Skip to main content

Metal Provisioning Overview

Bridge provisions bare metal servers (compute nodes) using Canonical MaaS (Metal as a Service) or BCM (Bridge Cluster Manager) over the Out-of-Band (OOB) management network. Once provisioned, servers are prepared with the GPU software stack and registered in the Bridge catalog, ready for tenant allocation.

Provisioning Methods

MethodToolHow It Works
Canonical MaaSMaaS micro-serviceRedfish-based discovery, DHCP/TFTP OS deployment via OOB
BCMBridge Cluster ManagerAlternative provisioning backend
CSV ImportManualUpload server details via CSV file — no network discovery needed

For large deployments, MaaS or BCM provides automated discovery and commissioning. For smaller deployments or when server details are already known, CSV import is simpler.

Day 0 Provisioning Sequence

Discovery and Commissioning

  1. Configure DHCP and TFTP — Bridge configures the TFTP server with OS images and the DHCP server for PXE boot, enabling network-based OS installation.
  2. Redfish discovery — Bridge uses Redfish (BMC API) to discover compute nodes over the OOB network. This retrieves the in-band interface MAC address and BMC credentials for each server.
  3. Fetch hardware details — Bridge queries each server for PCIe device details and other hardware properties (GPU model, memory, NIC type).
  4. Commission server — Bridge commissions the server in MaaS/BCM, which deploys the OS via PXE boot over the OOB network.
  5. Populate catalog — Discovered hardware properties are used to create metal and VM flavors in the Bridge catalog.

Post-Provisioning Preparation

After the OS is installed, Bridge prepares each server for GPU workloads:

Preparation StepPurpose
CUDA librariesEnable GPU compute workloads
MOFED (Mellanox OFED) stackEnable RDMA networking (RoCE, InfiniBand)
Kernel modulesLoad required GPU and NIC drivers

For VM-based deployments, Bridge additionally configures the host for KVM virtualization:

ConfigurationPurpose
intel_iommu=onEnable IOMMU for PCIe device passthrough
PCIe GPU passthroughAssign GPU directly to VM
Compute NIC passthrough1:1 mapping between GPUs and compute NICs
Macvlan virtual NICConverged network access for VMs

Security and Isolation

OOB Network Isolation

All provisioning traffic is carried on the OOB management network, which is isolated from tenant compute and converged networks. Tenant access is not granted until network isolation is fully configured.

Compute Deallocation

When a tenant's compute allocation is revoked, Bridge wipes the server disk via MaaS or BCM — writing patterns to ensure no tenant data remains before the server is returned to the free pool.

For VMs, the VM disk is similarly wiped upon deletion.

Metal Provisioning with BF3

For servers equipped with BlueField-3 DPUs in Zero Trust mode, the provisioning sequence differs. The host OS is provisioned through the DPU rather than directly via MaaS over the OOB network. See BlueField-3 (BF3) for the full DPU provisioning flow.