kubernetes集群使用GPU及安装kubeflow1.0.RC操作步骤
Kubeflow使用场景
-
希望训练tensorflow模型且可以使用模型接口发布应用服务在k8s环境中(eg.local,prem,cloud)
-
希望使用Jupyter notebooks来调试代码，多用户的notebook server
-
在训练的Job中，需要对的CPU或者GPU资源进行调度编排
-
希望Tensorflow和其他组件进行组合来发布服务
依赖库
-
ksonnet 0.11.0以上版本 /可以直接从github上下载，scp ks文件到usr/local/bin
-
kubernetes 1.8以上（直接使用CCE服务节点，需要创建一个CCE集群和若干节点，并为某个节点绑定EIP）
-
kubectl tools
1、安装ksonnet
ksonnet 安装过程，可以去网址里面查看ks最新版本
wget https://github.com/ksonnet/ksonnet/releases/download/v0.13.0/ks_0.13.0_linux_amd64.tar.gz
tar -vxf ks_0.13.0_linux_amd64.tar.gz
cd -vxf ks_0.13.0_linux_amd64
sudo cp ks /usr/local/bin
安装完成后
安装显卡驱动
sudo yum-config-manager --add-repo http://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo
sudo yum clean all
sudo yum -y install nvidia-driver-latest-dkms cuda
sudo yum -y install cuda-drivers
如缺少gcc依赖，则实行如下命令
yum install kernel-devel kernel-doc kernel-headers gcc\* glibc\* glibc-\*
rpm --import https://www.elrepo.org/RPM-GPG-KEY-elrepo.orgrpm -Uvh http://www.elrepo.org/elrepo-release-7.0-3.el7.elrepo.noarch.rpmyum install -y kmod-nvidia
###在GRUB_CMDLINE_LINUX添加 rdblacklist=nouveau 项
echo -e "blacklist nouveau\noptions nouveau modeset=0" > /etc/modprobe.d/blacklist.conf
重启，查看nouveau是否被禁用成功
lsmod|grep nouv
没有任何输出，则表示nouveau已被禁用
查看服务器显卡信息
[root@master ~]# nvidia-smi
Tue Jan 14 03:46:41 2020
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 440.44 Driver Version: 440.44 CUDA Version: 10.2 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 Tesla T4 Off | 00000000:18:00.0 Off | 0 |
| N/A 29C P8 10W / 70W | 0MiB / 15109MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
| 1 Tesla T4 Off | 00000000:86:00.0 Off | 0 |
| N/A 25C P8 9W / 70W | 0MiB / 15109MiB | 0% Default |
+-------------------------------+----------------------+----------------------++-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+安装NVIDIA-DOCKER
curl -s -L https://nvidia.github.io/nvidia-docker/centos7/x86_64/nvidia-docker.repo | sudo tee /etc/yum.repos.d/nvidia-docker.repo
- 查找NVIDIAdocker版本
yum search --showduplicates nvidia-docker
- 安装NVIDIA-docker
docker版本为docker18.09.7.ce，所以安装下述NVIDIAdocker版本
yum install -y nvidia-docker2
pkill -SIGHUP dockerd
nvidia-docker version 可查看已安装的nvidia docker版本
修改docker runtimes为nvidia-docker
[root@ks-allinone ~]# cat /etc/docker/daemon.json
{"default-runtime": "nvidia","runtimes": {"nvidia": {"path": "nvidia-container-runtime","runtimeArgs": []}},"registry-mirrors": ["https://o96k4rm0.mirror.aliyuncs.com"]
}重启docker及k8s
systemctl daemon-reload
systemctl restart docker.service
systemctl restart kubelet
cd /etc/kubernetes/
curl -O https://raw.githubusercontent.com/AliyunContainerService/gpushare-scheduler-extender/master/config/scheduler-policy-config.json
cd /tmp/
curl -O https://raw.githubusercontent.com/AliyunContainerService/gpushare-scheduler-extender/master/config/gpushare-schd-extender.yaml
kubectl create -f gpushare-schd-extender.yaml
kubectl create -f device-plugin-rbac.yaml
# rbac.yaml
---
kind: ClusterRole
apiVersion: rbac.authorization.k8s.io/v1
metadata:name: gpushare-device-plugin
rules:
- apiGroups:- ""resources:- nodesverbs:- get- list- watch
- apiGroups:- ""resources:- eventsverbs:- create- patch
- apiGroups:- ""resources:- podsverbs:- update- patch- get- list- watch
- apiGroups:- ""resources:- nodes/statusverbs:- patch- update
---
apiVersion: v1
kind: ServiceAccount
metadata:name: gpushare-device-pluginnamespace: kube-system
---
kind: ClusterRoleBinding
apiVersion: rbac.authorization.k8s.io/v1
metadata:name: gpushare-device-pluginnamespace: kube-system
roleRef:apiGroup: rbac.authorization.k8s.iokind: ClusterRolename: gpushare-device-plugin
subjects:
- kind: ServiceAccountname: gpushare-device-pluginnamespace: kube-system
kubectl create -f device-plugin-ds.yamlapiVersion: extensions/v1beta1
kind: DaemonSet
metadata:name: gpushare-device-plugin-dsnamespace: kube-system
spec:template:metadata:annotations:scheduler.alpha.kubernetes.io/critical-pod: ""labels:component: gpushare-device-pluginapp: gpusharename: gpushare-device-plugin-dsspec:serviceAccount: gpushare-device-pluginhostNetwork: truenodeSelector:gpushare: "true"containers:- image: registry.cn-hangzhou.aliyuncs.com/acs/k8s-gpushare-plugin:v2-1.11-35eccabname: gpushare# Make this pod as Guaranteed pod which will never be evicted because of node's resource consumption.command:- gpushare-device-plugin-v2- -logtostderr- --v=5#- --memory-unit=Miresources:limits:memory: "300Mi"cpu: "1"requests:memory: "300Mi"cpu: "1"env:- name: KUBECONFIGvalue: /etc/kubernetes/kubelet.conf- name: NODE_NAMEvalueFrom:fieldRef:fieldPath: spec.nodeNamesecurityContext:allowPrivilegeEscalation: falsecapabilities:drop: ["ALL"]volumeMounts:- name: device-pluginmountPath: /var/lib/kubelet/device-pluginsvolumes:- name: device-pluginhostPath:path: /var/lib/kubelet/device-plugins参考
https://github.com/AliyunContainerService/gpushare-scheduler-extender
https://github.com/AliyunContainerService/gpushare-device-plugin
为共享节点打上gpushare标签
kubectl label node mynode gpushare=true
安装扩展
curl -LO https://storage.googleapis.com/kubernetes-release/release/v1.12.1/bin/linux/amd64/kubectl
chmod +x ./kubectl
sudo mv ./kubectl /usr/bin/kubectlcd /usr/bin/
wget https://github.com/AliyunContainerService/gpushare-device-plugin/releases/download/v0.3.0/kubectl-inspect-gpushare
chmod u+x /usr/bin/kubectl-inspect-gpushare
kubectl inspect gpushare ##查看集群GPU使用情况
wget https://raw.githubusercontent.com/google/metallb/v0.7.3/manifests/metallb.yamlkubectl apply -f metallb.yaml
metallb-config.yaml
apiVersion: v1
kind: ConfigMap
metadata:namespace: metallb-systemname: config
data:config: |address-pools:- name: defaultprotocol: layer2addresses:- 10.18.5.30-10.18.5.50kubectl apply -f metallb-config.yaml
kubectl apply -f tensorflow.yaml
apiVersion: extensions/v1beta1
kind: Deployment
metadata:name: tensorflow-gpu
spec:replicas: 1template:metadata:labels:name: tensorflow-gpuspec:containers:- name: tensorflow-gpuimage: tensorflow/tensorflow:1.15.0-py3-jupyterimagePullPolicy: Neverresources:limits:aliyun.com/gpu-mem: 1024ports:- containerPort: 8888
---
apiVersion: v1
kind: Service
metadata:name: tensorflow-gpu
spec:ports:- port: 8888targetPort: 8888nodePort: 30888name: jupyterselector:name: tensorflow-gputype: NodePort
查看集群GPU使用情况
[root@master ~]# kubectl inspect gpushare
NAME IPADDRESS GPU0(Allocated/Total) GPU1(Allocated/Total) GPU Memory(MiB)
master 10.18.5.20 1024/15109 0/15109 1024/30218
node 10.18.5.21 0/15109 0/15109 0/30218
------------------------------------------------------------------
Allocated/Total GPU Memory In Cluster:
1024/60436 (1%)
[root@master ~]#
可通过动态伸缩tensorflow service 的节点数量以及修改单个节点的显存大小测试GPU使用情况
kubectl scale --current-replicas=1 --replicas=100 deployment/tensorflow-gpu
经测试，得出以下测试结果：
环境
| 节点 | GPU个数 | GPU内存 |
|---|---|---|
| master | 2 | 15109M*2=30218M |
| node | 2 | 15109M*2=30218M |
测试结果
| podGpu | pod个数 | gpu利用率 |
|---|---|---|
| 256M | 183 | 77% |
| 512M | 116 | 98% |
| 1024M | 56 | 94% |
安装kubeflow（V1.0.RC）
tar -vxf ks_0.12.0_linux_amd64.tar.gzcp ks_0.12.0_linux_amd64/* /usr/local/bin/
kfctl_v1.0-rc.3-1-g24b60e8_linux.tar.gz
tar -zxvf kfctl_v1.0-rc.3-1-g24b60e8_linux.tar.gz
cp kfctl /usr/bin/
准备工作，创建PV及PVC，使用NFS作为文件存储
创建storageclass
apiVersion: storage.k8s.io/v1
kind: StorageClass
metadata:name: local-pathnamespace: kubeflow
#provisioner: example.com/nfs
provisioner: kubernetes.io/gce-pd
parameters:type: pd-ssdkubectl create -f storage.yml
yum install nfs-utils rpcbind
#创建NFS挂载目录（至少需要四个）
mkdir -p /data/nfsvim /etc/exports
#添加上面的挂载目录
/data/nfs 192.168.122.0/24(rw,sync)systemctl restart nfs-server.service
创建PV，因多个pod挂载文件可能重名，所以最好创建多个PV由pod选择挂载（至少4个，分别供katib-mysql,metadata-mysql,minio,mysql挂载）
[root@master pv]# cat mysql-pv.yml
apiVersion: v1
kind: PersistentVolume
metadata:name: local-path #不同的PVC需要修改
spec:capacity:storage: 200GiaccessModes:- ReadWriteOncepersistentVolumeReclaimPolicy: RecyclestorageClassName: local-pathnfs:path: /data/nfs #不同的PVC需要修改server: 10.18.5.20
创建命名空间 kubeflow-anonymous
kubectl create namespace kubeflow-anonymous
下载kubeflow1.0.RC yml文件, https://github.com/kubeflow/manifests/blob/v1.0-branch/kfdef/kfctl_k8s_istio.yaml
[root@master 2020-0219]# cat kfctl_k8s_istio.yaml
apiVersion: kfdef.apps.kubeflow.org/v1
kind: KfDef
metadata:clusterName: kubernetescreationTimestamp: nullname: 2020-0219namespace: kubeflow
spec:applications:- kustomizeConfig:parameters:- name: namespacevalue: istio-systemrepoRef:name: manifestspath: istio/istio-crdsname: istio-crds- kustomizeConfig:parameters:- name: namespacevalue: istio-systemrepoRef:name: manifestspath: istio/istio-installname: istio-install- kustomizeConfig:parameters:- name: namespacevalue: istio-systemrepoRef:name: manifestspath: istio/cluster-local-gatewayname: cluster-local-gateway- kustomizeConfig:parameters:- name: clusterRbacConfigvalue: "OFF"repoRef:name: manifestspath: istio/istioname: istio- kustomizeConfig:parameters:- name: namespacevalue: istio-systemrepoRef:name: manifestspath: istio/add-anonymous-user-filtername: add-anonymous-user-filter- kustomizeConfig:repoRef:name: manifestspath: application/application-crdsname: application-crds- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: application/applicationname: application- kustomizeConfig:parameters:- name: namespacevalue: cert-managerrepoRef:name: manifestspath: cert-manager/cert-manager-crdsname: cert-manager-crds- kustomizeConfig:parameters:- name: namespacevalue: kube-systemrepoRef:name: manifestspath: cert-manager/cert-manager-kube-system-resourcesname: cert-manager-kube-system-resources- kustomizeConfig:overlays:- self-signed- applicationparameters:- name: namespacevalue: cert-managerrepoRef:name: manifestspath: cert-manager/cert-managername: cert-manager- kustomizeConfig:repoRef:name: manifestspath: metacontrollername: metacontroller- kustomizeConfig:overlays:- istio- applicationrepoRef:name: manifestspath: argoname: argo- kustomizeConfig:repoRef:name: manifestspath: kubeflow-rolesname: kubeflow-roles- kustomizeConfig:overlays:- istio- applicationrepoRef:name: manifestspath: common/centraldashboardname: centraldashboard- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: admission-webhook/bootstrapname: bootstrap- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: admission-webhook/webhookname: webhook- kustomizeConfig:overlays:- istio- applicationparameters:- name: userid-headervalue: kubeflow-useridrepoRef:name: manifestspath: jupyter/jupyter-web-appname: jupyter-web-app- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: spark/spark-operatorname: spark-operator- kustomizeConfig:overlays:- istio- application- dbrepoRef:name: manifestspath: metadataname: metadata- kustomizeConfig:overlays:- istio- applicationrepoRef:name: manifestspath: jupyter/notebook-controllername: notebook-controller- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: pytorch-job/pytorch-job-crdsname: pytorch-job-crds- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: pytorch-job/pytorch-operatorname: pytorch-operator- kustomizeConfig:overlays:- applicationparameters:- name: usageIdvalue: <randomly-generated-id>- name: reportUsagevalue: "true"repoRef:name: manifestspath: common/spartakusname: spartakus- kustomizeConfig:overlays:- istiorepoRef:name: manifestspath: tensorboardname: tensorboard- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: tf-training/tf-job-crdsname: tf-job-crds- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: tf-training/tf-job-operatorname: tf-job-operator- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: katib/katib-crdsname: katib-crds- kustomizeConfig:overlays:- application- istiorepoRef:name: manifestspath: katib/katib-controllername: katib-controller- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: pipeline/api-servicename: api-service- kustomizeConfig:overlays:- applicationparameters:- name: minioPvcNamevalue: minio-pv-claimrepoRef:name: manifestspath: pipeline/minioname: minio- kustomizeConfig:overlays:- applicationparameters:- name: mysqlPvcNamevalue: mysql-pv-claimrepoRef:name: manifestspath: pipeline/mysqlname: mysql- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: pipeline/persistent-agentname: persistent-agent- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: pipeline/pipelines-runnername: pipelines-runner- kustomizeConfig:overlays:- istio- applicationrepoRef:name: manifestspath: pipeline/pipelines-uiname: pipelines-ui- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: pipeline/pipelines-viewername: pipelines-viewer- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: pipeline/scheduledworkflowname: scheduledworkflow- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: pipeline/pipeline-visualization-servicename: pipeline-visualization-service- kustomizeConfig:overlays:- application- istioparameters:- name: adminvalue: johnDoe@acme.comrepoRef:name: manifestspath: profilesname: profiles- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: seldon/seldon-core-operatorname: seldon-core-operator- kustomizeConfig:overlays:- applicationparameters:- name: namespacevalue: knative-servingrepoRef:name: manifestspath: knative/knative-serving-crdsname: knative-crds- kustomizeConfig:overlays:- applicationparameters:- name: namespacevalue: knative-servingrepoRef:name: manifestspath: knative/knative-serving-installname: knative-install- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: kfserving/kfserving-crdsname: kfserving-crds- kustomizeConfig:overlays:- applicationrepoRef:name: manifestspath: kfserving/kfserving-installname: kfserving-installrepos:- name: manifestsuri: https://github.com/kubeflow/manifests/archive/master.tar.gzversion: master
status:reposCache:- localPath: '"../.cache/manifests/manifests-master"'name: manifests
[root@master 2020-0219]##进入你的kubeflowapp目录 执行
kfctl apply -V -f kfctl_k8s_istio.yaml
#安装过程中需要从GitHub下载配置文件，可能会失败，失败时重试
在kubeflowapp平级目录下会生成kustomize文件夹，为防重启时镜像拉取失败，需修改所有镜像拉取策略为IfNotPresent
然后再次执行 kfctl apply -V -f kfctl_k8s_istio.yaml
查看运行状态
kubectl get all -n kubeflow
#修改ingeress-gateway访问方式为LoadBalancer
kubectl -n istio-system edit svc istio-ingressgateway
#修改此处为LoadBalancer
selector:app: istio-ingressgatewayistio: ingressgatewayrelease: istio
sessionAffinity: None
type: LoadBalancer
保存，再次查看该svc信息
[root@master 2020-0219]# kubectl -n istio-system get svc istio-ingressgateway
NAME TYPE CLUSTER-IP EXTERNAL-IP PORT(S) AGE
istio-ingressgateway LoadBalancer 10.98.19.247 10.18.5.30 15020:32230/TCP,80:31380/TCP,443:31390/TCP,31400:31400/TCP,15029:31908/TCP,15030:31864/TCP,15031:31315/TCP,15032:30372/TCP,15443:32631/TCP 42h
[root@master 2020-0219]#
EXTERNAL-IP 即为外部访问地址，访问http://10.18.5.30 即可进入kubeflow主页
关于镜像拉取，gcr镜像国内无法拉取，可以通过如下方式拉取
curl -s https://zhangguanzhang.github.io/bash/pull.sh | bash -s -- 镜像信息
若上述方法也无法拉取，可以使用阿里云手动构建镜像方式使用海外服务器构建
