一系列合适的flags,使得容器运行后具有 GPU 能力。

GPU-Operator 安装说明

前提条件

在安装 GPU Operator 之前,请配置好安装环境如下:

所有节点 不需要预先安装NVIDIA组件(driver,container runtime,device plugin); 所有节点必须配置docker,cri-o, 或者containerd.对于 docker 来说,可以参考这里; 如果使用HWE内核(e.g. kernel 5.x) 的 Ubuntu 18.04 LTS 环境下,需要给nouveau driver添加黑名单,需要更新initramfs;
$ sudo vim /etc/modprobe.d/blacklist.conf # 在尾部添加黑名单 blacklist nouveau options nouveau modeset=0 $ sudo update-initramfs -u $ reboot $ lsmod | grep nouveau # 验证nouveau是否已禁用 $ cat /proc/cpuinfo | grep name | cut -f2 -d: | uniq -c #本文测试时处理器架构代号为Broadwell 16 Intel Core Processor (Broadwell) 节点发现(NFD) 需要在每个节点上配置,默认情况会直接安装,如果已经配置,请在Helm chart变量设置nfd.enabled为false, 再安装; 如果使用 Kubernetes 1.13和1.14, 需要激活 KubeletPodResources;

支持的linux版本

OS Name / Version

Identifier

amd64 / x86_64

ppc64le

arm64 / aarch64

Amazon Linux 1

amzn1

X



Amazon Linux 2

amzn2

X



Amazon Linux 2017.09

amzn2017.09

X



Amazon Linux 2018.03

amzn2018.03

X



Open Suse Leap 15.0

sles15.0

X



Open Suse Leap 15.1

sles15.1

X



Debian Linux 9

debian9

X



Debian Linux 10

debian10

X



Centos 7

centos7

X

X


Centos 8

centos8

X

X

X

RHEL 7.4

rhel7.4

X

X


RHEL 7.5

rhel7.5

X

X


RHEL 7.6

rhel7.6

X

X


RHEL 7.7

rhel7.7

X

X


RHEL 8.0

rhel8.0

X

X

X

RHEL 8.1

rhel8.1

X

X

X

RHEL 8.2

rhel8.2

X

X

X

Ubuntu 16.04

ubuntu16.04

X

X


Ubuntu 18.04

ubuntu18.04

X

X

X

Ubuntu 20.04

ubuntu20.04

X

X

X

支持的容器运行时

OS Name / Version

amd64 / x86_64

ppc64le

arm64 / aarch64

Docker 18.09

X

X

X

Docker 19.03

X

X

X

RHEL/CentOS 8 podman

X



CentOS 8 Docker

X



RHEL/CentOS 7 Docker

X



安装doker环境

可参考 Docker 官方文档

安装NVIDIA Docker

配置 stable 仓库和 GPG key :

$ distribution=$(. /etc/os-release;echo $ID$VERSION_ID) \ && curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add - \ && curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list

更新软件仓库后安装nvidia-docker2并添加运行时配置:

$ sudo apt-get update $ sudo apt-get install -y nvidia-docker2 ----- What would you like to do about it ? Your options are: Y or I : install the package maintainer's version N or O : keep your currently-installed version D : show the differences between the versions Z : start a shell to examine the situation ----- # 初次安装,遇到以上交互式问题可选择N # 如果选择Y会覆盖你的一些默认配置 # 选择N后,将以下配置添加到etc/docker/daemon.json { "runtimes": { "nvidia": { "path": "/usr/bin/nvidia-container-runtime", "runtimeArgs": [] } } }

重启docker:

$ sudo systemctl restart docker

安装Helm

$ curl -fsSL -o get_helm.sh https://raw.githubusercontent.com/helm/helm/master/scripts/get-helm-3 \ && chmod 700 get_helm.sh \ && ./get_helm.sh

添加helm仓库

$ helm repo add nvidia https://nvidia.github.io/gpu-operator \ && helm repo update

安装 NVIDIA GPU Operator

docker as runtime

$ kubectl create ns gpu-operator-resources $ helm install gpu-operator nvidia/gpu-operator -n gpu-operator-resources --wait

如果需要指定驱动版本,可参考如下:

$ helm install gpu-operator nvidia/gpu-operator -n gpu-operator-resources \ --set driver.version="450.80.02"

crio as runtime

helm install gpu-operator nvidia/gpu-operator -n gpu-operator-resources\ --set operator.defaultRuntime=crio

containerd as runtime

helm install gpu-operator nvidia/gpu-operator -n gpu-operator-resources\ --set operator.defaultRuntime=containerd Furthermore, when setting containerd as the defaultRuntime the following options are also available: toolkit: env: - name: CONTAINERD_CONFIG value: /etc/containerd/config.toml - name: CONTAINERD_SOCKET value: /run/containerd/containerd.sock - name: CONTAINERD_RUNTIME_CLASS value: nvidia - name: CONTAINERD_SET_AS_DEFAULT value: true

由于安装的镜像比较大,所以初次安装过程中可能会出现超时的情形,请检查你的镜像是否在拉取中!可以考虑使用离线安装解决该类问题,参考离线安装的链接。

使用 values.yaml 安装

$ helm install gpu-operator nvidia/gpu-operator -n gpu-operator-resources -f values.yaml

考虑离线安装

应用部署

检查已部署 operator 服务状态

检查 pods 状态

$ kubectl get pods -n gpu-operator-resources NAME READY STATUS RESTARTS AGE gpu-feature-discovery-4gk78 1/1 Running 0 35s gpu-operator-858fc55fdb-jv488 1/1 Running 0 2m52s gpu-operator-node-feature-discovery-master-7f9ccc4c7b-2sg6r 1/1 Running 0 2m52s gpu-operator-node-feature-discovery-worker-cbkhn 1/1 Running 0 2m52s gpu-operator-node-feature-discovery-worker-m8jcm 1/1 Running 0 2m52s nvidia-container-toolkit-daemonset-tfwqt 1/1 Running 0 2m42s nvidia-dcgm-exporter-mqns5 1/1 Running 0 38s nvidia-device-plugin-daemonset-7npbs 1/1 Running 0 53s nvidia-device-plugin-validation 0/1 Completed 0 49s nvidia-driver-daemonset-hgv6s 1/1 Running 0 2m47s

检查节点资源是否处于可分配

$ kubectl describe node worker-gpu-001 --- Allocatable: cpu: 15600m ephemeral-storage: 82435528Ki hugepages-2Mi: 0 memory: 63649242267 nvidia.com/gpu: 1 #check here pods: 110 ---

部署官方文档中的两个实例

实例一

$ cat cuda-load-generator.yaml apiVersion: v1 kind: Pod metadata: name: dcgmproftester spec: restartPolicy: OnFailure containers: - name: dcgmproftester11 image: nvidia/samples:dcgmproftester-2.0.10-cuda11.0-ubuntu18.04 args: ["--no-dcgm-validation", "-t 1004", "-d 120"] resources: limits: nvidia.com/gpu: 1 securityContext: capabilities: add: ["SYS_ADMIN"] EOF

实例二

$ curl -LO https://nvidia.github.io/gpu-operator/notebook-example.yml $ cat notebook-example.yml apiVersion: v1 kind: Service metadata: name: tf-notebook labels: app: tf-notebook spec: type: NodePort ports: - port: 80 name: http targetPort: 8888 nodePort: 30001 selector: app: tf-notebook --- apiVersion: v1 kind: Pod metadata: name: tf-notebook labels: app: tf-notebook spec: securityContext: fsGroup: 0 containers: - name: tf-notebook image: tensorflow/tensorflow:latest-gpu-jupyter resources: limits: nvidia.com/gpu: 1 ports: - containerPort: 8

基于 Jupyter Notebook 应用运行深度学习训练任务

部署应用

$ kubectl apply -f cuda-load-generator.yaml pod/dcgmproftester created $ kubectl apply -f notebook-example.yml service/tf-notebook created pod/tf-notebook created

查看 GPU 处于已分配状态:

$ kubectl describe node worker-gpu-001 --- Allocated resources: (Total limits may be over 100 percent, i.e., overcommitted.) Resource Requests Limits -------- -------- ------ cpu 1087m (6%) 1680m (10%) memory 1440Mi (2%) 1510Mi (2%) ephemeral-storage 0 (0%) 0 (0%) nvidia.com/gpu 1 1 #check this Events: <none>

当有 GPU 任务发布给平台时,GPU 资源从可分配状态转变为已分配状态,安装任务发布的先后顺序,第二个任务在第一个任务运行结束后开始运行:

$ kubectl get pods --watch NAME READY STATUS RESTARTS AGE dcgmproftester 1/1 Running 0 76s tf-notebook 0/1 Pending 0 58s ------ NAME READY STATUS RESTARTS AGE dcgmproftester 0/1 Completed 0 4m22s tf-notebook 1/1 Running 0 4m4s

获取应用端口信息:

$ kubectl get svc # get the nodeport of the svc, 30001 gpu-operator-1611672791-node-feature-discovery ClusterIP 10.233.10.222 <none> 8080/TCP 12h kubernetes ClusterIP 10.233.0.1 <none> 443/TCP 12h tf-notebook NodePort 10.233.53.116 <none> 80:30001/TCP 7m52s

查看日志,获取登录口令:

$ kubectl logs tf-notebook [I 21:50:23.188 NotebookApp] Writing notebook server cookie secret to /root/.local/share/jupyter/runtime/notebook_cookie_secret [I 21:50:23.390 NotebookApp] Serving notebooks from local directory: /tf [I 21:50:23.391 NotebookApp] The Jupyter Notebook is running at: [I 21:50:23.391 NotebookApp] http://tf-notebook:8888/?token=3660c9ee9b225458faaf853200bc512ff2206f635ab2b1d9 [I 21:50:23.391 NotebookApp] or http://127.0.0.1:8888/?token=3660c9ee9b225458faaf853200bc512ff2206f635ab2b1d9 [I 21:50:23.391 NotebookApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation). [C 21:50:23.394 NotebookApp] To access the notebook, open this file in a browser: file:///root/.local/share/jupyter/runtime/nbserver-1-open.html Or copy and paste one of these URLs: http://tf-notebook:8888/?token=3660c9ee9b225458faaf853200bc512ff2206f635ab2b1d9 or http://127.0.0.1:8888/?token=3660c9ee9b225458faaf853200bc512ff2206f635ab2b1d9

运行深度学习任务

进入jupyter notebook 环境后,尝试进入终端,运行深度学习任务:

...

进入terminal后拉取tersorflow测试代码并运行:

...

与此同时,开启另外一个终端运行nvidia-smi查看 GPU 监控使用情况:

...

利用 KubeSphere 自定义监控功能监控 GPU

部署 ServiceMonitor

gpu-operator帮我们提供了nvidia-dcgm-exporter这个exportor, 只需要将它集成到Prometheus的可采集对象中,也就是ServiceMonitor中,我们就能获取GPU监控数据了:

$ kubectl get pods -n gpu-operator-resources NAME READY STATUS RESTARTS AGE gpu-feature-discovery-ff4ng 1/1 Running 2 15h nvidia-container-toolkit-daemonset-2vxjz 1/1 Running 0 15h nvidia-dcgm-exporter-pqwfv 1/1 Running 0 5h27m #here nvidia-device-plugin-daemonset-42n74 1/1 Running 0 5h27m nvidia-device-plugin-validation 0/1 Completed 0 5h27m nvidia-driver-daemonset-dvd9r 1/1 Running 3 15h

可以构建一个busybox查看该exporter暴露的指标:

$ kubectl get svc -n gpu-operator-resources NAME TYPE CLUSTER-IP EXTERNAL-IP PORT(S) AGE gpu-operator-node-feature-discovery ClusterIP 10.233.54.111 <none> 8080/TCP 56m nvidia-dcgm-exporter ClusterIP 10.233.53.196 <none> 9400/TCP 54m $ kubectl exec -it busybox-sleep -- sh $ wget http://nvidia-dcgm-exporter.gpu-operator-resources:9400/metrics $ cat metrics ---- DCGM_FI_DEV_SM_CLOCK{gpu="0",UUID="GPU-eeff7856-475a-2eb7-6408-48d023d9dd28",device="nvidia0",container="tf-notebook",namespace="default",pod="tf-notebook"} 405 DCGM_FI_DEV_MEM_CLOCK{gpu="0",UUID="GPU-eeff7856-475a-2eb7-6408-48d023d9dd28",device="nvidia0",container="tf-notebook",namespace="default",pod="tf-notebook"} 715 DCGM_FI_DEV_GPU_TEMP{gpu="0",UUID="GPU-eeff7856-475a-2eb7-6408-48d023d9dd28",device="nvidia0",container="tf-notebook",namespace="default",pod="tf-notebook"} 30 ----

查看nvidia-dcgm-exporter暴露的svc和ep:

$ kubectl describe svc nvidia-dcgm-exporter -n gpu-operator-resources Name: nvidia-dcgm-exporter Namespace: gpu-operator-resources Labels: app=nvidia-dcgm-exporter Annotations: prometheus.io/scrape: true Selector: app=nvidia-dcgm-exporter Type: NodePort IP: 10.233.28.200 Port: gpu-metrics 9400/TCP TargetPort: 9400/TCP NodePort: gpu-metrics 31129/TCP Endpoints: 10.233.84.54:9400 Session Affinity: None External Traffic Policy: Cluster Events: <none>

配置ServiceMonitor定义清单:

$ cat custom/gpu-servicemonitor.yaml apiVersion: monitoring.coreos.com/v1 kind: ServiceMonitor metadata: name: nvidia-dcgm-exporter namespace: gpu-operator-resources labels: app: nvidia-dcgm-exporter spec: jobLabel: nvidia-gpu endpoints: - port: gpu-metrics interval: 15s selector: matchLabels: app: nvidia-dcgm-exporter namespaceSelector: matchNames: - gpu-operator-resources $ kubectl apply -f custom/gpu-servicemonitor.yaml

检查 GPU 指标是否被采集到(可选)

将servicemonitor提交给kubesphere平台后,通过暴露prometheus-k8s为NodePort,我们可以在Prometheus的UI上验证一下是否采集到的相关指标:

...

创建 KubeSphere GPU 自定义监控面板

KubeSphere 3.0

如果部署的 KubeSphere 版本是KubeSphere 3.0,需要简单地配置以下几个步骤,便可顺利完成可观察性监控。

首先, 登录kubsphere console后,创建一个企业空间名称为ks-monitoring-demo, 名称可按需创建;

其次,需要将ServiceMonitor所在的目标名称空间gpu-operator-resources分配为已存在的企业空间中,以便纳入监控。

...

最后,进入目标企业空间,在纳管的项目找到gpu-operator-resources, 点击后找到可自定义监控界面, 即可添加自定义监控。

... ...

后续版本

后续版本可选择添加集群监控

创建自定义监控

下载dashboard以及配置namespace:

$ curl -LO https://raw.githubusercontent.com/kubesphere/monitoring-dashboard/master/contrib/gallery/nvidia-gpu-dcgm-exporter-dashboard.yaml $ cat nvidia-gpu-dcgm-exporter-dashboard.yaml ---- apiVersion: monitoring.kubesphere.io/v1alpha1 kind: Dashboard metadata: name: nvidia-dcgm-exporter-dashboard-rev1 namespace: gpu-operator-resources # check here spec: -----

可以直接命令行apply或者在自定义监控面板中选择编辑模式进行导入:

...

正确导入后:

...

在上面创建的jupyter notebook运行深度学习测试任务后,可以明显地观察到相关GPU指标变化:

...

卸载

$ helm list -n gpu-operator-resources NAME NAMESPACE REVISION UPDATED STATUS CHART APP VERSION gpu-operator gpu-operator-resources 1 2021-02-20 11:50:56.162559286 +0800 CST deployed gpu-operator-1.5.2 1.5.2 $ helm uninstall gpu-operator -n gpu-operator-resources

重启无法使用 GPU

关于已部署正常运行的gpu-operator和AI应用的集群,重启GPU主机后会出现没法用上 GPU 的情况,极有可能是因为插件还没加载,应用优先进行了载入,就会导致这种问题。这时,只需要优先保证插件运行正常,然后重新部署应用即可。

GPU-Operator 常见问题

GPU-Operator 重启后无法使用

答:关于已部署正常运行的gpu-operator和 AI 应用的集群,重启 GPU 主机后会出现没法用上 GPU 的情况,极有可能是因为插件还没加载,应用优先进行了载入,就会导致这种问题。这时,只需要优先保证插件运行正常,然后重新部署应用即可。

Nvidia k8s-device-plugin 与 GPU-Operator 方案对比?

我之前针对GPU使用的是 https://github.com/NVIDIA/k8s-device-plugin 和 https://github.com/NVIDIA/gpu-monitoring-tools 相结合的方案来监控 GPU,请问这个方案与 GPU-Operator的方案相比,孰优孰劣一些?

答:个人认为 GPU-Operator 更简单易用,其自带 GPU 注入能力不需要构建专用的 OS,并且支持节点发现与可插拔,能够自动化集成管理 GPU 所需的 NVIDIA 设备组件,相对来说还是很省事的。

有没有 KubeSphere 自定义监控的详细使用教程?

答:可以参考 KubeSphere 官方文档来使用自定义监控。

参考资料

官方代码仓库

GitHub: https://github.com/NVIDIA/gpu-operator GitLab: https://gitlab.com/nvidia/kubernetes/gpu-operator

官方文档

GPU-Operator 快速入门:https://docs.nvidia.com/datacenter/cloud-native/gpu-operator/getting-started.html#install-nvidia-gpu-operator GPU-Operator 离线安装指南:https://docs.nvidia.com/datacenter/cloud-native/gpu-operator/getting-started.html#considerations-to-install-in-air-gapped-clusters KubeSphere 自定义监控使用文档:https://kubesphere.com.cn/docs/project-user-guide/custom-application-monitoring/examples/monitor-mysql/