NAME: DCGM Exporter - Generates GPU metrics in the prometheus format USAGE: DCGM Exporter [global options] command [command options] VERSION: 4.8.4 COMMANDS: help, h Shows a list of commands or help for one command GLOBAL OPTIONS: --collectors value, -f value Path to the file, that contains the DCGM fields to collect (default: "/etc/dcgm-exporter/default-counters.csv") [$DCGM_EXPORTER_COLLECTORS] --config-file value Path to a YAML configuration file. Legacy flags and environment variables explicitly set on startup override YAML values. [$DCGM_EXPORTER_CONFIG_FILE] --address value, -a value Listen address as :. For IPv6, use "[]:" (e.g., "[::]:9400") (default: ":9400") [$DCGM_EXPORTER_LISTEN] --web-read-timeout value Maximum duration for reading an HTTP request. (default: "10s") [$DCGM_EXPORTER_WEB_READ_TIMEOUT] --web-write-timeout value Maximum duration for generating and writing an HTTP response. (default: "30s") [$DCGM_EXPORTER_WEB_WRITE_TIMEOUT] --max-concurrent-scrapes value Maximum concurrent scrape requests. Excess requests receive HTTP 503 while admitted overlapping requests are coalesced. (default: 16) [$DCGM_EXPORTER_MAX_CONCURRENT_SCRAPES] --enable-exporter-metrics Expose Go runtime, process, and HTTP handler metrics about dcgm-exporter itself. (default: false) [$DCGM_EXPORTER_ENABLE_EXPORTER_METRICS] --collect-interval value, -c value Interval of time at which point metrics are collected. Unit is milliseconds (ms). (default: 30000) [$DCGM_EXPORTER_INTERVAL] --watch-max-keep-age value Maximum age of samples retained by DCGM field watches. Set to 0s to disable the age limit when a sample limit is configured. (default: 10m0s) [$DCGM_EXPORTER_WATCH_MAX_KEEP_AGE] --watch-max-keep-samples value Sample-count retention requested for DCGM field watches. Set to 0 for no sample-count limit when an age limit is configured. (default: 0) [$DCGM_EXPORTER_WATCH_MAX_KEEP_SAMPLES] --kubernetes, -k Enable kubernetes mapping metrics to kubernetes pods (default: false) [$DCGM_EXPORTER_KUBERNETES] --use-old-namespace, -o Use old 1.x namespace (default: false) [$DCGM_EXPORTER_USE_OLD_NAMESPACE] --cpu-devices value, -p value Specify which devices dcgm-exporter monitors. Possible values: f or g[:id1[,-id2...]] or i[:id1[,-id2...]] or g[:id1[,-id2...]]+i[:id1[,-id2...]]. If an id list is used, then devices with match IDs must exist on the system. For example: (default) = monitor all GPU instances in MIG mode, all GPUs if MIG mode is disabled. (See f) g = Monitor all GPUs i = Monitor all GPU instances f = Monitor all GPUs if MIG is disabled, or all GPU instances if MIG is enabled. Note: this rule will be applied to each GPU. If it has GPU instances, those will be monitored. If it doesn't, then the GPU will be monitored. This is our recommended option for single or mixed MIG Strategies. g:0,1 = monitor GPUs 0 and 1 i:0,2-4 = monitor GPU instances 0, 2, 3, and 4. g+i = monitor all GPUs and GPU instances. NOTE 1: Any time indices are specified, those indices must exist on the system. NOTE 2: The flex option f cannot be combined with g or i. (default: "f") [$DCGM_EXPORTER_CPU_DEVICES_STR] --configmap-data value, -m value ConfigMap : for metric data (default: "none") [$DCGM_EXPORTER_CONFIGMAP_DATA] --remote-hostengine-info value, -r value Connect to remote hostengine at : or a DCGM URI (tcp://:, unix:///, vsock://:). For IPv6, use "[]:" (e.g., "[::1]:5555") (default: "localhost:5555") [$DCGM_REMOTE_HOSTENGINE_INFO] --kubernetes-enable-pod-labels Enable kubernetes pod labels in metrics. This parameter is effective only when the '--kubernetes' option is set to 'true'. (default: false) [$DCGM_EXPORTER_KUBERNETES_ENABLE_POD_LABELS] --kubernetes-enable-pod-uid Enable kubernetes pod UID in metrics. This parameter is effective only when the '--kubernetes' option is set to 'true'. (default: false) [$DCGM_EXPORTER_KUBERNETES_ENABLE_POD_UID] --kubernetes-gpu-id-type value Choose Type of GPU ID to use to map kubernetes resources to pods. Possible values: 'uid', 'device-name' (default: "uid") [$DCGM_EXPORTER_KUBERNETES_GPU_ID_TYPE] --kubernetes-pod-label-allowlist-regex value [ --kubernetes-pod-label-allowlist-regex value ] Regex patterns for filtering pod labels to include in metrics (comma-separated). Empty means include all labels. This parameter is effective only when '--kubernetes-enable-pod-labels' is true. [$DCGM_EXPORTER_KUBERNETES_POD_LABEL_ALLOWLIST_REGEX] --devices value, -d value Specify which devices dcgm-exporter monitors. Possible values: f or g[:id1[,-id2...]] or i[:id1[,-id2...]] or g[:id1[,-id2...]]+i[:id1[,-id2...]]. If an id list is used, then devices with match IDs must exist on the system. For example: (default) = monitor all GPU instances in MIG mode, all GPUs if MIG mode is disabled. (See f) g = Monitor all GPUs i = Monitor all GPU instances f = Monitor all GPUs if MIG is disabled, or all GPU instances if MIG is enabled. Note: this rule will be applied to each GPU. If it has GPU instances, those will be monitored. If it doesn't, then the GPU will be monitored. This is our recommended option for single or mixed MIG Strategies. g:0,1 = monitor GPUs 0 and 1 i:0,2-4 = monitor GPU instances 0, 2, 3, and 4. g+i = monitor all GPUs and GPU instances. NOTE 1: Any time indices are specified, those indices must exist on the system. NOTE 2: The flex option f cannot be combined with g or i. NOTE 3: For GPU devices, i cannot be specified unless MIG mode is enabled. (default: "f") [$DCGM_EXPORTER_DEVICES_STR] --no-hostname, -n Omit the hostname information from the output, matching older versions. (default: false) [$DCGM_EXPORTER_NO_HOSTNAME] --switch-devices value, -s value Specify which devices dcgm-exporter monitors. Possible values: f or g[:id1[,-id2...]] or i[:id1[,-id2...]] or g[:id1[,-id2...]]+i[:id1[,-id2...]]. If an id list is used, then devices with match IDs must exist on the system. For example: (default) = monitor all GPU instances in MIG mode, all GPUs if MIG mode is disabled. (See f) g = Monitor all GPUs i = Monitor all GPU instances f = Monitor all GPUs if MIG is disabled, or all GPU instances if MIG is enabled. Note: this rule will be applied to each GPU. If it has GPU instances, those will be monitored. If it doesn't, then the GPU will be monitored. This is our recommended option for single or mixed MIG Strategies. g:0,1 = monitor GPUs 0 and 1 i:0,2-4 = monitor GPU instances 0, 2, 3, and 4. g+i = monitor all GPUs and GPU instances. NOTE 1: Any time indices are specified, those indices must exist on the system. NOTE 2: The flex option f cannot be combined with g or i. (default: "f") [$DCGM_EXPORTER_OTHER_DEVICES_STR] --fake-gpus Accept GPUs that are fake, for testing purposes only (default: false) [$DCGM_EXPORTER_USE_FAKE_GPUS] --web-config-file value Web configuration file following webConfig spec: https://github.com/prometheus/exporter-toolkit/blob/master/docs/web-configuration.md. [$DCGM_EXPORTER_WEB_CONFIG_FILE] --xid-count-window-size value, -x value Window size in milliseconds (ms) for counting XID errors in DCGM_EXP_XID_ERRORS_COUNT. (default: 300000) [$DCGM_EXPORTER_XID_COUNT_WINDOW_SIZE] --replace-blanks-in-model-name, --rbmn Replace every blank space in the GPU model name with a dash, ensuring a continuous, space-free identifier. (default: false) [$DCGM_EXPORTER_REPLACE_BLANKS_IN_MODEL_NAME] --debug Enable debug output (default: false) [$DCGM_EXPORTER_DEBUG] --clock-events-count-window-size value Window size in milliseconds (ms) for counting clock events in DCGM_EXP_CLOCK_EVENTS_COUNT. (default: 300000) [$DCGM_EXPORTER_CLOCK_EVENTS_COUNT_WINDOW_SIZE] --enable-dcgm-log Enable writing DCGM logs to standard output (stdout). (default: false) [$DCGM_EXPORTER_ENABLE_DCGM_LOG] --dcgm-log-level value Specify the DCGM log verbosity level. This parameter is effective only when the '--enable-dcgm-log' option is set to 'true'. Possible values: NONE, FATAL, ERROR, WARN, INFO, DEBUG and VERB (default: "NONE") [$DCGM_EXPORTER_DCGM_LOG_LEVEL] --log-format value Specify the log output format. Possible values: text, json (default: "text") [$DCGM_EXPORTER_LOG_FORMAT] --pod-resources-kubelet-socket value Path to the kubelet pod-resources socket file. (default: "/var/lib/kubelet/pod-resources/kubelet.sock") [$DCGM_POD_RESOURCES_KUBELET_SOCKET] --hpc-job-mapping-dir value Path to HPC job mapping file directory used for mapping GPUs to jobs. [$DCGM_HPC_JOB_MAPPING_DIR] --container-labels Enable runtime container labels in metrics. (default: false) [$DCGM_EXPORTER_CONTAINER_LABELS] --container-runtime-socket value Path to a container runtime socket used for container labels. [$DCGM_CONTAINER_RUNTIME_SOCKET] --nvidia-resource-names value [ --nvidia-resource-names value ] Nvidia resource names for specified GPU type like nvidia.com/a100, nvidia.com/a10. [$NVIDIA_RESOURCE_NAMES] --kubernetes-virtual-gpus Capture metrics associated with virtual GPUs exposed by Kubernetes device plugins when using GPU sharing strategies, e.g. time-sharing or MPS. (default: false) [$KUBERNETES_VIRTUAL_GPUS] --dump-enabled Enable file-based debugging dumps for troubleshooting (default: false) [$DCGM_EXPORTER_DUMP_ENABLED] --dump-directory value Directory to store debug dump files (default: "/tmp/dcgm-exporter-debug") [$DCGM_EXPORTER_DUMP_DIRECTORY] --dump-retention value Retention period for debug dump files in hours (0 = no cleanup) (default: 24) [$DCGM_EXPORTER_DUMP_RETENTION] --dump-compression Use gzip compression for debug dump files (default: true) [$DCGM_EXPORTER_DUMP_COMPRESSION] --kubernetes-enable-dra Capture metrics associated with GPUs managed by Kubernetes Dynamic Resource Allocation (DRA) API. (default: false) [$KUBERNETES_ENABLE_DRA] --disable-startup-validate Disable validation checks during startup. Can be useful for running in minimal environments or testing (default: false) [$DISABLE_STARTUP_VALIDATE] --enable-gpu-bind-unbind-watch Enable watching for GPU bind/unbind events to trigger automatic reloads (requires DCGM 4.5+) (default: false) [$DCGM_EXPORTER_ENABLE_GPU_BIND_UNBIND_WATCH] --gpu-bind-unbind-poll-interval value Interval for polling GPU bind/unbind events (DCGM recommends 1s) (default: "1s") [$DCGM_EXPORTER_GPU_BIND_UNBIND_POLL_INTERVAL] --enable-pprof Enable /debug/pprof/ HTTP endpoints for profiling and debugging (default: false) [$DCGM_EXPORTER_ENABLE_PPROF] --web-systemd-socket Use systemd socket activation listeners instead of port listeners (Linux only). (default: false) [$DCGM_EXPORTER_SYSTEMD_SOCKET] --help, -h show help --version, -v print the version