問題文
An operator must choose between building one large fabric that connects all nodes with the same non-blocking capacity, and building a fabric where nodes within a group have full capacity and groups share the upper layer. Which requirement argues for the second?
選択肢
- Most jobs fit within one group, so paying for full capacity between all groups would not be used.
- Jobs regularly span the entire cluster and synchronize their gradients on every single step of the run, so every pair of nodes exchanges data continuously.
- The cluster must support the largest possible single job with no slowdown at any point in the whole fabric.
- The operator wants to avoid any dependence on the fabric's topology when making job placement decisions.