問題文
Each compute node has a large local drive. A team asks whether the dataset should live there instead of on the shared tier, to avoid the network entirely. What is the correct division of roles?
選択肢
- The local drive should not be used at all, since caching happens in memory.
- The local drive can replace the shared tier if every node receives a full copy of the dataset before the run starts, and that keeps the network out of the read path.
- The shared tier gives every node one view of the data; the local drive is useful for caching or staging on top of that, not as a replacement.
- The local drive should hold the dataset and the shared tier should hold only the state written during training, because that removes the read path from the network and leaves the shared tier to a load it can absorb easily.