本頁面由 Cloud Translation API 翻譯而成。

NVIDIA Data Center GPU Manager (DCGM)

本文說明如何設定 Google Kubernetes Engine 部署作業，以便使用 Google Cloud Managed Service for Prometheus 從 NVIDIA Data Center GPU Manager 收集指標。本文說明如何執行下列操作：

設定 DCGM 的匯出工具，以回報指標。

只有在使用 Managed Service for Prometheus 的代管收集作業時，才需要按照這些操作說明進行設定。如果您使用自行部署的收集作業，請參閱 DCGM Exporter 的來源存放區，瞭解安裝資訊。

這些操作說明僅供參考，適用於大多數 Kubernetes 環境。如要瞭解代管 DCGM 服務，請參閱「收集及查看 DCGM 指標」。

如果因嚴格的安全或機構政策而無法安裝應用程式或匯出工具，建議您參閱開放原始碼文件尋求支援。

如需 NVIDIA Data Center GPU Manager 的相關資訊，請參閱 NVIDIA DCGM。

必要條件

如要使用 Managed Service for Prometheus 和代管收集作業從 DCGM 收集指標，部署作業必須符合下列規定：

叢集必須執行 Google Kubernetes Engine 1.28.15-gke.2475000 以上版本。
您必須執行 Managed Service for Prometheus，並啟用代管收集作業。詳情請參閱「開始使用代管集合」一文。

確認您有足夠的 NVIDIA GPU 配額。

如要列舉 GKE 叢集中的 GPU 節點及其 GPU 類型，請在相關叢集中執行下列指令：

kubectl get nodes -l cloud.google.com/gke-gpu -o jsonpath='{range .items[*]}{@.metadata.name}{" "}{@.metadata.labels.cloud\.google\.com/gke-accelerator}{"\n"}{end}'

請注意，如果自動安裝功能已停用，或您的 GKE 版本不支援這項功能，您可能必須在節點上安裝相容的 NVIDIA GPU 驅動程式。如要確認 NVIDIA GPU 裝置外掛程式是否正在執行，請執行下列指令：
```
kubectl get pods -n kube-system | grep nvidia-gpu-device-plugin
```

安裝 DCGM 匯出工具

建議您使用下列設定安裝 DCGM 匯出工具 DCGM-Exporter：

# Copyright 2023 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: nvidia-dcgm
  namespace: gmp-public
  labels:
    app: nvidia-dcgm
spec:
  selector:
    matchLabels:
      app: nvidia-dcgm
  updateStrategy:
    type: RollingUpdate
  template:
    metadata:
      labels:
        name: nvidia-dcgm
        app: nvidia-dcgm
    spec:
      affinity:
        nodeAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
            nodeSelectorTerms:
            - matchExpressions:
              - key: cloud.google.com/gke-accelerator
                operator: Exists
      tolerations:
      - operator: "Exists"
      volumes:
      - name: nvidia-install-dir-host
        hostPath:
          path: /home/kubernetes/bin/nvidia
          type: Directory
      containers:
      - image: "nvcr.io/nvidia/cloud-native/dcgm:3.3.0-1-ubuntu22.04"
        command: ["nv-hostengine", "-n", "-b", "ALL"]
        ports:
        - containerPort: 5555
          hostPort: 5555
        name: nvidia-dcgm
        securityContext:
          privileged: true
        volumeMounts:
        - name: nvidia-install-dir-host
          mountPath: /usr/local/nvidia
---
apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: nvidia-dcgm-exporter
  namespace: gmp-public
  labels:
    app.kubernetes.io/name: nvidia-dcgm-exporter
spec:
  selector:
    matchLabels:
      app.kubernetes.io/name: nvidia-dcgm-exporter
  updateStrategy:
    type: RollingUpdate
  template:
    metadata:
      labels:
        app.kubernetes.io/name: nvidia-dcgm-exporter
    spec:
      affinity:
        nodeAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
            nodeSelectorTerms:
            - matchExpressions:
              - key: cloud.google.com/gke-accelerator
                operator: Exists
      tolerations:
      - operator: "Exists"
      volumes:
      - name: nvidia-dcgm-exporter-metrics
        configMap:
          name: nvidia-dcgm-exporter-metrics
      - name: nvidia-install-dir-host
        hostPath:
          path: /home/kubernetes/bin/nvidia
          type: Directory
      - name: pod-resources
        hostPath:
          path: /var/lib/kubelet/pod-resources
      containers:
      - name: nvidia-dcgm-exporter
        image: nvcr.io/nvidia/k8s/dcgm-exporter:3.3.0-3.2.0-ubuntu22.04
        command: ["/bin/bash", "-c"]
        args:
        - hostname $NODE_NAME; dcgm-exporter --remote-hostengine-info $(NODE_IP) --collectors /etc/dcgm-exporter/counters.csv
        ports:
        - name: metrics
          containerPort: 9400
        securityContext:
          privileged: true
        env:
        - name: NODE_NAME
          valueFrom:
            fieldRef:
              fieldPath: spec.nodeName
        - name: "DCGM_EXPORTER_KUBERNETES_GPU_ID_TYPE"
          value: "device-name"
        - name: LD_LIBRARY_PATH
          value: /usr/local/nvidia/lib64
        - name: NODE_IP
          valueFrom:
            fieldRef:
              fieldPath: status.hostIP
        - name: DCGM_EXPORTER_KUBERNETES
          value: 'true'
        - name: DCGM_EXPORTER_LISTEN
          value: ':9400'
        volumeMounts:
        - name: nvidia-dcgm-exporter-metrics
          mountPath: "/etc/dcgm-exporter"
          readOnly: true
        - name: nvidia-install-dir-host
          mountPath: /usr/local/nvidia
        - name: pod-resources
          mountPath: /var/lib/kubelet/pod-resources
---
apiVersion: v1
kind: ConfigMap
metadata:
  name: nvidia-dcgm-exporter-metrics
  namespace: gmp-public
data:
  counters.csv: |
    # Utilization (the sample period varies depending on the product),,
    DCGM_FI_DEV_GPU_UTIL, gauge, GPU utilization (in %).
    DCGM_FI_DEV_MEM_COPY_UTIL, gauge, Memory utilization (in %).

    # Temperature and power usage,,
    DCGM_FI_DEV_GPU_TEMP, gauge, Current temperature readings for the device in degrees C.
    DCGM_FI_DEV_MEMORY_TEMP, gauge, Memory temperature for the device.
    DCGM_FI_DEV_POWER_USAGE, gauge, Power usage for the device in Watts.

    # Utilization of IP blocks,,
    DCGM_FI_PROF_SM_ACTIVE, gauge, The ratio of cycles an SM has at least 1 warp assigned
    DCGM_FI_PROF_SM_OCCUPANCY, gauge, The fraction of resident warps on a multiprocessor
    DCGM_FI_PROF_PIPE_TENSOR_ACTIVE, gauge, The ratio of cycles the tensor (HMMA) pipe is active (off the peak sustained elapsed cycles)
    DCGM_FI_PROF_PIPE_FP64_ACTIVE, gauge, The fraction of cycles the FP64 (double precision) pipe was active.
    DCGM_FI_PROF_PIPE_FP32_ACTIVE, gauge, The fraction of cycles the FP32 (single precision) pipe was active.
    DCGM_FI_PROF_PIPE_FP16_ACTIVE, gauge, The fraction of cycles the FP16 (half precision) pipe was active.

    # Memory usage,,
    DCGM_FI_DEV_FB_FREE, gauge, Framebuffer memory free (in MiB).
    DCGM_FI_DEV_FB_USED, gauge, Framebuffer memory used (in MiB).
    DCGM_FI_DEV_FB_TOTAL, gauge, Total Frame Buffer of the GPU in MB.

    # PCIE,,
    DCGM_FI_PROF_PCIE_TX_BYTES, gauge, Total number of bytes transmitted through PCIe TX
    DCGM_FI_PROF_PCIE_RX_BYTES, gauge, Total number of bytes received through PCIe RX

    # NVLink,,
    DCGM_FI_PROF_NVLINK_TX_BYTES, gauge, The number of bytes of active NvLink tx (transmit) data including both header and payload.
    DCGM_FI_PROF_NVLINK_RX_BYTES, gauge, The number of bytes of active NvLink rx (read) data including both header and payload.

如要確認 DCGM Exporter 是否在預期端點上發出指標，請按照下列步驟操作：

使用下列指令設定通訊埠轉送：

kubectl -n gmp-public port-forward POD_NAME 9400

在另一個終端機工作階段中，使用瀏覽器或 curl 公用程式存取端點 localhost:9400/metrics。

您可以自訂 ConfigMap 區段，選取要發布的 GPU 指標。

或者，您也可以考慮使用官方 Helm 資訊套件安裝 DCGM Exporter。

如要從本機檔案套用設定變更，請執行下列指令：

kubectl apply -n NAMESPACE_NAME -f FILE_NAME

您也可以使用 Terraform 管理設定。

定義 PodMonitoring 資源

如要探索目標，Managed Service for Prometheus Operator 需要與 DCGM Exporter 對應的 PodMonitoring 資源 (位於相同命名空間)。

您可以使用下列 PodMonitoring 設定：

# Copyright 2023 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

apiVersion: monitoring.googleapis.com/v1
kind: ClusterPodMonitoring
metadata:
  name: nvidia-dcgm-exporter
  labels:
    app.kubernetes.io/name: nvidia-dcgm-exporter
    app.kubernetes.io/part-of: google-cloud-managed-prometheus
spec:
  selector:
    matchLabels:
      app.kubernetes.io/name: nvidia-dcgm-exporter
  endpoints:
  - port: metrics
    interval: 30s
  targetLabels:
    metadata: []

如要從本機檔案套用設定變更，請執行下列指令：

kubectl apply -n NAMESPACE_NAME -f FILE_NAME

您也可以使用 Terraform 管理設定。

驗證設定

您可以使用 Metrics Explorer 驗證 DCGM Exporter 是否設定正確。Cloud Monitoring 可能需要一到兩分鐘才能擷取指標。

如要確認指標已擷取，請按照下列步驟操作：

前往 Google Cloud 控制台的「Metrics Explorer」頁面：
前往 Metrics Explorer

如果您是使用搜尋列尋找這個頁面，請選取子標題為「Monitoring」的結果。
在查詢建構工具窗格的工具列中，選取名稱為 MQL 或 PromQL 的按鈕。
確認已在「Language」(語言) 切換按鈕中選取「PromQL」。語言切換按鈕位於同一工具列，可供你設定查詢格式。

輸入並執行下列查詢：

DCGM_FI_DEV_GPU_UTIL{cluster="CLUSTER_NAME", namespace="gmp-public"}

疑難排解

如要瞭解如何排解指標擷取問題，請參閱「排解擷取端問題」中的「無法從匯出工具收集資料」。