A housing authority colo cage must size parallel file throughput so many trainers can read the same corpus concurrently. Which storage metric should drive that sizing?

A housing authority colo cage that must serve the same corpus to many concurrent trainers should size storage by aggregate parallel file read throughput and concurrency under multi-trainer load. Parallel file systems or scale-out NAS succeed when the shared aisle can deliver enough simultaneous readers without collapsing into per-client serialization. For civic AI, that metric—not vanity single-stream demos—gates whether GPUs stay fed during corpus-wide epochs.

Sizing only by aggregate parallel file metadata throughput and concurrency under multi-trainer load fails by concept: metadata operations measure namespace access, not bulk data delivery, so it says nothing about whether the file system can sustain concurrent read GB/s to keep trainers fed. Aggregate parallel file read IOPS and concurrency under multi-trainer load fails similarly—it measures operation rate for small reads, not the sustained bulk GB/s throughput the concurrent multi-reader corpus access the stem requires. Aggregate parallel file read throughput and concurrency in single-client warm-cache load fails representativeness: a single warm-cache client hides backend and fabric limits and ignores contention under many readers and the aggregate parallel read pattern the stem describes.

Exam caveat: Domain 2.3 file performance evaluation for AI emphasizes aggregate throughput and concurrency. Single-client warm-cache read throughput can be marketing-useful but is an insufficient acceptance criterion for shared corpora.

Operational check: run multi-reader benchmarks matching trainer count and corpus layout; capture aggregate GB/s and per-client tail latency; expand file nodes or network before housing-authority jobs show GPU idle-from-I/O during concurrent corpus reads.