NVIDIA Minerva gathers metric, log, and topology telemetry from GPUs, NICs, DPUs, scale-up NVLink switches, scale-out Ethernet switches, north-south Ethernet switches, NCCL, schedulers, and AI workloads.
Configuration requirements for most data sources are in the Reference Configuration Package under the endpoint-integrations directory, where each data source has a README.md file with support information, setup instructions, and troubleshooting tips. The NCCL dashboards and their prerequisites are in minerva-configs/dashboards/user-tenant/nccl. To obtain the Reference Configuration Package, contact your NVIDIA sales or support representative.
|
Data Source |
Hardware or Component |
What NVIDIA Minerva Collects |
Transport |
Reference Config Path |
|---|---|---|---|---|
|
DOCA Telemetry Service |
NICs, DPUs, compute nodes |
NIC and DPU metrics (ethtool, sysfs); compute node syslogs |
OTEL; rsyslog |
|
|
Cumulus Linux |
Scale-out and N/S Ethernet switches |
Interface metrics, platform statistics, BGP state, LLDP; switch syslogs; packet drop events (WJH) |
OTEL; rsyslog; gNMI dial-out |
|
|
NMX-Telemetry |
NVLink switches (NMX-T) in GB and Vera Rubin racks |
NVLink switch telemetry. Telemetry is on by default on NMX-T. To collect it, configure NVIDIA Minerva as described in the README in |
Prometheus scrape (TCP 9352) |
|
|
Slurm |
Job scheduler |
Job events and node state events; Slurm daemon logs |
OTLP; rsyslog |
|
|
NCCL |
GPU-to-GPU fabric (distributed training) |
Per-device, per-collective, and event telemetry via the NCCL Net Inspector Profiler Plugin; provides visibility into fabric health, collective-level performance, and transport reliability |
OTEL |
|
Collecting NMX-T telemetry is separate from NVLink management. To register NMX-C and NMX-T, view the NVLink inventory, and manage GPU partitions through the NVIDIA Minerva NVLink API, see NVLink Management.
For WJH (What Just Happened) configuration on Cumulus Linux switches, see the Cumulus Linux WJH documentation.
Last updated: