Multi-Node Diagnostics#
The multi-node diagnostics module coordinates a named diagnostic across
multiple DCGM host engines. These diagnostics are selected with
dcgmi mndiag --run. They are distinct from the single-node diagnostic
plugins selected with dcgmi diag --run and should not be described as a
plugin interface.
DCGM 4.6 accepts one multi-node diagnostic per request and provides one
diagnostic, mnubergemm. Keeping the diagnostic catalogue separate from
the command reference distinguishes the multi-node framework from the
diagnostic that happens to be available in the current release.
Available diagnostics#
Canonical name |
Purpose |
Availability in DCGM 4.6 |
|---|---|---|
Run a coordinated MPI workload across the selected hosts and GPUs and report GPU-scoped errors emitted while the workload exercises the supported multi-node system. |
GB200 NVL and GB300 NVL systems |
The multi-node diagnostic module was introduced in DCGM 4.3. GB300 NVL support was added in DCGM 4.4. The diagnostic page gives the complete version, product, parameter, result, and failure-condition reference.
Selection and parameters#
The --run value is a diagnostic name from the table above. The
dcgmi 4.6 client compares the name without regard to case and defaults to
mnubergemm when --run is omitted. It does not provide a command that
enumerates multi-node diagnostics at runtime; this catalogue supplies the
canonical names for the documented release.
Parameters use the form diagnostic.parameter=value. Syntactic acceptance
by the common dcgmi parameter parser does not establish that a parameter
is supported: the selected diagnostic can validate, ignore, or pass through
an assignment. Use the diagnostic page for supported parameter names,
values, and defaults.
Results and command status#
A multi-node request has two different kinds of outcome:
An orchestration error means DCGM could not connect to the participating host engines, reserve resources, start or monitor the workload, or collect a usable response. The command reports a DCGM error and returns a nonzero status.
A diagnostic result describes what the workload reported after it ran. A completed request can contain a failed diagnostic result while
dcgmiitself returns status zero.
The current response contains an overall diagnostic result together with per-host and per-GPU information when the workload supplies it. See dcgmi mndiag for invocation, host-list, output, signal, and exit-status behavior. See Multi-Node Diagnostics for the environment setup and an end-to-end qualification workflow.