問題文
A platform team is asked to design the cluster so that adding capacity later does not require re-architecting. They want the design principle that reference architectures for accelerated clusters follow. Which principle is it?
選択肢
- Start with a single flat switch and replace it with a larger one each time capacity doubles, so the topology stays one hop deep.
- Build in repeatable units of a fixed number of nodes with their own leaf switches, and size the upper switching layer for the intended maximum.
- Connect every node to every other node directly so that no switching layer is needed at any scale.
- Add nodes only in the same rack as the existing ones, so the fabric never spans racks.