Prometheus

IMPORTANT

From version v0.11.0, this configuration is not anymore necessary and if recording rules are generated ceems_tool, the relabeling will be handled directly in the rules.

In order to use the dashboards provided in the repository, minor metric_relabel_configs configuration must be provided for all target groups that have NVIDIA GPUs where the dcgm-exporter exports GPU metrics to Prometheus.

TIP

ceems_tool can be used to generate the relabel configs based on your current Prometheus targets.

The following example shows scrape configurations where the target nodes contain NVIDIA GPUs:

scrape_configs:
  # Scrape job containing NVIDIA DCGM exporter targets
  - job_name: <job-name>
    metric_relabel_configs:
      - source_labels:
          - modelName
          - UUID
        target_label: gpuuuid
        regex: NVIDIA(.*);(.*)
        replacement: $2
        action: replace
      - source_labels:
          - modelName
          - GPU_I_ID
        target_label: gpuiid
        regex: NVIDIA(.*);(.*)
        replacement: $2
        action: replace
      - regex: UUID
        action: labeldrop
      - regex: GPU_I_ID
        action: labeldrop

  # Scrape job containing AMD SMI exporter targets
  # https://github.com/amd/amd_smi_exporter
  - job_name: <job-name>
    metric_relabel_configs:
      - source_labels:
          - gpu_power
        target_label: index
        regex: (.*)
        replacement: $1
        action: replace
      - source_labels:
          - index
          - gpu_use_percent
        target_label: index
        regex: ;(.+)
        replacement: $1
        action: replace
      - source_labels:
          - index
          - gpu_memory_use_percent
        target_label: index
        regex: ;(.+)
        replacement: $1
        action: replace
      - regex: gpu_power
        action: labeldrop
      - regex: gpu_use_percent
        action: labeldrop
      - regex: gpu_memory_use_percent
        action: labeldrop

  # Scrape job containing AMD device metrics exporter targets
  # https://github.com/ROCm/device-metrics-exporter/
  - job_name: <job-name>
    metric_relabel_configs:
      - source_labels:
          - gpu_id
        target_label: index
        regex: (.*)
        replacement: $1
        action: replace
      - source_labels:
          - gpu_partition_id
        target_label: gpuiid
        regex: (.*)
        replacement: $1
        action: replace
      - regex: gpu_id
        action: labeldrop
      - regex: gpu_partition_id
        action: labeldrop

The metric_relabel_configs section renames the UUID and GPU_I_ID labels (which represent the UUID and MIG instance ID of the NVIDIA GPU, respectively) to gpuuuid and gpuiid, making them compatible with the CEEMS exporter. Moreover, the configuration also drops the unused UUID and GPU_I_ID labels to reduce storage usage.

Similarly, for AMD SMI exporter targets, the metric_relabel_configs section extracts the GPU index from the gpu_power, gpu_use_percent, and gpu_memory_use_percent labels and maps it to the index label, which is compatible with the CEEMS exporter. In the case of AMD Device Metrics Exporter, the GPU index is exported as gpu_id and GPU partition index is exported by gpu_partition_id which are replaced by index and gpuiid.