問題文
A team is specifying shared storage for a new training cluster. A vendor proposes a system tuned for one large sequential pass on the grounds that training reads the dataset once. Which correction should the team make?
選択肢
- Streaming reads are the right target, but the capacity should be multiplied by the number of iterations, because each pass needs its own copy of the data staged ahead of the accelerators.
- Staging the dataset onto each node's local disk once removes the requirement.
- The dataset is written more often than it is read, so write performance dominates the requirement.
- Training iterates, so the same data is read again and again; the decisive characteristic is re-read performance rather than a single pass.