A data center design must keep training jobs close together. Which network layout helps?
Why B is the answer
A non-blocking fabric, with few hops
Training traffic is heavy and synchronized.
In detail
- Training traffic is heavy and synchronized.
- Training traffic is heavy and synchronized.
Use case
An illustrative scenario, not a real incident.
A training cluster uses a non-blocking fabric, so synchronized traffic moves in few hops.
See it in different ways
mindmap
root((AI storage))
Reads
High throughput
Parallel file system
Writes
Fast checkpoints
Direct path
GPUDirect Storageflowchart LR S0["Place jobs near each other"] --> S1["Few hops"] --> S2["Non-blocking links"]
Key point: Training needs non-blocking fabrics.
Why the other options are wrong
- A. A chain of unmanaged switches
- A chain adds hops.
- C. One oversubscribed uplink per rack
- Oversubscription slows it.
- D. A campus-style three-tier network
- That is not an AI fabric.