dcgmi dmon#

NAME#

dcgmi-dmon - stream DCGM field values for selected entities

SYNOPSIS#

dcgmi dmon --list [--host <address>]
dcgmi dmon (--field-id <field-id-list> |
            --field-group-id <field-group-id>)
           [--entity-id <entities> | --group-id <group>]
           [--delay <milliseconds>] [--count <count>]
           [--host <address>]

DESCRIPTION#

dcgmi dmon periodically reads watched DCGM fields and prints one row per entity. Select fields directly or through a persistent field group. Select entities directly or through a persistent entity group.

When no entity or group is supplied, the command monitors all supported GPUs. It runs until interrupted unless --count sets a finite number of samples.

OPTIONS#

Field selection#

-l, --list: List the long name, short column name, and field ID for fields known to the local dcgmi. Do not combine this operation with monitoring options.
-e <field-id-list>, --field-id <field-id-list>: Monitor a comma-separated list of numeric DCGM field IDs.
-f <field-group-id>, --field-group-id <field-group-id>: Monitor every field in a persistent field group. The ID must be nonnegative; obtain current IDs with dcgmi fieldgroup --list.

Entity selection#

-i <entities>, --entity-id <entities>: Monitor a comma-separated entity list. See dcgmi for entity, range, MIG, and link selector syntax.
-g <group>, --group-id <group>: Monitor a persistent numeric group. The special values all_gpus and all_nvswitches select the corresponding built-in groups. Do not combine this option with --entity-id.

Sampling#

-d <milliseconds>, --delay <milliseconds>: Wait this many milliseconds between output samples. The default is 1,000; the minimum is 1.
-c <count>, --count <count>: Exit after this many samples. Zero, the default, runs until interrupted.

Common options#

--host <address>: Connect to the specified host engine.
-h, --help: Print generated usage information and exit.

FIELD CATALOGUE AND OUTPUT#

dcgmi dmon --list is the command-line field catalogue for the local DCGM release. It enumerates local field metadata; it does not probe the selected GPU, driver, or entity for support. Use the numeric IDs from this output with --field-id or dcgmi fieldgroup. The full static definitions are in Field Identifiers; consult Field Scope before pairing a field with a non-GPU entity and Field Types when interpreting a raw value.

Each sample contains one row per selected entity and one column per requested field. Floating-point values are printed with three decimal places; integers and timestamps are printed as decimal integers; strings are printed directly. Binary fields and all DCGM blank sentinel values are printed as N/A. Thus N/A can mean not supported, not found, not permissioned, or not yet sampled; dmon does not preserve that distinction in its table.

A valid field ID is not necessarily valid for every entity type or GPU generation. A failed watch request is a command failure. An accepted watch can still produce N/A when the field is unavailable on a particular row.

WATCH LIFETIME#

Monitoring creates connection-scoped watches at the requested delay and asks the host engine to retain at most two samples for this client. Those requests end when dmon exits. Other clients can independently keep the same field watched, so stopping dmon does not imply that system-wide collection of the field has stopped.

EXAMPLES#

List monitorable fields:

$ dcgmi dmon --list

Read fields 100 and 101 for every GPU ten times:

$ dcgmi dmon --field-id 100,101 --count 10

Read field 1200 for links 0 through 5 on GPU 0:

$ dcgmi dmon --field-id 1200 --entity-id 'gpu_link:0:{0-5}'

Monitor field group 5 for persistent entity group 2 every five seconds:

$ dcgmi dmon --field-group-id 5 --group-id 2 --delay 5000