Building a Private Cloud on Bare Metal with OpenStack

Private Cloud OpenStack Data Center banner

An OpenStack private cloud on bare metal uses physical servers to provide virtual machines, networks, and storage through a common set of APIs. To build one, you need more than an installer. You must plan the hosts, confirm that their physical network supports the IP design, choose where VM data lives, deploy the services, and test what happens when a machine fails.

A single-server installation can teach you how OpenStack works. Production needs a design that accounts for controller availability, guest traffic, storage loss, backups, and staff time. This guide follows those decisions in the order you will face them, then gives you a checklist to use with a dedicated server provider.

What does OpenStack on bare metal actually run?

OpenStack runs on physical hosts, but most of the resources it creates are virtual. A controller hosts cloud APIs and supporting services. A compute host runs a hypervisor, commonly KVM with QEMU, that starts and manages tenant VMs. Users create instances, networks, and other resources through an API, command-line client, or optional dashboard.

“Bare metal” describes the infrastructure below OpenStack. It does not mean every instance is a physical machine. OpenStack has an optional service called Ironic for teams that want to provision physical servers as tenant resources. You do not need Ironic just to install OpenStack on dedicated servers.

This distinction matters when you write a hardware request. If your plan is to sell or assign VMs, ask for hosts that can run the chosen hypervisor. If you also want to provision physical machines through Ironic, add the required management access and provisioning networks to the plan.

What makes it a private cloud?

A private cloud gives a defined organization or group control over its cloud resources and access rules. OpenStack can pool physical hosts and offer users a way to request VMs and networks without asking an administrator to configure each one by hand. Projects, roles, quotas, and images help divide that pool among teams.

Hardware alone does not create that service model. A rack of dedicated servers running separate VMs can meet a workload need without OpenStack. OpenStack earns its place when users need shared capacity, an API, repeatable provisioning, and virtual networks across multiple hosts.

Is OpenStack the right choice for your team?

OpenStack is a good fit when a team needs many VMs across multiple hosts, wants separate projects with controlled access, and has people who can run the cloud platform. A smaller virtualization cluster may be easier to operate when there are only a few fixed workloads. A managed OpenStack offering may fit when the team needs the platform but does not want to maintain its control services.

选项 Fits best when Work the team still owns
One dedicated server with VMs A few workloads can share one host and accept its failure risk Host, hypervisor, VMs, and backups
Small virtualization cluster Several hosts and a simpler VM management layer meet the need Cluster, networking, storage, and recovery
Self-managed OpenStack on bare metal Many users need pooled hosts, APIs, quotas, and self-service networking Full cloud stack, physical infrastructure, and day-to-day operation
Managed OpenStack OpenStack features are needed, but platform operation can be assigned by contract Workloads and the duties the service agreement leaves to the customer

A proof of concept should answer a business question. For example: Can two internal teams create VMs on separate networks, attach persistent volumes, and recover service within the required time after a host fails? If a few manually managed VMs already meet the need, adding OpenStack may create more work than value.

Which OpenStack services do you need first?

Start with the services that let an authorized user launch a VM and connect it to a network. OpenStack’s Installation Guide introduces identity, image, placement, compute, and networking before optional services. Each component has a clear job.

服务 Job What a failure can look like
Keystone Confirms users and services; supports projects, roles, and service endpoints Authentication or API requests fail
Glance Registers and serves VM images New instances cannot retrieve the selected image
Placement Tracks resource inventories and allocations Nova cannot find a suitable host for an instance
Nova Schedules and manages compute instances Instances fail to build or host state is missing
Neutron Provides virtual networks and related connectivity Instances boot but cannot reach the required network
Horizon, optional Gives users a web interface The web interface is down even if APIs still work
Cinder, optional Supplies block volumes A VM cannot attach or access the planned persistent volume

The services also depend on components such as a database, message broker, reliable name resolution, and synchronized clocks. An API that answers a health check does not prove the whole cloud works. A tenant should be able to authenticate, launch an instance, receive an address, reach the network, and use storage where required.

Other services belong to specific needs. Swift offers object storage, Ironic can provision physical machines, and Octavia can provide load balancing for tenant applications. Add each service because the workload needs it, then include it in the backup and upgrade plan.

How many physical servers should you plan for?

The number depends on your failure goal and deployment method. One host can support a lab. OpenStack’s manual Installation Guide says its basic example is a proof of concept, not a production design. For production, a common starting layout uses multiple controllers for the API and supporting services, separate compute hosts for VMs, and a storage setup chosen for the data.

Kolla Ansible’s production guide describes control, network, compute, storage, monitoring, and deployment roles. These are roles, not a rule that every role needs its own server. You can combine some roles in a small cloud, but that increases the effect of a single host failure and puts more load on each machine.

角色 Main work Question before combining roles
Controller APIs, database, and messaging, based on deployment design Can it stay available if one controller fails?
网络 Guest connectivity, routers, API traffic path, and related agents Can clients still reach the cloud when one network host fails?
计算 Hypervisor and tenant VMs Can other hosts take the required workload after a failure?
存储 Image or volume back ends and data replication, if used Which disks and paths are shared failure points?
配置 Runs automation, stores inventory and configuration Can another operator rebuild from a secure copy?

As an illustration, a team might plan three controller nodes, two or more compute nodes, and separately designed storage. That layout is a starting sketch. Its actual size depends on VM memory use, network routes, storage layout, and acceptable downtime. Three controllers in one rack still share rack-level risk.

Size compute for one host loss

Normal capacity is an incomplete sizing target. Suppose two compute hosts each have 96 GiB of memory available for tenant VMs after host reservations. Together they offer 192 GiB. If the VMs need 150 GiB, the cloud fits on a normal day. After one host fails, only 96 GiB remains, so it cannot run all the same VMs at once.

Add a third host with the same assumed VM capacity. After losing one host, the remaining two offer 192 GiB. The arithmetic now supports the 150 GiB VM demand, but you still need to check CPU placement, disk access, network throughput, VM restart procedures, and workload dependencies. This is a planning example, not an Atal Networks hardware specification or a test result.

The formula is simple: capacity after the largest planned host failure must meet the demand you need to keep online, with room for host and platform services. If the business can tolerate some VMs staying offline during a host outage, document which ones and size for that target instead.

Plan the physical network before installing OpenStack

The network is often the part that separates a working lab from a usable hosted private cloud. A server provider may give you a private interface and extra public IPs, yet those features alone do not show that OpenStack can move an API VIP between controllers or deliver a floating IP to a tenant VM.

Plan these paths separately:

路径 目的 确认什么
Host management and control OpenStack service traffic, administration, and supporting services Reachability among nodes, DNS, time sync, and restricted access
Tenant data VM traffic between computes Allowed VLANs or overlay traffic, routes, and MTU
External or provider network Public access and egress for selected tenant VMs How the provider routes addresses and permits network bridging
存储 VM disk, image, or replication traffic Capacity, latency, separation, and host failure behavior
Out-of-band management Recovery when the operating system or host network fails Console access, power control, and who holds the credentials

One or more paths may share a physical interface if the provider and design allow it. Keeping the traffic types distinct in the plan still helps you set firewall rules, measure load, and find failures. Do not change the only management interface for a Neutron bridge without a working out-of-band recovery path.

Provider networks, self-service networks, and floating IPs

一台 provider network connects OpenStack networking to a defined physical network. A self-service network lets a project create an internal virtual network, often using an overlay. Neutron can connect that internal network to outside networks through a router when the chosen network back end and physical paths support it.

一台 floating IP is an address that users can associate with a VM for external access. It needs a route from the upstream network to the OpenStack network path. Its existence in an OpenStack address pool does not make the upstream switch or router deliver packets to it.

Before you choose an IP range or bridge, get answers to these questions from the provider:

  • Do the intended hosts share the layer 2 segment required by the VIP or provider-network design?
  • Which VLAN tags are allowed, and are tagged interfaces or switch port trunks available?
  • Does the public address block route to a single server, a gateway, a VLAN, or a movable endpoint?
  • Can an API VIP and tenant floating IPs move between network hosts under the plan terms?
  • Does the provider offer a routed VIP, external load balancer, or BGP arrangement if a shared layer 2 path is unavailable?
  • Which NIC and port can be used for the guest external network without interrupting host management?
  • Which network MTU works for both physical packets and any tenant overlay?

Ask for a small test range and prove the path with one VM before assigning it to production applications. A private VLAN between servers is useful for host traffic, but it is not automatically a complete Neutron tenant network.

Choose storage for images, VM disks, and volumes

OpenStack storage is not one disk pool. Glance stores VM images. Nova can store an instance’s local, temporary disk on its compute host. Cinder, if enabled, provides block volumes that a VM can attach. These data types may use different back ends, so each needs its own failure and backup answer.

Data type Possible location Host failure question
Glance image data A supported image store or shared back end Can a new instance still read the image?
Nova instance disk Local compute disk or supported shared storage Can the VM’s disk be accessed or recovered elsewhere?
Cinder volume Local, shared, or distributed block back end, depending on setup Does the volume remain available if its host fails?
Database and configuration Control-plane data store plus protected configuration copies Can operators restore cloud state and credentials?
Application data Inside VM disks, attached volumes, or external systems How do application backups restore consistent records?

Local storage can be a sound choice for disposable VMs or applications that manage their own data copies. It is a weaker fit if a VM must restart on another compute host with its data intact and there is no tested way to move or restore the disk.

A separately operated Ceph cluster can provide shared storage for selected OpenStack services. It adds storage hosts, replication traffic, capacity planning, monitoring, and recovery work. In the documented Kolla Ansible external Ceph path, Kolla connects OpenStack services to an existing Ceph cluster. It does 不 create and operate that Ceph cluster for you. Check the chosen release’s storage guidance before selecting a back end.

Shared storage does not mean a failed compute host’s VMs start elsewhere by themselves. Check the supported VM recovery workflow and test it. Live migration also has its own prerequisites and is not a substitute for an unplanned host-failure test. RAID can protect against some disk failures when configured for redundancy, but it does not replace backups of VM and application data.

Choose a deployment method you can maintain

OpenStack is a collection of services, and a production install needs a repeatable way to configure and upgrade them. The right method is one your team can support after the first VM boots.

Route Where it fits Check before choosing
Kolla Ansible Team wants containerized OpenStack services deployed with Ansible Supported OpenStack release, host OS, NIC and VIP design, storage integration, and upgrade process
OpenStack-Ansible Team has Ansible skills and wants this project’s deployment model Supported release, production example, host roles, and upgrade path
Canonical OpenStack Team wants Canonical’s documented product and operating model Hardware and network requirements, support scope, and its installation route
Managed OpenStack Team needs OpenStack features but plans to assign platform operations Exact split of hardware, API, storage, security, upgrades, and restore duties

Kayobe is another route when you want to automate the physical host lifecycle around Kolla Ansible. If you choose it, include management controllers, host provisioning, and access control in the design. Do not add a second deployment framework just because it appears in another company’s reference architecture.

OpenStack 2026.2 is the latest upstream documentation release as of November 2026. A current upstream release does not prove that every deployment tool, OS version, storage driver, and add-on supports the same set of versions. Select a tested combination and record it in a release matrix. Keep copyable commands in a separate tutorial that can be updated when that matrix changes.

How to build a private cloud with OpenStack: seven steps

The build order starts with the workload and the physical network. It ends with a service and recovery test, not a dashboard login.

1. Write the service requirements

List the teams that need access, expected VM profiles, image types, storage needs, public IP needs, expected traffic, and security boundaries. Set a target for how long each service may be down after a host failure and how much data it may lose. Assign owners for the platform and its backups.

A lab can accept one controller and manual repair. Production may need multiple controllers, spare compute capacity, an external backup location, and on-call coverage. Write down which standard applies before you buy hosts.

2. Reserve hosts and prove the underlay

Assign proposed controller, compute, network, and storage roles. Record CPU virtualization support, RAM, usable disks, NICs, addresses, and out-of-band access. Confirm the provider’s network answers in writing. Test host-to-host connectivity, name resolution, time sync, IP routing, and access to the management console.

Keep a record of which physical rack or failure domain each machine occupies if your provider can supply that detail. If all roles share one device or link, make that dependency visible in the architecture.

3. Pin the software and deployment path

Choose one deployment tool and a supported OpenStack release. Check the supported host OS, Python and Ansible dependencies, container runtime where used, hypervisor, network back end, and storage driver. Decide how to keep the inventory, secrets, and overrides in secure version control or a protected backup.

For a Kolla Ansible path, use its release-specific quickstart to learn the tool and its multinode and production guidance for the real layout. Do not copy an all-in-one example into a production environment unchanged.

4. Deploy and check the control services

Bring up the controllers, database, message broker, Keystone, Glance, Placement, and Nova control services according to the chosen tool. Add Neutron in the planned network design. If the setup uses an internal API VIP, test that the address stays reachable when one eligible controller or network host stops.

Confirm authentication and the service catalog with a client, not only with Horizon. Back up the control-plane configuration and establish a tested recovery method for the database and secrets.

5. Add compute and the tenant network

Register compute hosts and confirm they report usable CPU and memory to Placement. Create a test project, flavor, image, security group, and network. Boot a VM. Check its private IP, DNS, and connection to another VM where applicable.

If a VM must be public, test a floating IP from outside the provider network. A VM that boots but cannot accept traffic has passed only part of the build.

6. Add the storage and access controls you need

Choose the Glance image store and the intended VM disk back end. Add Cinder if workloads require attachable block volumes. If you use Ceph, build and test it as a separate system, then connect OpenStack services through the supported integration.

Set project access, roles, quotas, image permissions, security groups, and administrative access. Keep physical management credentials away from tenant networks. Test a volume attachment and a restore from a protected backup.

7. Rehearse failure, restore, and upgrade

Stop one test controller, one compute host, and one network path in planned windows. Record what stays available, what needs a manual action, and the time needed to restore service. Restore a sample VM and its application data in an isolated environment. Rehearse an upgrade on a copy of the deployment plan before a production change.

A first successful VM is a useful milestone. The cloud is ready for production only when its behavior under loss and recovery meets your written service targets.

A production acceptance worksheet

Use this table as a test record. Add the observed result, time, evidence, owner, and next action for each row. Do not mark a test complete because an API reports “healthy” while the tenant workload is down.

测试 Action in a safe test window Pass condition to define before testing
Controller loss Stop one controller Required APIs remain reachable through the planned endpoint
Database or queue member loss Stop one member in the supported topology Control services behave as designed, and operators can identify any reduced capacity
Compute host loss Stop a host running test VMs The documented VM recovery process works within the workload’s target
Network node loss Stop the node carrying the test traffic path API and tenant traffic use the expected alternate path, if one was designed
Floating IP reachability Test from a network outside the provider Traffic reaches the intended VM before and after the planned failover
Storage loss Remove a test storage node or path VM disks, images, and volumes behave as the storage design states
恢复 Restore cloud data and one application from protected backups Operator can verify the recovered resources and application records
Upgrade rehearsal Run the supported procedure in a test environment Services and tenant connectivity pass the same checks after the change

A recovery time objective, or RTO, is the downtime the service aims to meet. A recovery point objective, or RPO, is the data loss window it aims to meet. Only a measured restore shows whether the design meets either goal. A provider’s server or network uptime commitment does not set guest VM or application availability by itself.

What does an OpenStack private cloud cost?

Compare full operating cost, not just the price of the compute hosts. A practical monthly model has these lines:

Controller and network nodes + compute nodes + storage and replication + public IPs and network transfer + backup capacity + software support + staff and on-call time + a share of setup and replacement costs.

Keep the assumptions next to each line. State how many hosts are dedicated to control services, how much compute headroom remains after a failure, and whether storage traffic uses a separate port or billed transfer. If the same team operates the OS, OpenStack, and Ceph, include that work. If a provider manages any of them, read the service agreement for the exact boundary.

OpenStack being open source does not remove the cost of hardware, upgrades, backup storage, or incident response. A fair comparison with a managed cloud also accounts for the work the managed provider takes over. Use current quotes rather than a general claim that private cloud always costs less.

What should you ask a bare metal provider?

Ask these questions before making the final host order. They affect whether your OpenStack design can work on that provider’s network:

  • Which private network paths can connect the proposed servers? Can the provider place them across separate physical failure domains?
  • Are the required VLANs, tagged ports, bridges, or routes supported on the chosen plan?
  • How is a public IP range routed? Can the planned API VIP or tenant floating IPs move between hosts?
  • Is there an external load balancer or approved BGP option if the network cannot support the selected VIP design?
  • How many NICs are available, at what port speeds, and what traffic is included or billed separately?
  • Can we use the host OS, virtualization settings, and disk layout required by our deployment tool?
  • What console, KVM, IPMI, or other out-of-band access is included, and who can use it?
  • How are a failed disk and host replaced? What happens to local data during replacement?
  • Which layers will the provider support, and which will our team operate?

Atal Networks 提供 裸机服务器 和 专用服务器配置 that can provide the physical hosts for a private cloud. Teams bringing their own machines can also review 托管. We should confirm VLANs, routed IPs, floating IP behavior, out-of-band access, and support terms for the exact plan and location before recommending an OpenStack layout. A dedicated server order alone does not mean OpenStack or Ceph is managed for you.

经常问的问题

Can I build an OpenStack private cloud on rented dedicated servers?

Yes, if the servers and network support your chosen deployment design. You need control, compute, external IP, and storage paths, plus a way to manage failed hosts. Confirm VLAN, routing, VIP, NIC, and out-of-band access with the provider before selecting configuration examples.

Do I need Ironic to install OpenStack on bare metal?

No. You can run OpenStack controllers and compute hypervisors on physical hosts without Ironic. Add Ironic when the cloud must provision physical machines as tenant resources. That adds its own management and provisioning network requirements.

How many servers does a first OpenStack cloud need?

A single host can support an evaluation build. Production host count depends on which service failures you must survive, how much compute capacity you need after losing a node, and how storage works. A three-controller design can address one class of control-plane failure, but application recovery needs its own plan.

Which services do I need to launch a VM?

The core flow uses Keystone for identity, Glance for images, Placement for resource inventory, Nova for compute, and Neutron for networking. The setup also relies on a database, message broker, and working host network. Horizon is useful for web access, but it is not required to prove the APIs work.

Does Kolla Ansible install and manage Ceph?

No. In the documented external Ceph setup, the team operates an existing Ceph cluster and connects supported OpenStack services to it through Kolla Ansible. You need a separate plan for Ceph installation, capacity, replication, monitoring, upgrades, and backups.

Can I use my provider’s public IP block for floating IPs?

Possibly. The address block must be delivered in a way that works with your Neutron external network and upstream routing. If the provider routes every address to one host, placing the addresses into a floating IP pool will not make them move to another host. Ask the provider how the range is delivered and test it.

What happens when an OpenStack compute server fails?

The VMs on that physical host may stop. OpenStack control services and shared storage alone do not guarantee that the workloads restart on other hosts. Recovery depends on spare compute capacity, disk access, the chosen high availability or operator process, and tests of the actual applications.

Is an all-in-one OpenStack install suitable for production?

Treat it as a learning or proof-of-concept path unless your service explicitly accepts one host as a single failure point. OpenStack’s manual Installation Guide identifies its example as a proof of concept. Use production guidance from your selected deployment method when designing a cloud that must survive host failures.

How is OpenStack different from a small virtualization cluster?

Both can run VMs on physical hosts. OpenStack adds a broad set of APIs and services for pooled capacity, projects, virtual networking, quotas, and user self-service. Those features bring more components to run. If your need is a small set of fixed VMs, compare the operational work before choosing OpenStack.

Can Kubernetes run on VMs created by OpenStack?

Yes. OpenStack can supply virtual machines and networks that host Kubernetes nodes. Kubernetes then manages containers on those VMs. The two systems solve different problems, so Kubernetes is not required to build the first OpenStack cloud.

推荐的下一步

Write the physical network and storage plan before you install OpenStack. Confirm the provider’s answers on VLANs, routed IPs, VIP behavior, disk replacement, and management access. Then use a supported deployment method to build a small test cloud and complete the acceptance worksheet with observed results.

If the design calls for dedicated hardware, send us the node roles, preferred location, VM count, CPU and RAM targets, NIC and VLAN plan, public IP needs, storage choice, and recovery goals. We can help you compare 裸机服务器选项 for the physical part of the cloud.

滚动至顶部