NCCL

conceptstub

Collectives, topology detection, transports, and tuning.

Planned content

• Ring vs tree vs NVLS/CollNet algorithms
• Topology detection and NCCL_TOPO_FILE
• Key env vars: NCCL_IB_HCA, NCCL_SOCKET_IFNAME, NCCL_IB_GID_INDEX
• nccl-tests: running and reading all_reduce_perf
• Detecting silent TCP fallback (NET/Socket vs NET/IB)

When collectives are slow, start with the decision tree.