乐趣区

关于云计算:使用-GPUOperator-与-KubeSphere-简化深度学习训练与监控-GPU

本文将从 GPU-Operator 概念介绍、装置部署、深度训练测试利用部署,以及在 KubeSphere 应用自定义监控面板对接 GPU 监控,从原理到实际,逐渐浅析介绍与实际 GPU-Operator。

GPU-Operator 简介

家喻户晓,Kubernetes 平台通过设施插件框架提供对非凡硬件资源的拜访,如 NVIDIA GPU、网卡、Infiniband 适配器和其余设施。然而,应用这些硬件资源配置和治理节点须要配置多个软件组件,如驱动程序、容器运行时或其余依赖库,这是艰难的和容易出错的。

NVIDIA GPU Operator 由 Nvidia 公司开源,利用了 Kubernetes 平台的 Operator 管制模式,不便地自动化集成治理 GPU 所需的 NVIDIA 设施组件,无效地解决了上述 GPU 设施集成的痛点。这些组件包含 NVIDIA 驱动程序(用于启用 CUDA)、用于 GPU 的 Kubernetes 设施插件、NVIDIA Container 运行时、主动节点标签、基于 DCGM 的监控等。

NVIDIA GPU Operator 的不仅实现了设施和组件一体化集成,而且它治理 GPU 节点就像治理 CPU 节点一样不便,无需独自为 GPU 节点提供非凡的操作系统。值得关注的是,它将 GPU 各组件容器化,提供 GPU 能力,非常适合疾速扩大和治理规模 GPU 节点。当然,对于曾经为 GPU 组件构建了非凡操作系统的利用场景来说,显得并不是那么适合了。

GPU-Operator 架构原理

前文提到,NVIDIA GPU Operator 治理 GPU 节点就像治理 CPU 节点一样不便,那么它是如何实现这一能力呢?

咱们一起来看看 GPU-Operator 运行时的架构图:

通过图中的形容,咱们能够晓得,GPU-Operator 是通过实现了 Nvidia 容器运行时,以 runC 作为输出,在 runCpreStart hook中注入了一个名叫 nvidia-container-toolkit 的脚本,该脚本调用 libnvidia-container CLI 设置一系列适合的flags,使得容器运行后具备 GPU 能力。

GPU-Operator 装置阐明

前提条件

在装置 GPU Operator 之前,请配置好装置环境如下:

  • 所有节点 不须要 事后装置 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,能够参考这里;
  • 如果应用 HWE 内核 (e.g. kernel 5.x) 的 Ubuntu 18.04 LTS 环境下, 须要给nouveau driver 增加黑名单,须要更新initramfs
$ sudo vim /etc/modprobe.d/blacklist.conf # 在尾部增加黑名单
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) 须要在每个节点上配置,默认状况会间接装置,如果曾经配置,请在Helm chart 变量设置 nfd.enabledfalse, 再装置;
  • 如果应用 Kubernetes 1.13 和 1.14, 须要激活 KubeletPodResources;

反对的 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

反对的容器运行时

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

装置 doker 环境

可参考 Docker 官网文档

装置 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

更新软件仓库后装置 nvidia-docker2 并增加运行时配置:

$ 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
-----
# 首次装置,遇到以上交互式问题可抉择 N
# 如果抉择 Y 会笼罩你的一些默认配置
# 抉择 N 后,将以下配置增加到 etc/docker/daemon.json
{
  "runtimes": {
      "nvidia": {
          "path": "/usr/bin/nvidia-container-runtime",
          "runtimeArgs": []}
  }
}

重启docker:

$ sudo systemctl restart docker

装置 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

增加 helm 仓库

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

装置 NVIDIA GPU Operator

docker as runtime

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

如果须要指定驱动版本,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

因为装置的镜像比拟大,所以首次装置过程中可能会呈现超时的情景,请查看你的镜像是否在拉取中!能够思考应用离线装置解决该类问题,参考离线装置的链接。

应用 values.yaml 装置

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

思考离线装置

利用部署

查看已部署 operator 服务状态

查看 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

查看节点资源是否处于可调配

$ 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
---

部署官网文档中的两个实例

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 利用运行深度学习训练任务

部署利用

$ 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 处于已调配状态:

$ 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 工作公布给平台时,GPU 资源从可调配状态转变为已调配状态,装置工作公布的先后顺序,第二个工作在第一个工作运行完结后开始运行:

$ 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

获取利用端口信息:

$ 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

运行深度学习工作

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

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 监控应用状况:

利用 KubeSphere 自定义监控性能监控 GPU

部署 ServiceMonitor

gpu-operator帮咱们提供了 nvidia-dcgm-exporter 这个 exportor, 只须要将它集成到Prometheus 的可采集对象中,也就是 ServiceMonitor 中,咱们就能获取 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

能够构建一个 busybox 查看该 exporter 裸露的指标:

$ 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 裸露的 svcep

$ 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

查看 GPU 指标是否被采集到(可选)

servicemonitor 提交给 kubesphere 平台后,通过裸露 prometheus-k8sNodePort,咱们能够在 PrometheusUI上验证一下是否采集到的相干指标:

创立 KubeSphere GPU 自定义监控面板

KubeSphere 3.0

如果部署的 KubeSphere 版本是KubeSphere 3.0,须要简略地配置以下几个步骤,便可顺利完成可察看性监控。

首先,登录 kubsphere console 后,创立一个企业空间名称为ks-monitoring-demo, 名称可按需创立;

其次,须要将 ServiceMonitor 所在的指标名称空间 gpu-operator-resources 调配为已存在的企业空间中,以便纳入监控。

最初,进入指标企业空间,在纳管的我的项目找到gpu-operator-resources, 点击后找到可自定义监控界面, 即可增加自定义监控。

后续版本

后续版本可抉择增加集群监控

创立自定义监控

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:
-----

能够间接命令行 apply 或者在自定义监控面板中抉择编辑模式进行导入:

正确导入后:

在下面创立的 jupyter notebook 运行深度学习测试工作后,能够显著地察看到相干 GPU 指标变动:

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

重启无奈应用 GPU

对于已部署失常运行的 gpu-operator 和 AI 利用的集群,重启 GPU 主机后会呈现没法用上 GPU 的状况,极有可能是因为插件还没加载,利用优先进行了载入,就会导致这种问题。这时,只须要优先保障插件运行失常,而后重新部署利用即可。

GPU-Operator 常见问题

GPU-Operator 重启后无奈应用

答:对于已部署失常运行的 gpu-operator 和 AI 利用的集群,重启 GPU 主机后会呈现没法用上 GPU 的状况,极有可能是因为插件还没加载,利用优先进行了载入,就会导致这种问题。这时,只须要优先保障插件运行失常,而后重新部署利用即可。

Nvidia k8s-device-plugin 与 GPU-Operator 计划比照?

我之前针对 GPU 应用的是 https://github.com/NVIDIA/k8s… 和 https://github.com/NVIDIA/gpu… 相结合的计划来监控 GPU,请问这个计划与 GPU-Operator 的计划相比,孰优孰劣一些?

答:集体认为 GPU-Operator 更简略易用,其自带 GPU 注入能力不须要构建专用的 OS,并且反对节点发现与可插拔,可能自动化集成治理 GPU 所需的 NVIDIA 设施组件,相对来说还是很省事的。

有没有 KubeSphere 自定义监控的具体应用教程?

答:能够参考 KubeSphere 官网文档来应用自定义监控。

参考资料

官网代码仓库

  • GitHub: https://github.com/NVIDIA/gpu…
  • GitLab: https://gitlab.com/nvidia/kub…

官网文档

  • GPU-Operator 疾速入门:https://docs.nvidia.com/datac…
  • GPU-Operator 离线装置指南:https://docs.nvidia.com/datac…
  • KubeSphere 自定义监控应用文档:https://kubesphere.com.cn/doc…

本文由博客一文多发平台 OpenWrite 公布!

退出移动版