UFM for InfiniBand fabric
UFM (NVIDIA Unified Fabric Manager) manages InfiniBand fabrics:
- Subnet manager (SM) functionality
- Fabric health monitoring and congestion control
- Performance analysis
UFM HA (High Availability):
- Primary/standby setup with SSH trust between nodes
- Split-brain recovery uses fencing to ensure only one primary is active
- CIDR separation documents specific recovery steps
Upgrading UFM: Always stop services → upgrade package → restart. Follow the version-specific upgrade path from NVIDIA docs.
