問題文
A multi-node training job synchronizes gradients after every step. The platform team must state what property of the network matters most for this pattern. Which property is it?
選択肢
- High capacity for north-south traffic and a fast uplink, because the gradients leave the cluster on every step to be aggregated by a central service.
- Consistently low latency and high bandwidth for all-to-all collective patterns, because every step waits for the slowest participant.
- Support for many simultaneous small connections, since each accelerator opens a session per step.
- Tolerance of variable latency, because the framework retries whichever transfer is slow.