← Back to reference

Supported Network Matrix

NetworkStatusNotes
RoCEv2Fullethtool, mlxlink, devlink, CNP
Ethernet (TCP)Best effortPSI + TCP retransmit eBPF
SONiC / gNMIFullgNMI/JSON-IETF subscriptions
UFMFullREST
Fabric ManagerFull/fm/v2/nvlinks, /fm/v2/sxid/events

Per-network Notes

InfiniBand

Per-port counters: symbol_error, port_xmit_wait, link_downed, link_error_recovery, excessive_buffer_overrun, port_rcv_remote_physical_errors, local_link_integrity_errors, VL15_dropped, plus per-VL counters.

ibdiagnet / ibstat / ibqueryerrors integrated.

RoCEv2

ethtool -S <iface> for per-prio PFC and counter stats, mlxlink -p <dev> for module telemetry, devlink dev <iface> eswitch show for switch QoS.

CNP detection: cnp_rcv, cnp_send, rp_ecn_marked, rp_ecn_handled.

gNMI/SONiC switch collector behind --roce-gnmi (uses tonic + protobuf).

sFlow/NetFlow export parsing for loss + retransmit.

GPUDirect / RDMA

/sys/module/nvidia_peermem/version, modinfo nvidia_peermem, dmesg | grep peer, NCCL env NCCL_IB_HCA, NCCL_P2P_LEVEL, NCCL_NET_GDR_LEVEL.

Compatibility Tested

  • mlx5 driver 5.x and 6.x.
  • Mellanox ConnectX-5/6/7.
  • Broadcom PEX89048 switches.
  • Cisco N9K switches (gNMI).

Unsupported

  • Pre-mlx5 hardware (CX-3) — basic sysfs only.
  • Non-Mellanox RDMA (AWS EFA, libfabric-based) — best effort.