gpu-resource-management

name	gpu-resource-management
description	Design GPU orchestration on EKS using Karpenter v1.2+ NodePools, KEDA scale-to-zero, and DRA 1.35 GA for multi-instance GPU (MIG) partitioning. Right-size NodePool for p5/g6e/trn2 instance mix, spot/on-demand split, consolidation, and topology-aware scheduling.
argument-hint	[workload profile — training/inference, GPU mix, QPS pattern]
user-invocable	true
model	claude-sonnet-4-6
allowed-tools	Read,Write,Edit,Bash,Grep,Glob,mcp__eks,mcp__aws-pricing,mcp__aws-documentation,mcp__prometheus

When to Use

• GPU NodePool 설계 및 Karpenter v1.2+ consolidation 정책 조정이 필요할 때
• KEDA 로 scale-to-zero, HPA 로 Pod 수준 스케일링 정책을 수립할 때
• DRA 1.35 GA 로 MIG(Multi-Instance GPU) 파티셔닝을 적용할 때
• Spot vs On-Demand 비율, Reserved Instance vs Savings Plans 구성 검토

When NOT to Use

• CPU-only 워크로드 — 본 스킬은 GPU 전용
• Managed 환경(SageMaker endpoints) 사용 중 — 별도 스킬
• GPU Operator 미설치 — agentic-eks-bootstrap 선행

Preconditions

• EKS 1.32+, Kubernetes DRA 1.35 GA 활성(--feature-gates=DynamicResourceAllocation=true)
• Karpenter v1.2+, NVIDIA GPU Operator, DCGM Exporter, Prometheus 설치
• 대상 리전에서 p5/g6e/trn2 쿼터 확인 완료

Procedure

Step 1. NodePool 설계

apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: gpu-inference-pool
spec:
  template:
    metadata:
      labels:
        node-type: gpu-inference
        workload: genai
    spec:
      requirements:
        - key: kubernetes.io/arch
          operator: In
          values: ["amd64"]
        - key: karpenter.sh/capacity-type
          operator: In
          values: ["on-demand", "spot"]
        - key: node.kubernetes.io/instance-type
          operator: In
          values: ["p5.48xlarge", "g6e.12xlarge", "g6e.48xlarge"]
      taints:
        - key: nvidia.com/gpu
          value: "true"
          effect: NoSchedule
      nodeClassRef:
        name: default
  disruption:
    consolidationPolicy: WhenEmptyOrUnderutilized
    consolidateAfter: 30s
  limits:
    nvidia.com/gpu: 64

Step 2. Spot / On-Demand 비율 전략

• 추론 (inference): Spot 70% + On-Demand 30%, HPA min=1 유지
• 학습 (training): Savings Plans 또는 Capacity Blocks for ML
• Reserved Instance: 24/7 baseline 워크로드에만 적용

Step 3. KEDA ScaledObject (Scale-to-Zero)

apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: vllm-scaler
  namespace: inference
spec:
  scaleTargetRef:
    name: vllm-llama3
  minReplicaCount: 0
  maxReplicaCount: 8
  triggers:
    - type: prometheus
      metadata:
        serverAddress: http://prometheus.observability.svc:9090
        metricName: vllm_num_requests_running
        threshold: "4"
        query: sum(rate(vllm_num_requests_running[1m]))

Step 4. DRA 1.35 GA - MIG 파티셔닝

• H100 80GB 를 1g.10gb × 7 로 분할하면 소형 모델 7개 동시 서빙
• ResourceClaim + DeviceClass 리소스로 Pod 에 동적 할당

apiVersion: resource.k8s.io/v1beta1
kind: ResourceClaimTemplate
metadata:
  name: mig-1g10gb
spec:
  spec:
    devices:
      requests:
        - name: gpu
          deviceClassName: nvidia.com/mig-1g.10gb

Step 5. Topology-Aware Routing

• NCCL 통신 최소화: 같은 AZ, 같은 placement group 내 TP 그룹 배치
• topologySpreadConstraints 로 AZ 분산 DP, 동일 노드 내 TP

Step 6. 비용 관측 + 알림

• Prometheus: DCGM_FI_DEV_GPU_UTIL, DCGM_FI_DEV_MEM_COPY_UTIL
• KubeCost + AWS Cost Explorer 태그
• 알림: GPU util < 20% 30분 지속 시 Slack 알림, consolidation 검토

Good Examples

• 추론 NodePool: p5 + g6e 혼합, Spot 70%, consolidation 30s, KEDA scale-to-zero
• MIG: H100 80GB → 1g.10gb × 7 → 7B-14B 모델 7개 동시
• DRA 1.35 로 ResourceClaim 기반 fractional GPU 할당

Bad Examples (금지)

• Karpenter v0.x legacy API (provisioner.karpenter.sh) — v1 migration 필수
• 단일 인스턴스 타입 제약 — 쿼터 부족 시 Pending 무한 대기
• GPU idle 상태 방치 — Spot 인스턴스로도 비용 누수
• MIG 없이 7B 모델 7개를 동일 GPU 에 공유 (device plugin time-slicing 의존) — latency 변동

References

• GPU Resource Management (community resource)
• NVIDIA GPU Stack (community resource)
• AWS Neuron Stack (community resource)
• Karpenter 공식 문서
• KEDA 공식 문서
• Kubernetes DRA 1.35 GA
• NVIDIA MIG 가이드
• 플러그인 내부: ../../references/vllm-performance-tuning.md