What does single-node NCCL testing verify?
Why A is the answer
GPU-to-GPU traffic over NVLink and NVSwitch
Within a node, NCCL traffic travels over NVLink.
In detail
- Within a node, NCCL traffic uses NVLink and NVSwitch.
- Traffic stays in the node.
Use case
An illustrative scenario, not a real incident.
Inside a node, an NCCL test checks traffic between GPUs over NVLink and NVSwitch.
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 --> [*]
flowchart TD
Q{"What are you checking?"}
Q -->|"Inside a node"| A0["Single-node NCCL"]
Q -->|"Between nodes"| A1["Multi-node NCCL"]Key point: Single-node NCCL tests NVLink.
Why the other options are wrong
- B. Traffic between racks over the storage network
- That is not inside a node.
- C. The speed of the out-of-band management network
- That is management traffic.
- D. The switch uplink to the data center's core
- That is outside the node.