Public-health InfiniBand latency complaints need a measured write path between GPU nodes. Which command fits?
Select an answer to reveal the explanation.
Short Explanation
Latency complaints need a stopwatch built for InfiniBand writes. Use ib_write_lat between public-health GPU nodes to measure the civic AI write path.
Full Explanation
Public-health InfiniBand latency complaints need a measured write path between GPU nodes. The ib_write_lat command measures InfiniBand write latency on civic AI paths, which matches a latency complaint with a purpose-built RDMA latency tool rather than bandwidth or management-plane substitutes.
Why the other options fail by concept: ib_write_bw alone answers bandwidth, not latency, so it misfits a pure latency complaint. ib_read_lat measures read latency rather than write latency, so it does not exercise the write path the complaint asks about. ib_send_lat measures send latency rather than write latency, so it does not measure the write-path latency the complaint requires.
Exam caveat: use comparable message sizes, queue pairs, and directions to the production path; a tiny idle latency number may not represent loaded public-health training traffic. Pair ib_write_lat with fabric health context so you do not chase host tuning while SM or switch health is degraded.
Operational check: between the implicated public-health GPU nodes, run ib_write_lat with documented parameters, compare to civic baselines, archive results with the incident, and escalate fabric or host owners based on whether latency exceeds the agreed write-path budget.