如何在kubernetes环境中共享GPU-数字星空

随着人工智能和大模型的快速发展，云上GPU资源共享变得必要，因为它可以降低硬件成本，提升资源利用效率，并满足模型训练和推理对大规模并行计算的需求。

在kubernetes内置的资源调度功能中，GPU调度只能根据“核数”进行调度，但是深度学习等算法程序执行过程中，资源占用比较高的是显存，这样就形成了很多的资源浪费。

目前的GPU资源共享方案有两种。一种是将一个真正的GPU分解为多个虚拟GPU，即vGPU，这样就可以基于vGPU的数量进行调度；另一种是根据GPU的显存进行调度。

本文将讲述如何安装kubernetes组件实现根据GPU显存调度资源。

系统信息

系统：centos stream8
内核：4.18.0-490.el8.x86_64
驱动：NVIDIA-Linux-x86_64-470.182.03
docker：20.10.24
kubernetes版本：1.24.0

1. 驱动安装

请登录nvida官网自行安装：https://www.nvidia.com/Download/index.aspx?lang=en-us

2. docker安装

请自行安装docker或其他容器运行时，如果使用其他容器运行时，第三步配置请参考NVIDA官网 https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html#installation-guide

注意：官方支持docker、containerd、podman，但本文档只验证过docker的使用，如果使用其他容器运行时，请注意差异性。

3. NVIDIA Container Toolkit 安装

设置仓库与GPG Key

distribution=$(. /etc/os-release;echo $ID$VERSION_ID) \
   && curl -s -L https://nvidia.github.io/libnvidia-container/$distribution/libnvidia-container.repo | sudo tee /etc/yum.repos.d/nvidia-container-toolkit.repo

开始安装

sudo dnf clean expire-cache --refresh
sudo dnf install -y nvidia-container-toolkit

修改docker配置文件添加容器运行时实现

sudo nvidia-ctk runtime configure --runtime=docker

修改/etc/docker/daemon.json,设置nvidia为默认容器运行时（必需）

{
    "default-runtime": "nvidia",
    "runtimes": {
        "nvidia": {
            "path": "/usr/bin/nvidia-container-runtime",
            "runtimeArgs": []
        }
    }
}

重启docker并开始验证是否生效

sudo systemctl restart docker
sudo docker run --rm --runtime=nvidia --gpus all nvidia/cuda:11.6.2-base-ubuntu20.04 nvidia-smi

如果返回如下数据，说明配置成功

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.51.06    Driver Version: 450.51.06    CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla T4            On   | 00000000:00:1E.0 Off |                    0 |
| N/A   34C    P8     9W /  70W |      0MiB / 15109MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

4. 安装K8S GPU调度器

首先执行以下yaml，部署调度器

# rbac.yaml
---
kind: ClusterRole
apiVersion: rbac.authorization.k8s.io/v1
metadata:
  name: gpushare-schd-extender
rules:
  - apiGroups:
      - ""
    resources:
      - nodes
    verbs:
      - get
      - list
      - watch
  - apiGroups:
      - ""
    resources:
      - events
    verbs:
      - create
      - patch
  - apiGroups:
      - ""
    resources:
      - pods
    verbs:
      - update
      - patch
      - get
      - list
      - watch
  - apiGroups:
      - ""
    resources:
      - bindings
      - pods/binding
    verbs:
      - create
  - apiGroups:
      - ""
    resources:
      - configmaps
    verbs:
      - get
      - list
      - watch
---
apiVersion: v1
kind: ServiceAccount
metadata:
  name: gpushare-schd-extender
  namespace: kube-system
---
kind: ClusterRoleBinding
apiVersion: rbac.authorization.k8s.io/v1
metadata:
  name: gpushare-schd-extender
  namespace: kube-system
roleRef:
  apiGroup: rbac.authorization.k8s.io
  kind: ClusterRole
  name: gpushare-schd-extender
subjects:
  - kind: ServiceAccount
    name: gpushare-schd-extender
    namespace: kube-system

# deployment yaml
---
kind: Deployment
apiVersion: apps/v1
metadata:
  name: gpushare-schd-extender
  namespace: kube-system
spec:
  replicas: 1
  strategy:
    type: Recreate
  selector:
    matchLabels:
      app: gpushare
      component: gpushare-schd-extender
  template:
    metadata:
      labels:
        app: gpushare
        component: gpushare-schd-extender
      annotations:
        scheduler.alpha.kubernetes.io/critical-pod: ''
    spec:
      hostNetwork: true
      tolerations:
        - effect: NoSchedule
          operator: Exists
          key: node-role.kubernetes.io/master
        - effect: NoSchedule
          key: node-role.kubernetes.io/control-plane
          operator: Exists
        - effect: NoSchedule
          operator: Exists
          key: node.cloudprovider.kubernetes.io/uninitialized
      nodeSelector:
        node-role.kubernetes.io/control-plane: ""
      serviceAccount: gpushare-schd-extender
      containers:
        - name: gpushare-schd-extender
          image: registry.cn-hangzhou.aliyuncs.com/acs/k8s-gpushare-schd-extender:1.11-d170d8a
          env:
            - name: LOG_LEVEL
              value: debug
            - name: PORT
              value: "12345"

# service.yaml
---
apiVersion: v1
kind: Service
metadata:
  name: gpushare-schd-extender
  namespace: kube-system
  labels:
    app: gpushare
    component: gpushare-schd-extender
spec:
  type: NodePort
  ports:
    - port: 12345
      name: http
      targetPort: 12345
      nodePort: 32766
  selector:
    # select app=ingress-nginx pods
    app: gpushare
    component: gpushare-schd-extender

在/etc/kubernetes目录下添加调度策略配置文件

#scheduler-policy-config.yaml
---
apiVersion: kubescheduler.config.k8s.io/v1beta2
kind: KubeSchedulerConfiguration
clientConnection:
  kubeconfig: /etc/kubernetes/scheduler.conf
extenders:
    # 不知道为什么不支持svc的方式调用，必须用nodeport
  - urlPrefix: "http://gpushare-schd-extender.kube-system:12345/gpushare-scheduler"
    filterVerb: filter
    bindVerb: bind
    enableHTTPS: false
    nodeCacheCapable: true
    managedResources:
      - name: aliyun.com/gpu-mem
        ignoredByScheduler: false
    ignorable: false

上面的 http://gpushare-schd-extender.kube-system:12345 注意要替换为你本地部署的{nodeIP}:{gpushare-schd-extender的nodeport端口}，否则会访问不到

查询命令如下：

kubectl get service gpushare-schd-extender -n kube-system -o jsonpath='{.spec.ports[?(@.name=="http")].nodePort}'

修改kubernetes调度配置 /etc/kubernetes/manifests/kube-scheduler.yaml

1. 在commond中添加
 - --config=/etc/kubernetes/scheduler-policy-config.yaml

2. 添加pod挂载目录
在volumeMounts:中添加
- mountPath: /etc/kubernetes/scheduler-policy-config.yaml
  name: scheduler-policy-config
  readOnly: true
在volumes:中添加
- hostPath:
      path: /etc/kubernetes/scheduler-policy-config.yaml
      type: FileOrCreate
  name: scheduler-policy-config

注意：这里千万不要改错，否则可能会出现莫名其妙的错误
示例如下：

配置rbac及安装device插件

kubectl create -f https://raw.githubusercontent.com/AliyunContainerService/gpushare-device-plugin/master/device-plugin-rbac.yaml
kubectl create -f https://raw.githubusercontent.com/AliyunContainerService/gpushare-device-plugin/master/device-plugin-ds.yaml

5. 在GPU节点上添加标签

kubectl label node <target_node> gpushare=true

6. 安装kubectl Gpu 插件

cd /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

7. 验证

使用kubectl查询GPU资源使用情况

# kubectl inspect gpushare
NAME                                IPADDRESS     GPU0(Allocated/Total)  GPU Memory(GiB)
cn-shanghai.i-uf61h64dz1tmlob9hmtb  192.168.0.71  6/15                   6/15
cn-shanghai.i-uf61h64dz1tmlob9hmtc  192.168.0.70  3/15                   3/15
------------------------------------------------------------------------------
Allocated/Total GPU Memory In Cluster:
9/30 (30%)

创建一个有GPU需求的资源，查看其资源调度情况

apiVersion: apps/v1
kind: Deployment
metadata:
  name: binpack-1
  labels:
    app: binpack-1
spec:
  replicas: 1
  selector: # define how the deployment finds the pods it manages
    matchLabels:
      app: binpack-1
  template: # define the pods specifications
    metadata:
      labels:
        app: binpack-1
    spec:
      tolerations:
        - effect: NoSchedule
          key: cloudClusterNo
          operator: Exists        
      containers:
        - name: binpack-1
          image: cheyang/gpu-player:v2
          resources:
            limits:
              # 单位GiB
              aliyun.com/gpu-mem: 3

8. 问题排查

如果在安装过程中发现资源未安装成功，可以通过pod查看日志

kubectl get po -n kube-system -o=wide | grep gpushare-device 
kubecl logs -n kube-system <pod_name>

参考地址：
NVIDA官网container-toolkit安装文档: https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html#docker
阿里云GPU插件安装：https://github.com/AliyunContainerService/gpushare-scheduler-extender/blob/master/docs/install.md