> ## Documentation Index
> Fetch the complete documentation index at: https://gcore-doc-1046.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.
# Manage a Bare Metal GPU cluster
After creating a GPU cluster, use the cluster details page to monitor nodes, resize the cluster, manage power state, configure network interfaces, and delete cluster resources.
## Access cluster details
To view and manage an existing cluster, open the cluster details page.
1. In the [Gcore Customer Portal](https://portal.gcore.com), navigate to **GPU Cloud**.
2. In the sidebar, expand **GPU Clusters** and select **Bare Metal GPU Clusters**.
3. Click on a cluster name to open the details page.
The cluster details page displays summary information in the header panel:
| Field | Description |
| ---------- | ------------------------------------------------------------------------------------------------------------ |
| Cluster ID | Unique identifier for the cluster |
| Pkey ID | InfiniBand Partition Key ID. Displayed as "-" if InfiniBand is not configured or during cluster provisioning |
| OS Distro | Operating system image installed on all nodes |
| Status | Current cluster state |
| Region | Data center location |
| Plan | Monthly pricing plan for the cluster |
The page is organized into tabs: **Overview**, **Power**, **Networking**, **Tags**, **User Actions**, and **Delete**.
## View cluster nodes
The **Overview** tab lists all nodes (servers) in the cluster. Each node entry shows the node name, flavor, assigned IP addresses, status, timestamps, cost, and other metadata. The table supports filtering by name, date range, and status.
All nodes in a cluster share the same configuration (image, network settings, file shares) defined at cluster creation. These settings cannot be changed after the cluster is created.
## Resize a cluster
Cluster size can be adjusted after creation by adding or removing nodes.
To resize a cluster:
1. On the **Overview** tab, click **Resize Cluster**.
2. Adjust the instance count using the **+** and **-** buttons.
3. Click **Resize**.
New nodes inherit the cluster's original configuration defined at creation time. The maximum number of nodes is limited by the current stock availability in the selected region.
While a cluster is in the Resizing state, most management actions are unavailable. Bare Metal GPU nodes, especially H100 configurations, may take 15–40 minutes to provision.
Scaling down to zero nodes is not allowed and returns an error. A cluster must contain at least one node. To remove the cluster entirely, use the **Delete** tab.
When scaling down, the system removes a random node from the cluster. To delete a specific node, use the per-node delete action instead of resize.
### Delete a specific node
To remove a specific node without random selection:
1. Locate the node in the cluster list.
2. Click the actions menu (three dots) on the node row.
3. Select **Delete**.
4. Confirm the deletion.
Deleting the last node in a cluster automatically deletes the entire cluster. This applies to both UI and API/Terraform operations. When using the API or Terraform to delete nodes, no warning is displayed before the cluster is removed. Ensure at least one node remains if you want to preserve the cluster.
## Power actions
Power actions control the running state of cluster nodes. Actions can be applied to individual nodes or in bulk.
### Individual node actions
To control a single node:
1. Locate the node in the cluster list.
2. Click the actions menu (three dots) on the node row.
3. Select the desired action:
* **Power on**: Start the node
* **Power off**: Shut down the node
* **Soft reboot**: Graceful restart of the operating system
* **Hard reboot**: If soft reboot fails, force restart the node
* **Rebuild**: Reinstall the original operating system image used at cluster creation. All data on local storage is deleted.
* **Replace**: Delete the current node and provision a new one with the same configuration (flavor, image, network settings). Available in the UI and via the [Replace bare metal GPU cluster server](/api-reference/cloud/gpu-bare-metal/replace-bare-metal-gpu-cluster-server) API endpoint. Use this to recover a node stuck in a failed state without changing the cluster setup.
### Bulk actions
To apply actions to multiple nodes simultaneously:
1. Select nodes using the checkboxes in the nodes table.
2. Click **Group actions** in the toolbar.
3. Select the action to apply to all selected nodes.
Alternatively, use the **Power** tab to perform soft or hard reboot on all cluster nodes at once.
## Network interfaces
The **Networking** tab displays network interfaces for each node.
Click on a node name to expand its interface details.
Interface types include:
| Type | Description |
| ---------- | ----------------------------------------------------------------------- |
| Public | External IP address for internet access |
| Private | Internal network for communication with other cloud resources |
| InfiniBand | High-speed, low-latency inter-node network for GPU-to-GPU communication |
For flavors with InfiniBand, multiple InfiniBand interfaces are created automatically (by default, 8 for H100 configurations). These appear as "GPU-cluster ib-subnet" entries in the interface list.
Click on an interface to expand its details, including IP address, network configuration, and other network information.
### Modify network interfaces
To add network interfaces on a specific node:
1. Navigate to the **Networking** tab.
2. Click on a node to expand its details.
3. Click **Add Interface** or **Add Sub-Interface**.
4. Configure the interface type (public or private) and IP allocation settings as supported by the service.
Only network interfaces can be modified on individual nodes after cluster creation. Image and storage configuration cannot be changed at the node level. InfiniBand interfaces are managed automatically and cannot be modified or deleted.
## Console access
For troubleshooting or when SSH access is unavailable, use the browser-based console.
1. In the **Overview** tab, locate the node you want to access.
2. Click **Open Console** in the node row.
3. The console opens in a new browser tab using noVNC.
4. Log in using the same credentials as for SSH access.
The console provides direct terminal access to the node, useful when network connectivity issues prevent SSH access.
## Tags
Tags are key-value pairs used to organize and categorize clusters. Tags are applied at the cluster level and inherited by all nodes.
To manage tags:
1. Navigate to the **Tags** tab on the cluster details page.
2. Enable the **Add custom tags** checkbox.
3. Enter the key and value for each tag.
4. Click **Save changes**.
Use tags for billing allocation, environment identification, or organizational purposes.
Tags can be modified even while the cluster is in the Resizing state, unlike most other management actions.
## User actions
The **User Actions** tab displays a log of all operations performed on the cluster, including creation, deletion, resize, power, and network actions.
Use the date and action type filters to narrow results.
This log is useful for auditing and troubleshooting.
## Delete a cluster
When a cluster is deleted, local NVMe storage is permanently erased; file shares and object storage remain intact.
To delete a cluster:
1. Navigate to the **Delete** tab on the cluster details page.
2. Click **Delete Cluster**.
3. Confirm the deletion.
Cluster deletion is irreversible. Data on local NVMe storage cannot be recovered.
## Automating cluster management
The [Gcore Customer Portal](https://portal.gcore.com) is suitable for managing individual clusters. For automated workflows—such as CI/CD pipelines, infrastructure-as-code, or batch provisioning—use the GPU Bare Metal API.
The API allows:
* Starting, stopping, and rebooting cluster nodes
* Resizing clusters and replacing individual nodes
* Managing network interfaces on cluster servers
* Deleting clusters and individual nodes programmatically
The [GPU Bare Metal API reference](/api-reference/cloud/gpu-bare-metal/create-bare-metal-gpu-cluster) provides endpoint details and code examples.