Public-health InfiniBand switches need baseline services before GPU hosts attach. What initial steps should operators complete?
Select an answer to reveal the explanation.
Short Explanation
Don't activate OpenSM on the GPU hosts for initial discovery or wait to sync switch firmware until after the GPUs are attached; the subnet manager and firmware need to be ready before the fabric comes up. Complete subnet manager, GUID, and port provisioning first, then attach the public-health GPU hosts. Don't set PKey partitions or HCA MTU during NCCL tuning with GPUs already attached—those are baseline fabric settings, not performance tweaks.
Full Explanation
Public-health InfiniBand switches need baseline services before GPU hosts attach, so operators should complete baseline InfiniBand configuration including subnet manager, GUID, and port provisioning before GPU attach. Running OpenSM on the GPU hosts for initial discovery and syncing switch firmware only after attachment misplaces the subnet manager on endpoints and defers required firmware baseline work past the point where GPU hosts join the fabric. Depending on GPU-host OpenSM discovery and post-attach firmware sync fails by concept: it uses the production hosts themselves to bring up the subnet and treats necessary switch firmware as an afterthought, extending outage on an unfinished subnet. Establishing InfiniBand PKey partitions and HCA MTU settings during NCCL performance tuning with GPUs attached fails because partition keys and MTU are baseline fabric parameters that must be provisioned before hosts attach, not adjusted while GPUs are running performance tests. Configuring only Spectrum BGP-EVPN and assuming it manages InfiniBand partitions fails because BGP-EVPN is Ethernet overlay control plane—it does not manage IB partitions or replace the IB SM.
Exam caveat: IB baseline = SM + GUID/port provisioning before GPU attach; Spectrum EVPN does not manage InfiniBand, and PKey/MTU tuning is not a substitute for pre-attach baseline provisioning.
Operational check: bring up SM, record GUIDs and port maps for public-health IB switches, verify subnet management is healthy, then attach GPU hosts and confirm IB links and partitions before enabling clinic inference.