A 311 model-farm InfiniBand bring-up needs a complete initial checklist. What should that checklist cover?
Select an answer to reveal the explanation.
Short Explanation
GPU driver versions and Kubernetes labels don't tell you whether the InfiniBand fabric itself is healthy. Your initial checklist needs subnet manager HA, GUID inventory, and management reachability before those 311 models go live.
Full Explanation
A complete InfiniBand initial-configuration checklist for a 311 model-farm civic fabric must cover subnet manager HA, GUID planning/inventory, and management reachability—not GPU driver versions, NVLink status, and compute node health checks alone. SM HA proves the fabric can be managed through a node failure; GUIDs identify HCAs and switches for membership and troubleshooting; management reachability ensures operators can reach SMs and IB switches when models are live. Omitting these items leaves 311 racks with mystery outages and no municipal ops footing. A checklist of only GPU driver versions, NVLink status, and compute node health checks fails by concept: host-side driver, NVLink, and compute health data do not validate IB SM HA, GUID layout, or management path. Treating Ethernet WJH telemetry, interface counters, and BGP session states as a full substitute for InfiniBand HA checks fails because Ethernet What-Just-Happened telemetry, counters, and BGP data do not confirm IB SM master/standby behavior or IB management reachability. Checking only the Network Operator version, node labels, and SR-IOV policies for Kubernetes fails by scope: Kubernetes networking settings assume an underlying fabric; they do not replace InfiniBand switch bring-up or SM HA verification. Exam caveat: candidates mix Ethernet WJH telemetry, interface counters, BGP states, and Kubernetes Network Operator details with InfiniBand initial-config essentials. Operational check: before 311 models go live, verify SM master and standby, inventory switch and HCA GUIDs against the civic plan, and prove SSH/API reachability to IB management endpoints from the ops network.