Why it matters

Topology shapes performance. Understanding shapes efficient placement.

Advertisement

The architecture

Discover topology: NVML, PCIe tree.

Place: minimize cross-domain comm.

Topology-aware schedulingDiscover topologyNVML + PCIePlace processesaffinity groupsCollective algotopology-awareK8s device plugin exposes topology; NCCL auto-tunes based on topology
Topology-aware.
Advertisement

How it works end to end

NVIDIA GPU topology: nvidia-smi topo.

K8s: device plugin with topology info.

NCCL: topology auto-detected.