What does multi-node NCCL verify?
Why B is the answer
East-west fabric bandwidth
Traffic between nodes crosses the compute fabric.
In detail
- Traffic between nodes crosses the east-west fabric.
- Run it across all pairs that matter.
Use case
An illustrative scenario, not a real incident.
A multi-node NCCL test checks east-west bandwidth on the compute fabric.
See it in different ways
flowchart LR A["GPU"] -->|"Single-node NCCL"| B["NVLink and NVSwitch"] B --> C["GPU"] A -->|"Multi-node NCCL"| D["East-west fabric"] D --> E["GPU in another node"]
stateDiagram-v2 [*] --> SingleNodeStress SingleNodeStress --> SingleNodeNccl SingleNodeNccl --> SingleNodeHpl SingleNodeHpl --> MultiNodeNccl MultiNodeNccl --> MultiNodeHpl MultiNodeHpl --> BurnIn BurnIn --> [*]
sequenceDiagram participant P0 as Node A GPUs participant P1 as Compute fabric participant P2 as Node B GPUs P0->>P1: Collective traffic P1->>P2: East-west bandwidth
Key point: Multi-node NCCL tests the fabric.
Why the other options are wrong
- A. The BMC password on each of the nodes
- That is a security check.
- C. The temperature of the GPUs in one node
- That is a thermal check.
- D. Whether Slurm accounting has been enabled
- That is a scheduler setting.