You can now read an InfiniBand fabric from the ONYX CLI and correlate what you find with UFM. You know that InfiniBand's centralised SM model means most routing problems require UFM to fix, not individual switch configuration. You understand that error counters -- especially SymbolErrors -- are the most sensitive early warning of physical layer problems.
You know ibdiagnet as the tool that sees the whole fabric at once: speed mismatches, error hotspots, routing inconsistencies, and performance bottlenecks. And you know how to read the UFM event log to understand not just what failed but when and why.
The next chapter covers the performance layer: NCCL and training communication, how to benchmark the fabric, and how to tune DCQCN parameters to optimise AllReduce performance.