問題文
A data center team is asked to describe how a distributed training run and an online inference service differ in what they demand from the infrastructure. Which description is correct?
選択肢
- Neither workload places any particular demand on the network, so the same oversubscribed access layer that serves office applications is appropriate for both.
- Both are dominated by sustained east-west traffic, on the grounds that inference also has to exchange gradients between the accelerators after every request.
- Training is dominated by north-south traffic to the storage system, while inference generates no network traffic at all once the model has been loaded into memory.
- Training is dominated by sustained east-west traffic between the accelerators that hold the model, while inference is dominated by many small north-south requests whose response time is what users notice.