Deploying 360Stack Bare Metal Servers on OpenStack: Architecture and Network Modes

Bare metal services provide physical servers as resources to tenants in cloud environments; tenants can request, manage, and configure these unprovisioned machines.

In OpenStack, the Ironic component handles bare metal provisioning and interacts with Keystone, Nova, Neutron, Cinder, and Swift. Nova creates instances via the same API as virtual machines, but uses different scheduling flavors and drivers (PXE/IPMI for physical nodes).

The deployment (Provision/Deploy) phase assigns a bare metal node, sets it to PXE boot, uses TFTP to fetch kernel and ramdisk, runs the ironic-python-agent, and retrieves the user image from the Ironic control node.

The runtime (Tenant) phase starts the installed system image and runs tenant workloads; after image download, Ironic triggers a reboot to join the tenant network.

Three network modes are supported: flat, VLAN, and VXLAN. Flat mode uses a single external network for both deploy and tenant traffic. VLAN mode (via networking‑generic‑switch) tags the switch ports with deploy and tenant VLANs, optionally using a single VLAN for both phases. VXLAN mode is enabled by installing networking‑l2gw, configuring L2 gateways, and establishing tunnels between compute nodes and the switch to bridge VM and bare metal networks.

Key configuration steps include ensuring the management network can reach the IPMI interfaces, using the same flat network for deploy and tenant traffic, creating a virtual NIC on the Ironic control node for image transfer, and verifying tunnel and MAC learning via ovsdb-client.

Related documentation: OpenStack Ironic official docs and a blog post on neutron‑l2gw integration.