Kubernetes & Amazon EKS Training

中文

한국어
English
日本語
Español

中文

한국어
English
日本語
Español

Appearance

工作负载特定优化 ​

支持的版本: EKS 1.29+, EKS Auto Mode GA 最后更新: February 19, 2026

本指南介绍如何针对不同工作负载类型优化 EKS Auto Mode 配置，包括 Web 服务、批处理、GPU 工作负载以及 AI/ML 训练。

Web 服务（可用性优先） ​

yaml
# web-service-optimized.yaml
apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: web-tier
spec:
  template:
    metadata:
      labels:
        tier: web
    spec:
      requirements:
        # General-purpose instances
        - key: karpenter.k8s.aws/instance-category
          operator: In
          values: ["m"]
        - key: karpenter.k8s.aws/instance-size
          operator: In
          values: ["large", "xlarge", "2xlarge"]
        # Use only On-Demand (availability first)
        - key: karpenter.sh/capacity-type
          operator: In
          values: ["on-demand"]
      taints:
        - key: tier
          value: web
          effect: NoSchedule
      nodeClassRef:
        group: eks.amazonaws.com
        kind: NodeClass
        name: default
  disruption:
    consolidationPolicy: WhenEmptyOrUnderutilized
    consolidateAfter: 5m
    budgets:
      - nodes: "10%"
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: web-frontend
spec:
  replicas: 10
  selector:
    matchLabels:
      app: web-frontend
  template:
    metadata:
      labels:
        app: web-frontend
    spec:
      tolerations:
        - key: tier
          value: web
          effect: NoSchedule
      nodeSelector:
        tier: web
      affinity:
        podAntiAffinity:
          preferredDuringSchedulingIgnoredDuringExecution:
            - weight: 100
              podAffinityTerm:
                labelSelector:
                  matchLabels:
                    app: web-frontend
                topologyKey: kubernetes.io/hostname
      containers:
        - name: web
          image: my-web-app:latest
          resources:
            requests:
              cpu: 500m
              memory: 512Mi
            limits:
              cpu: 1000m
              memory: 1Gi
          readinessProbe:
            httpGet:
              path: /health
              port: 8080
            initialDelaySeconds: 5
            periodSeconds: 10
          livenessProbe:
            httpGet:
              path: /health
              port: 8080
            initialDelaySeconds: 15
            periodSeconds: 20

Web 服务优化摘要 ​

方面建议理由
容量类型On-Demand高可用性要求
实例系列M-series（通用型）CPU/内存均衡
反亲和性按 hostname分散到不同节点
健康检查同时使用 readiness 和 liveness快速故障检测
PDBminAvailable: N-1在更新期间维持服务

批处理（成本优先，Spot） ​

yaml
# batch-processing-optimized.yaml
apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: batch-tier
spec:
  template:
    metadata:
      labels:
        tier: batch
    spec:
      requirements:
        # Compute-optimized
        - key: karpenter.k8s.aws/instance-category
          operator: In
          values: ["c"]
        - key: karpenter.k8s.aws/instance-size
          operator: In
          values: ["xlarge", "2xlarge", "4xlarge"]
        # Use only Spot (cost first)
        - key: karpenter.sh/capacity-type
          operator: In
          values: ["spot"]
        # Various instance types for better Spot availability
        - key: karpenter.k8s.aws/instance-generation
          operator: In
          values: ["5", "6", "7"]
      taints:
        - key: tier
          value: batch
          effect: NoSchedule
      nodeClassRef:
        group: eks.amazonaws.com
        kind: NodeClass
        name: default
  disruption:
    consolidationPolicy: WhenEmpty
    consolidateAfter: 30s
---
apiVersion: batch/v1
kind: Job
metadata:
  name: data-processing
spec:
  parallelism: 20
  completions: 100
  backoffLimit: 10
  template:
    spec:
      tolerations:
        - key: tier
          value: batch
          effect: NoSchedule
      nodeSelector:
        tier: batch
      restartPolicy: OnFailure
      terminationGracePeriodSeconds: 30
      containers:
        - name: processor
          image: my-batch-processor:latest
          resources:
            requests:
              cpu: 2000m
              memory: 4Gi
            limits:
              cpu: 4000m
              memory: 8Gi
          env:
            - name: SPOT_AWARE
              value: "true"

批处理优化摘要 ​

方面建议理由
容量类型仅 Spot最大化节省成本
实例系列C-series（计算优化型）CPU 密集型工作负载
实例多样性多个世代更好的 Spot 可用性
重启策略OnFailure处理 Spot 中断
整合激进（30s）Job 完成后快速清理

GPU 工作负载（p5, g5） ​

yaml
# gpu-workload-optimized.yaml
apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: gpu-tier
spec:
  template:
    metadata:
      labels:
        tier: gpu
        accelerator: nvidia
    spec:
      requirements:
        # GPU instances
        - key: karpenter.k8s.aws/instance-category
          operator: In
          values: ["g", "p"]
        - key: karpenter.k8s.aws/instance-gpu-manufacturer
          operator: In
          values: ["nvidia"]
        # Specific GPU instance types
        - key: node.kubernetes.io/instance-type
          operator: In
          values: ["g5.xlarge", "g5.2xlarge", "g5.4xlarge", "p5.48xlarge"]
        # On-Demand (GPU Spot availability is low)
        - key: karpenter.sh/capacity-type
          operator: In
          values: ["on-demand"]
      taints:
        - key: nvidia.com/gpu
          value: "true"
          effect: NoSchedule
      nodeClassRef:
        group: eks.amazonaws.com
        kind: NodeClass
        name: gpu-nodeclass
  limits:
    nvidia.com/gpu: 16
  disruption:
    consolidationPolicy: WhenEmpty
    consolidateAfter: 10m  # GPU takes longer to start
---
apiVersion: eks.amazonaws.com/v1
kind: NodeClass
metadata:
  name: gpu-nodeclass
spec:
  amiFamily: AL2023
  blockDeviceMappings:
    - deviceName: /dev/xvda
      ebs:
        volumeSize: 200Gi  # Large volume for model caching
        volumeType: gp3
        iops: 6000
        throughput: 250
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: ml-inference
spec:
  replicas: 2
  selector:
    matchLabels:
      app: ml-inference
  template:
    metadata:
      labels:
        app: ml-inference
    spec:
      tolerations:
        - key: nvidia.com/gpu
          value: "true"
          effect: NoSchedule
      nodeSelector:
        tier: gpu
      containers:
        - name: inference
          image: my-ml-model:latest
          resources:
            limits:
              nvidia.com/gpu: 1
            requests:
              cpu: 4000m
              memory: 16Gi

GPU 实例选择指南 ​

实例GPUGPU 内存使用场景
g5.xlarge1x A10G24GB小型推理
g5.2xlarge1x A10G24GB中型推理
g5.4xlarge1x A10G24GB大型推理
g5.12xlarge4x A10G96GB多模型服务
p5.48xlarge8x H100640GB大规模训练

GPU 优化摘要 ​

方面建议理由
容量类型On-DemandGPU Spot 可用性有限
存储200GB+ gp3模型缓存、检查点
整合宽松（10m）GPU 启动较慢
限制设置 nvidia.com/gpu 限制防止 GPU 成本失控

AI/ML 训练工作负载 ​

yaml
# ml-training-optimized.yaml
apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: ml-training
spec:
  template:
    metadata:
      labels:
        tier: ml-training
    spec:
      requirements:
        # Large-scale GPU instances
        - key: node.kubernetes.io/instance-type
          operator: In
          values: ["p5.48xlarge", "p4d.24xlarge"]
        - key: karpenter.sh/capacity-type
          operator: In
          values: ["on-demand"]
      taints:
        - key: ml-training
          value: "true"
          effect: NoSchedule
      nodeClassRef:
        group: eks.amazonaws.com
        kind: NodeClass
        name: ml-training-nodeclass
  limits:
    nvidia.com/gpu: 64
---
apiVersion: eks.amazonaws.com/v1
kind: NodeClass
metadata:
  name: ml-training-nodeclass
spec:
  amiFamily: AL2023
  # Enable EFA networking
  blockDeviceMappings:
    - deviceName: /dev/xvda
      ebs:
        volumeSize: 500Gi
        volumeType: gp3
        iops: 16000
        throughput: 1000
    # Additional volume for training data
    - deviceName: /dev/xvdb
      ebs:
        volumeSize: 2000Gi
        volumeType: gp3
---
apiVersion: kubeflow.org/v1
kind: PyTorchJob
metadata:
  name: distributed-training
spec:
  pytorchReplicaSpecs:
    Master:
      replicas: 1
      template:
        spec:
          tolerations:
            - key: ml-training
              value: "true"
              effect: NoSchedule
          nodeSelector:
            tier: ml-training
          containers:
            - name: pytorch
              image: my-training-image:latest
              resources:
                limits:
                  nvidia.com/gpu: 8
    Worker:
      replicas: 3
      template:
        spec:
          tolerations:
            - key: ml-training
              value: "true"
              effect: NoSchedule
          nodeSelector:
            tier: ml-training
          containers:
            - name: pytorch
              image: my-training-image:latest
              resources:
                limits:
                  nvidia.com/gpu: 8

ML 训练优化摘要 ​

方面建议理由
实例类型p5.48xlarge, p4d.24xlarge最大 GPU 容量
存储500GB+ root, 2TB+ data大型数据集、检查点
IOPS16000+快速写入检查点
网络已启用 EFA分布式训练
框架PyTorchJob, TFJob原生分布式支持

工作负载类型快速参考 ​

工作负载NodePool 策略实例类型容量整合
Web 服务可用性优先m-seriesOn-Demand中等（5m）
API 后端混合m/c-series混合中等（5m）
批处理成本优先c-series仅 Spot激进（30s）
CI/CD成本优先c/m-series优先 Spot激进（1m）
数据库稳定性优先r-seriesOn-Demand保守（10m）
GPU 推理可用性优先g5-seriesOn-Demand宽松（10m）
ML 训练性能优先p5/p4dOn-Demand宽松（15m）
流处理均衡m/c-series混合中等（5m）

Pod 资源指南 ​

CPU 密集型工作负载 ​

yaml
resources:
  requests:
    cpu: 2000m      # Request what you need
    memory: 2Gi
  limits:
    cpu: 4000m      # Allow some burst
    memory: 4Gi

内存密集型工作负载 ​

yaml
resources:
  requests:
    cpu: 500m
    memory: 8Gi     # Request what you need
  limits:
    cpu: 1000m
    memory: 8Gi     # Limit = request (no overcommit)

GPU 工作负载 ​

yaml
resources:
  requests:
    cpu: 4000m
    memory: 16Gi
  limits:
    nvidia.com/gpu: 1   # GPU limits are always required
    cpu: 8000m
    memory: 32Gi

< 上一页：Node 生命周期 | 目录 | 下一页：迁移指南 >