Checkpoint writes are slow on the city training farm. How should the UCS storage policy evolve?

When checkpoint writes are slow on the city training farm, the UCS storage policy should evolve to add an NVMe-backed datastore suited to high-performance checkpoints. Checkpoint I/O is latency- and throughput-sensitive; media and path quality dominate step time and overall training efficiency on municipal GPU farms. Domain 3.2.c covers storage-policy changes that introduce higher-performance targets for AI workloads instead of leaving checkpoints on slow tiers. Adding a read-cache accelerated pool suited to high-performance checkpoints doesn't fix slow checkpoint writes because read caching accelerates reads, not write throughput and latency. Adding a SAS-backed datastore suited to high-performance checkpoints keeps checkpoints on slower, higher-latency media than NVMe and won't deliver the throughput needed for checkpoint writes. Enabling inline compression tuned for high-performance checkpoints adds processing overhead to the write path and increases latency instead of providing faster media for checkpoint writes. Exam caveat: NVMe-backed may be local NVMe, NVMe-oF, or an NVMe datastore depending on platform—exam intent is faster media via storage policy, not inline compression tuned for high-performance checkpoints or a SAS-backed datastore suited to high-performance checkpoints. Operational check: measure checkpoint write duration under the current policy, update the UCS storage policy to present an NVMe-backed datastore to GPU nodes, redeploy or reassociate profiles, rerun checkpoint-heavy training, and confirm write time and job step time improve.