NCCL

Ingest NCCL Inspector Prometheus Exposition (2.29.2+, 2.31.2 names)

Accept NCCL Inspector Prometheus-format output (supported since NCCL 2.29.2) as a text/plain body — e.g. curl -X POST --data-binary @nccl_inspector.prom this-endpoint. The four official NCCL 2.31.2 metric names are mapped exactly: nccl_bus_bandwidth_gbs / nccl_p2p_bus_bandwidth_gbs → bandwidth_gbps, nccl_collective_exec_time_microseconds / nccl_p2p_exec_time_microseconds → duration_ms (µs→ms); event_type collective|p2p comes from the metric name. Labels honoured: version (v5.1 aggregated / v5.2 verbose), message_size bucket strings (stored verbatim, e.g. "4-5GB"), coll_sn/p2p_sn (per-op identity in verbose mode), p2p_peer, comm_id, rank, coll, hostname. Other nccl_* names fall back to a tolerant token mapping (bandwidth/duration/count); unknown metrics are ignored. Samples sharing (comm, rank, collective, node, event_type, message_size, sn, peer) merge into one event. ?node=/?job= supply defaults when the labels lack them (typical for the textfile route).

post/api/nccl/ingest-prometheus

Request

  • Base URL: https://api.fiber.ac (the document declares no server; this is the origin it was published from)
  • URL: https://api.fiber.ac/api/nccl/ingest-prometheus
  • Auth: none declared

Query parameters

nodestring

Default hostname when samples carry no node/hostname label

jobstring

Default job id when samples carry no job label

Response

Count of parsed/inserted events + Inspector schema version histogram

unknown required

Changes

No recorded changes to this operation across all 1 revision of this API.