Topology

The Topology page maps your network in a hierarchical, real-time display, from data centers and pods down to switches, racks, NVLink switch trays, and servers. Select a job to see only the devices it used, and overlay anomalies, events, and metrics to see which components are affecting performance. To open it, go to Topology > Search.

Topology Page Layout

• View: select a view, and create, save, and manage views. The controls are the same as on Trends. See Manage Views.

• Filter: the time range and filters that limit what the topology shows. See Filter the Topology.

• Show: the hierarchy and the overlays drawn on the topology. See Choose What the Topology Shows.

• Job Explorer: select a job to see its topology. See Job Topology.

• Legend: the node types on the canvas.

• Selected Entities: details of the nodes and links you select. See Device and Link Details.

• Toolbar (below the right panel): Reset, Auto Layout, Zoom In, Zoom Out, Settings, and layers.

• Minimap (lower left): an overview of the whole topology.

Filter the Topology

Click + next to Filter and select a filter:

To change the time range, click the time chip, for example Latest day.

Choose What the Topology Shows

Click + next to Show and select:

• Hierarchy: organize the topology into hierarchical groups. Links are drawn only while Hierarchy is on.

• Anomaly: mark nodes and interfaces that have anomalies.

• Event: mark nodes and interfaces that have events.

• Metric: select a measure to show its values on links, for example ifHCOutOctets.

Multi-Layer Hierarchy

With Hierarchy on, the topology is organized in levels. The Legend shows each node type:

Each group's title lists the node types it contains, for example Leaf / Spine / VPN. An NVLink rack (GB200, GB300, or Vera Rubin) expands to an NVLink switch / Server group. An NVLink switch expands to its ASICs, and a server expands to its GPUs. A server shows its GPUs only after its compute tray hostname is set. See Set Compute Tray Hostnames.

To hide a node type, click it in the Legend. Click it again to show it.

Navigation

• Expand or collapse a node: click the ⊕ or ⊖ button under it.

• Choose which nodes a group shows: click the group's filter icon (Grid nodes filter), select the nodes, and click OK.

• Show a group's links: click Show connections in the group's header.

• Zoom and pan: use the mouse wheel and drag the canvas, or use Zoom In and Zoom Out. As you zoom out, node names are hidden and only the icons remain.

• Reset the view: click Reset.

To arrange the groups in tiers, click Auto Layout in the toolbar. To keep the topology arranged as nodes expand and collapse, select Auto Layout in the Settings menu. It's off by default. While it's on, positions you save or move by hand are ignored. Auto Layout is available only when Hierarchy is in Show.

The Settings menu also has:

• Display nodes with Anomalies or Events only

• Display interfaces with Anomalies or Events only

• Display expanded groups only

• Highlight Events

• New Nodes: what the canvas does when nodes are added: Zoom to fit (the default), Pan to new nodes, or Keep camera in place.

Device and Link Details

Click a node to add its details under Selected Entities, for example Name, Management IPv4 Address, Role, Fabric Type, SU, and Oper Status. A rack shows its Name, Domain UID, Family (for example, gb300), and Topology (for example, 72x1), with the rack's members listed under Servers and NVLink switches. An NVLink switch also shows its Tray, Slot ID, and ASICs. An ASIC shows its Component Name, Health (NMX-C), Firmware (NMX-T transceiver), Product (NMX-T), and Node GUID. A GPU also shows its GPU UUID and PCI location (DCGM). Each selection adds a card. To remove one, click its X.

Click the connection between two devices to list the links between them, with the number of links. Each row shows the interface at each end, for example swp13s0 and eth_p0_r0. For BGP sessions, the list is titled BGP Links, and each row shows the BGP address at each end. Click a row to add the link under Selected Entities. The card has a tab for each end.

Show the GPUs in a Partition

1. Expand a server to show its GPU group.

2. Click the group's filter icon.

3. Under GPU partition, select a partition, or Unassigned for GPUs that aren't in a partition. Each GPU in the list shows its partition.

4. Click OK.

Job Topology

Use the Job Explorer panel to select a job ID and see that job's topology: the switches, racks, and servers it used, with the rest of the cluster removed. Selecting a job also sets the Job Topology filter. The job list comes from Slurm scheduler data, so it's empty until NVIDIA Minerva receives job data from Slurm. To filter the list by Subscriber Name or Job Status, click the filter icon in Job Explorer. Click Show Anomalies / Events to highlight what happened during the job, for example an XID error, and click Hide Anomalies / Events to hide them. Servers are shaded by how many events and anomalies they had; links aren't shaded. To see only the nodes that had problems, select Display nodes with Anomalies or Events only in the Settings menu.

The job topology goes down to the server level. Events and anomalies are shown per server, not per GPU or interface.

Integrated Anomaly Mapping

NVIDIA Minerva overlays performance telemetry directly onto topology nodes, so you can visually correlate job performance with network health without switching views. Common overlays include:

• Anomalous PFC (Priority Flow Control): visualizes where pause frames are being generated, indicating downstream congestion

• RoCE retransmits: highlights links or adapters experiencing packet loss that triggers RDMA retransmission, often a precursor to job timeouts

• Rare log anomalies: surfaces unique log patterns via NVIDIA Minerva's ML engine

Save and Share Topology Views

1. Set the filters, overlays, and layout you want to save.

2. In the View panel, click + (Create).

3. Enter a View Name and click OK. The view is private. To make it visible to all users in your tenant, select Publish (visible to all users).

To switch views, select one in the View panel, for example topology_v1. For the other view controls, see Manage Views.

BGP Sessions

NVIDIA Minerva discovers every BGP session in the fabric and models it end to end, with no manual configuration. Each switch reports its own side of a session, and NVIDIA Minerva joins the two sides into one session. You can see sessions as links in the topology and look them up in the BGP Sessions table. For fabric-wide BGP health and recovery, see the BGP Operations dashboard.

BGP session modeling requires:

• Cumulus Linux 5.18 or later on the switches, with BGP routing telemetry exported over OpenTelemetry. Version 5.18 adds the router ID and local address to BGP telemetry, which NVIDIA Minerva needs to join the two sides of a session.

• LLDP between peers that use unnumbered BGP. NVIDIA Minerva identifies an unnumbered peer through its Layer 2 (LLDP) neighbor. Without LLDP, the session is listed without a peer device and isn't drawn as a link.

Numbered sessions (IPv4 or IPv6 peer address), unnumbered sessions, and dual-stack peerings are all supported.

How a BGP Session Is Modeled

Each switch reports its side of every session: its router ID, the VRF, its local address, the peer address, the AS number, and the session state. When the switch at the other end reports its side, NVIDIA Minerva joins the two into one session. It doesn't matter which side reports first.

• Until both ends report, the session shows only one side, and the peer device is blank. NVIDIA Minerva doesn't guess the peer device from the peer address.

• A session that is down but still reported stays visible with its current state. Shutting down a session doesn't remove it.

• A session is removed after its telemetry stops arriving. If a switch stops reporting, its side of each session is removed; the other end keeps its side until it also stops reporting.

View BGP Sessions in the Topology

1. Go to Topology > Search. Make sure Hierarchy is in Show.

2. Expand the hierarchy to the switches you want to check, and click Show connections in their group's header. Each BGP session is a link labeled BGP followed by its address.

3. Click the connection between two switches. The BGP Links list shows every session between them.

4. Click a session to see it under Selected Entities. The card has a tab for each end, with its Device Name, Peer ID, Type (for example, BGP:EBGP), VRF / VPN, and BGP State (for example, ESTABLISHED).

Look Up Sessions in the BGP Sessions Table

1. Go to Topology > Provision > Metadata > BGP Sessions.

2. Check Total BGP Sessions at the top of the page.

3. Search, filter, or sort the table to find the sessions you need. For example, filter Device Name to see all of one switch's sessions.

The table has one row per side, so a session that both ends report appears twice, once from each switch. A blank Peer Device Name means the far end hasn't reported its side yet or, for an unnumbered session, that LLDP hasn't identified the neighbor.

Maintenance Mode

Place a device in Maintenance Mode during scheduled maintenance windows to avoid false alerts and skewed health metrics. When a device is in Maintenance Mode, NVIDIA Minerva does not send notifications or events for that device.

Put a Device in Maintenance Mode

1. Go to Topology > Provision > Metadata > Devices and find the device.

2. At the end of the device's row, click .

3. Set Admin Status to Maintenance, click OK, and then click Yes to confirm.

To end Maintenance Mode, set Admin Status back to Enabled.

Last updated: October 29, 2026