使用 EKS Auto Mode 和 Karpenter 管理 AI/ML 工作負載的運算

如需完整清單，請參閱 EKS Auto Mode 支援的標籤。

如需完整清單，請參閱 Karpenter Well-Known 標籤。

以下範例顯示使用預設 EKS Auto Mode NodeClass 的靜態容量 NodePool，並建立具有 4 個節點 (replicas) 的靜態 NodePool，最多可達 6 個節點 ()limits.nodes。

apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: gpu-static-inference
spec:
  replicas: 4
  template:
    spec:
      nodeClassRef:
        group: eks.amazonaws.com
        kind: NodeClass
        name: default
      requirements:
        - key: "karpenter.sh/capacity-type"
          operator: In
          values: ["on-demand"]
        - key: "karpenter.k8s.aws/instance-family"
          operator: In
          values: ["g6e"]
        - key: "topology.kubernetes.io/zone"
          operator: In
          values: ["us-east-1a"]
      taints:
        - key: nvidia.com/gpu
          value: "true"
          effect: NoSchedule
  limits:
    nodes: 6   # Allow temporary headroom during node replacement

使用自我管理的 Karpenter 時，靜態容量由 alpha StaticCapacity功能 （在 Karpenter 1.8 版中啟動） 鎖定，必須在 Helm 值中啟用：

settings:
  featureGates:
    staticCapacity: true

NodePool 參考EC2NodeClass名為 的自訂my-nodeclass，並建立具有 4 個節點 (replicas) 的靜態 NodePool，最多可達 6 個節點 ()limits.nodes。

apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: gpu-static-inference
spec:
  replicas: 4
  template:
    spec:
      nodeClassRef:
        group: karpenter.k8s.aws
        kind: EC2NodeClass
        name: my-nodeclass
      requirements:
        - key: "karpenter.sh/capacity-type"
          operator: In
          values: ["on-demand"]
        - key: "karpenter.k8s.aws/instance-family"
          operator: In
          values: ["g6e"]
        - key: "topology.kubernetes.io/zone"
          operator: In
          values: ["us-east-1a"]
      taints:
        - key: nvidia.com/gpu
          value: "true"
          effect: NoSchedule
  limits:
    nodes: 6   # Allow temporary headroom during node replacement

建立參考容量區塊保留NodeClass的 ，然後建立使用它的 NodePool。

consolidateAfter: Never 設定後，Karpenter 不會嘗試取代、合併或終止節點，以更有效地降低成本或封裝工作負載。這建議用於容量區塊，因為容量已預付。

apiVersion: eks.amazonaws.com/v1
kind: NodeClass
metadata:
  name: capacity-block-gpu
spec:
  capacityReservationSelectorTerms:
    - id: "cr-0123456789abcdef0"   # Your Capacity Block reservation ID
    # Alternative: select by tags
    # - tags:
    #     role: "production-inference"
    #   owner: "012345678901"
---
apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: gpu-capacity-block
spec:
  disruption:
    consolidationPolicy: WhenEmpty
    consolidateAfter: Never
  template:
    spec:
      nodeClassRef:
        group: eks.amazonaws.com
        kind: NodeClass
        name: capacity-block-gpu
      requirements:
        - key: "karpenter.sh/capacity-type"
          operator: In
          values: ["reserved"]
        - key: "karpenter.k8s.aws/instance-family"
          operator: In
          values: ["p5", "p5e", "p5en", "p4d"]
      taints:
        - key: nvidia.com/gpu
          value: "true"
          effect: NoSchedule

除了 之外，建立EC2NodeClass包含 AMI、子網路和安全群組選取器的 capacityReservationSelectorTerms，然後建立使用它的 NodePool。

consolidateAfter: Never 設定後，Karpenter 不會嘗試取代、合併或終止節點，以更有效地降低成本或封裝工作負載。這建議用於容量區塊，因為容量已預付。

apiVersion: karpenter.k8s.aws/v1
kind: EC2NodeClass
metadata:
  name: capacity-block-gpu
spec:
  amiSelectorTerms:
    - alias: al2023@latest
  subnetSelectorTerms:
    - tags:
        karpenter.sh/discovery: ml-cluster
  securityGroupSelectorTerms:
    - tags:
        karpenter.sh/discovery: ml-cluster
  capacityReservationSelectorTerms:
    - id: "cr-0123456789abcdef0" # Your Capacity Block reservation ID
    # Alternative: select by tags
    # - tags:
    #     role: "production-inference"
    #   owner: "012345678901"
---
apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: gpu-capacity-block
spec:
  disruption:
    consolidationPolicy: WhenEmpty
    consolidateAfter: Never
  template:
    spec:
      nodeClassRef:
        group: karpenter.k8s.aws
        kind: EC2NodeClass
        name: capacity-block-gpu
      requirements:
        - key: "karpenter.sh/capacity-type"
          operator: In
          values: ["reserved"]
        - key: "karpenter.k8s.aws/instance-family"
          operator: In
          values: ["p5", "p5e", "p5en", "p4d"]
      taints:
        - key: nvidia.com/gpu
          value: "true"
          effect: NoSchedule

NodeClass 使用 capacityReservationSelectorTerms 和以on-demand備用reserved為優先順序的 NodePool 建立 。釘topology.kubernetes.io/zone選至 ODCR 的 AZ：

apiVersion: eks.amazonaws.com/v1
kind: NodeClass
metadata:
  name: odcr-gpu-production
spec:
  capacityReservationSelectorTerms:
    - id: "cr-0987654321fedcba0"
    # Alternative: select by tags
    # - tags:
    #     Purpose: "production-inference"
    #   owner: "012345678901"
---
apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: gpu-reserved-production
spec:
  disruption:
    consolidationPolicy: WhenEmpty
    consolidateAfter: Never
  template:
    spec:
      nodeClassRef:
        group: eks.amazonaws.com
        kind: NodeClass
        name: odcr-gpu-production
      requirements:
        - key: "karpenter.sh/capacity-type"
          operator: In
          values: ["reserved", "on-demand"]
        - key: "karpenter.k8s.aws/instance-family"
          operator: In
          values: ["p5", "g6e"]
        - key: "topology.kubernetes.io/zone"
          operator: In
          values: ["us-east-1a"]
      taints:
        - key: nvidia.com/gpu
          value: "true"
          effect: NoSchedule

除了 之外，EC2NodeClass使用 AMI、子網路和安全群組選取器建立 capacityReservationSelectorTerms，然後建立 NodePool：

apiVersion: karpenter.k8s.aws/v1
kind: EC2NodeClass
metadata:
  name: odcr-gpu-production
spec:
  amiSelectorTerms:
    - alias: al2023@latest
  subnetSelectorTerms:
    - tags:
        karpenter.sh/discovery: ml-cluster
  securityGroupSelectorTerms:
    - tags:
        karpenter.sh/discovery: ml-cluster
  capacityReservationSelectorTerms:
    - id: "cr-0987654321fedcba0"
---
apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: gpu-reserved-production
spec:
  disruption:
    consolidationPolicy: WhenEmpty
    consolidateAfter: Never
  template:
    spec:
      nodeClassRef:
        group: karpenter.k8s.aws
        kind: EC2NodeClass
        name: odcr-gpu-production
      requirements:
        - key: "karpenter.sh/capacity-type"
          operator: In
          values: ["reserved", "on-demand"]
        - key: "karpenter.k8s.aws/instance-family"
          operator: In
          values: ["p5", "g6e"]
        - key: "topology.kubernetes.io/zone"
          operator: In
          values: ["us-east-1a"]
      taints:
        - key: nvidia.com/gpu
          value: "true"
          effect: NoSchedule

apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: gpu-ondemand
spec:
  template:
    spec:
      nodeClassRef:
        group: eks.amazonaws.com
        kind: NodeClass
        name: default
      requirements:
        - key: "karpenter.sh/capacity-type"
          operator: In
          values: ["on-demand"]
        - key: "karpenter.k8s.aws/instance-family"
          operator: In
          values: ["g6", "g6e", "g7e"]
        - key: "karpenter.k8s.aws/instance-gpu-manufacturer"
          operator: In
          values: ["nvidia"]
      taints:
        - key: nvidia.com/gpu
          value: "true"
          effect: NoSchedule

apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: gpu-ondemand
spec:
  template:
    spec:
      nodeClassRef:
        group: karpenter.k8s.aws
        kind: EC2NodeClass
        name: default
      requirements:
        - key: "karpenter.sh/capacity-type"
          operator: In
          values: ["on-demand"]
        - key: "karpenter.k8s.aws/instance-family"
          operator: In
          values: ["g6", "g6e", "g7e"]
        - key: "karpenter.k8s.aws/instance-gpu-manufacturer"
          operator: In
          values: ["nvidia"]
      taints:
        - key: nvidia.com/gpu
          value: "true"
          effect: NoSchedule

EKS Auto 模式會以原生方式處理 Spot 中斷。不需要 SQS 佇列或節點終止處理常式。

apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: gpu-spot
spec:
  disruption:
    budgets:
      - nodes: 10%
    consolidationPolicy: WhenEmpty
    consolidateAfter: 1h
  template:
    spec:
      nodeClassRef:
        group: eks.amazonaws.com
        kind: NodeClass
        name: default
      requirements:
        - key: "karpenter.sh/capacity-type"
          operator: In
          values: ["spot"]
        - key: "karpenter.k8s.aws/instance-family"
          operator: In
          values: ["g6", "g6e", "g7e"]
        - key: "karpenter.k8s.aws/instance-gpu-manufacturer"
          operator: In
          values: ["nvidia"]
      taints:
        - key: nvidia.com/gpu
          value: "true"
          effect: NoSchedule
  limits:
    resources:
      nvidia.com/gpu: "64"

自我管理的 Karpenter 要求您透過設定中斷佇列，在 Karpenter 控制器 （而非 NodePool) 上啟用原生中斷處理：接收 EC2 Spot 中斷和重新平衡建議事件的 SQS 佇列。您可以在安裝時設定一次。

如果您直接使用 Helm 安裝 Karpenter，請在 settings.interruptionQueue中設定 values.yaml：

# karpenter values.yaml (Helm)
settings:
  clusterName: my-cluster
  interruptionQueue: my-queue   # Name of the SQS queue receiving Spot events

如果您使用 引導 Karpentereksctl，withSpotInterruptionQueue: true請在叢集組態檔案中設定 。 eksctl會建立 SQS 佇列和 EventBridge 規則，並將 Karpenter 控制器設定為使用它們。

# eksctl ClusterConfig
karpenter:
  version: "${KARPENTER_VERSION}"
  withSpotInterruptionQueue: true

一旦控制器設定為使用您的佇列，個別 NodePool 資源就不需要額外的組態。中斷處理會套用至整個叢集。

apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: gpu-spot
spec:
  disruption:
    budgets:
      - nodes: 10%
    consolidationPolicy: WhenEmpty
    consolidateAfter: 1h
  template:
    spec:
      nodeClassRef:
        group: karpenter.k8s.aws
        kind: EC2NodeClass
        name: default
      requirements:
        - key: "karpenter.sh/capacity-type"
          operator: In
          values: ["spot"]
        - key: "karpenter.k8s.aws/instance-family"
          operator: In
          values: ["g6", "g6e", "g7e"]
        - key: "karpenter.k8s.aws/instance-gpu-manufacturer"
          operator: In
          values: ["nvidia"]
      taints:
        - key: nvidia.com/gpu
          value: "true"
          effect: NoSchedule
  limits:
    resources:
      nvidia.com/gpu: "64"