A transit ML cluster finishes InfiniBand leaf provisioning and must prove subnet manager HA. What should operators test?
Select an answer to reveal the explanation.
Short Explanation
HA on paper is not HA on the rails. After SM configuration, force a failover and confirm the transit InfiniBand fabric keeps managing itself.
Full Explanation
After a transit ML cluster finishes InfiniBand leaf provisioning, operators must prove subnet manager HA by testing InfiniBand HA failover behavior after initial subnet manager configuration. A failover test confirms the standby SM takes over fabric management so transit training jobs are not stranded if the primary SM fails—paper HA is insufficient for civic ML rails.
Ethernet management network reachability and latency checks across leaf switches fail by concept: management-plane reachability and latency do not prove InfiniBand subnet manager failover. Subnet manager priority values in configuration files on both switches fail: reviewing static priority settings does not demonstrate an actual standby takeover or prove HA failover behavior. InfiniBand link speed and width validation across leaf to spine uplinks fails: validating link rate and width checks physical fabric parameters, not subnet manager HA failover behavior.
Exam caveat: validate IB HA with an actual SM failover test after configuration; Ethernet management reachability and latency checks, SM priority file reviews, or IB link speed and width validation do not demonstrate subnet manager HA.
Operational check: in a transit maintenance window, trigger controlled primary SM failure or takeover, confirm standby assumes SM mastership, verify IB fabric management continuity and host path records, restore HA pair health, and log results for municipal change control.