How Bare Metal Servers Combine Physical Power with Cloud Flexibility

Introduction

Bare metal servers combine the dedicated hardware of physical servers with the flexibility of cloud services. Users get exclusive access to CPU, memory, and storage while benefiting from on‑demand provisioning, rapid deployment, and cloud‑style management.

Key Characteristics

Hardware exclusivity : CPU, memory, and storage are not shared with other tenants.

Cloud‑style management : Provisioning, configuration changes, and OS reinstallations are performed through a cloud platform.

No virtualization overhead : The server runs directly on physical hardware, delivering near‑native performance.

Comparison with Traditional Physical Servers

Unlike ordinary physical servers, bare metal servers offer cloud‑native features such as elastic scaling and automated lifecycle management, while retaining the performance advantages of dedicated hardware.

Typical Use Cases

AI/LLM inference and training

Big data processing

High‑performance computing (HPC)

Database workloads

Middleware services

OpenStack Ironic Overview

Ironic is an OpenStack project that provides management capabilities for bare metal servers. It handles enrollment, deployment, lifecycle management, and monitoring without a hypervisor layer.

Ironic Architecture

Ironic‑api: RESTful API for interaction.

Ironic‑conductor: Executes operations such as deployment and power control.

Nova: Manages the bare metal lifecycle.

Neutron: Supplies flexible networking.

Cinder: Provides scalable block storage.

Glance: Supplies OS images.

Drivers: Communicate with hardware via IPMI, Redfish, etc.

Business Workflow

Cloud administrators register bare metal nodes and connect them to the physical network.

Users request bare metal instances; the platform installs the chosen OS.

The platform configures the required cloud network.

Appropriate cloud disks are attached to the bare metal server.

Users manage the server lifecycle (start, stop, reboot, delete).

Network and Storage Challenges

Ironic’s native support is limited to flat and VLAN networks, leading to VLAN ID limits, inflexible networking, and coupled cloud‑on‑cloud routing.

DPU‑Based Solution

MaShang Cloud extends Ironic with a Data Processing Unit (DPU). The DPU runs its own OS, hosts OVS and storage agents, and uses its PCI configuration capabilities to expose network and storage devices to the bare metal server. This enables overlay networking (OVN) and direct storage access via NVMe or Ceph.

Network Stack

Physical NICs (PF) are virtualized via SR‑IOV into VFs, which connect through OVS to the DPU’s underlay ports. Overlay connectivity is achieved by deploying OVN on the DPU, establishing VXLAN tunnels to other DPU nodes.

Storage Stack

Cloud disks are attached to the DPU (via iSCSI or Ceph RBD) and presented to the bare metal server as virtio‑blk devices using NVIDIA DOCA and SPDK, allowing the server to access storage without host‑side virtualization overhead.

Implementation Details

The overall architecture integrates OpenStack Ironic, Nova, and the DPU agents to manage the full lifecycle of bare metal instances, including network and storage provisioning.