Skip to content

Chapter 4: InfiniBand Operations -- The ONYX CLI and Fabric Management · Part 3 of 9

Reading InfiniBand port state

Building the model…

Where the commands in this section run, marked on the cluster. Switch the overlay to see the rest of the fabric.

show interfaces ib status -- the overview

show interfaces ib status gives you the state of all InfiniBand ports on the switch at once. This is your starting point.

switch # show interfaces ib status

IB Interface  Admin  Oper   Width  Speed  PortGUID
------------- -----  ------  -----  -----  --------
IB1/1         Up     Active  4x     NDR    0x506b4b0300a1b200
IB1/2         Up     Active  4x     NDR    0x506b4b0300a1b201
IB1/3         Up     Down    --     --     0x000000000000000
IB1/4         Up     Polling --     --     0x000000000000000
IB1/5         Up     Active  4x     NDR    0x506b4b0300a1b204
...

IB port states — click each to understand it

Admin: Up
Oper: Active
Width: 4x
Speed: NDR
PortGUID: 0x506b4b0300a1b200

Port fully operational. SM assigned LID. Routing programmed. Traffic flows.

Cause: Normal operating state.

Action: No action.

Reading the columns:

Admin: Up -- An administrator has not manually disabled this port. If you see Admin: Down, someone ran shutdown on that interface. Check if it was intentional.

Oper: Active -- The operational state. This is what matters. States you will see:

  • Active -- Port is fully operational. InfiniBand state machine has completed. SM has assigned a LID and programmed the routing table for this port. Traffic flows.
  • Down -- No physical signal. Cable unplugged, broken, or the device on the other end is powered off.
  • Polling -- Physical signal detected but InfiniBand init has not completed. The port is sending InfiniBand polling signals. Usually seen when a device is booting or when the SM has not yet discovered this port.
  • Init -- Physical link established, port entered IB state machine Init state. SM has discovered it but not yet armed it. Brief transient state.
  • Armed -- SM has armed the port. One more transition to Active. Should be very brief.

Width: 4x -- The lane count. NDR InfiniBand uses 4 lanes per port. If you see 1x or 2x, lanes have failed -- check the cable or transceiver. Width directly affects bandwidth: 4x NDR = 400 Gb/s, 2x NDR = 200 Gb/s, 1x NDR = 100 Gb/s.

Speed: NDR -- The link speed tier. NDR (Next Data Rate) = 400 Gb/s at 4x. You might also see HDR (200 Gb/s), EDR (100 Gb/s), or FDR (56 Gb/s) if the link has negotiated down. Negotiation down to a lower speed usually means one side cannot run NDR -- check cable compatibility and NIC firmware.

PortGUID: 0x506b4b... -- The hardware-burned GUID of the device on the other end of this cable. A GUID of all zeros means no device is connected or the device is in a non-IB state.

show interface ib 1/1 -- the detail

When you need more than the overview provides, inspect a specific port:

switch # show interface ib 1/1
IB1/1 is up, line protocol is up (connected)
  Hardware is InfiniBand, address is 506b:4b03:00a1:b200
  Interface index: 1
  MTU 4096 bytes
  Current Speed: NDR, Width: 4x
  Supported Speeds: NDR, HDR, EDR
  Auto-negotiate: enabled
  Link layer: InfiniBand
  Local Identifier (LID): 12
  Remote Identifier (Remote LID): 24
  SM LID: 1
  Capability mask: 0x2514afc8
  Description: "dgx-node-01-rail-0"
  Input rate: 142,847 Mb/s (35.7% utilisation)
  Output rate: 141,293 Mb/s (35.3% utilisation)
  5 minute input rate: 139,482 Mb/s
  5 minute output rate: 138,917 Mb/s

LID: 12 -- This port's Local Identifier as assigned by the SM. LIDs are the routing addresses in InfiniBand -- they are analogous to IP addresses but assigned by the SM, not configured statically.

Remote LID: 24 -- The LID of the device on the other end of this cable. You can use this to find the port's counterpart in UFM.

SM LID: 1 -- The LID of the Subnet Manager. If this shows 0, no SM is active -- the fabric cannot route traffic.

Description: "dgx-node-01-rail-0" -- Interface descriptions are gold for troubleshooting. If your team has labelled ports with the connected device and rail number, a problem report immediately tells you which DGX node and which GPU rail is affected.

Input rate: 142,847 Mb/s (35.7%) -- Current bandwidth utilisation as a percentage of the 400 Gb/s port capacity. 35% utilisation on a training cluster means AllReduce is in progress. Near 100% sustained means a bottleneck.


Read aloudPage 3