Common Considerations

This section covers considerations that are equally important to all described architectures.


As explained in Bare Metal service overview, the Bare Metal service has three components.

  • The Bare Metal API service (ironic-api) should be deployed in a similar way as the control plane API services. The exact location will depend on the architecture used.

  • The Bare Metal conductor service (ironic-conductor) is where most of the provisioning logic lives. The following considerations are the most important when deciding on the way to deploy it:

    • The conductor manages a certain proportion of nodes, distributed to it via a hash ring. This includes constantly polling these nodes for their current power state and hardware sensor data (if enabled and supported by hardware, see Collecting sensor data for an example).

    • The conductor needs access to the management controller of each node it manages.

    • The conductor co-exists with TFTP (for PXE) and/or HTTP (for HTTPBoot and iPXE) services that provide the kernel and ramdisk to boot the nodes. The conductor manages them by writing files to their root directories.

    • If serial console is used, the conductor launches console processes locally. If the nova-serialproxy service (part of the Compute service) is used, it has to be able to reach the conductors. Otherwise, they have to be directly accessible by the users.

    • There must be mutual connectivity between the conductor and the nodes being deployed or cleaned. See Networking for details.

  • The provisioning ramdisk which runs the ironic-python-agent service on start up.


    The ironic-python-agent service is not intended to be used or executed anywhere other than a provisioning/cleaning/rescue ramdisk.

Hardware and drivers

The Bare Metal service strives to provide the best support possible for a variety of hardware. However, not all hardware is supported equally well. It depends on both the capabilities of hardware itself and the available drivers. This section covers various considerations related to the hardware interfaces. See Enabling drivers and hardware types for a detailed introduction into hardware types and interfaces before proceeding.

Power and management interfaces

The minimum set of capabilities that the hardware has to provide and the driver has to support is as follows:

  1. getting and setting the power state of the machine

  2. getting and setting the current boot device

  3. booting an image provided by the Bare Metal service (in the simplest case, support booting using PXE and/or iPXE)


Strictly speaking, it is possible to make the Bare Metal service provision nodes without some of these capabilities via some manual steps. It is not the recommended way of deployment, and thus it is not covered in this guide.

Once you make sure that the hardware supports these capabilities, you need to find a suitable driver. Most of enterprise-grade hardware has support for IPMI and thus can utilize IPMI driver. Some newer hardware also supports Redfish driver. Several vendors provide more specific drivers that usually provide additional capabilities. Check Drivers, Hardware Types and Hardware Interfaces to find the most suitable one.

Boot interface

The boot interface of a node manages booting of both the deploy ramdisk and the user instances on the bare metal node. The deploy interface orchestrates the deployment and defines how the image gets transferred to the target disk.

The main alternatives are to use PXE/iPXE or virtual media - see Boot interfaces for a detailed explanation. If a virtual media implementation is available for the hardware, it is recommended using it for better scalability and security. Otherwise, it is recommended to use iPXE, when it is supported by target hardware.

Hardware specifications

The Bare Metal services does not impose too many restrictions on the characteristics of hardware itself. However, keep in mind that

  • By default, the Bare Metal service will pick the smallest hard drive that is larger than 4 GiB for deployment. Another hard drive can be used, but it requires setting root device hints.


    This device does not have to match the boot device set in BIOS (or similar firmware).

  • The machines should have enough RAM to fit the deployment/cleaning ramdisk to run. The minimum varies greatly depending on the way the ramdisk was built. For example, tinyipa, the TinyCoreLinux-based ramdisk used in the CI, only needs 400 MiB of RAM, while ramdisks built by diskimage-builder may require 3 GiB or more.

Image types

The Bare Metal service can deploy two types of images:

  • Whole-disk images that contain a complete partitioning table with all necessary partitions and a bootloader. Such images are the most universal, but may be harder to build.

  • Partition images that contain only the root partition. The Bare Metal service will create the necessary partitions and install a boot loader, if needed.


    Partition images are only supported with GNU/Linux operating systems.

    For the Bare Metal service to set up the bootloader during deploy, your partition images must container either GRUB2 bootloader or ready-to-use EFI artifacts.


There are several recommended network topologies to be used with the Bare Metal service. They are explained in depth in specific architecture documentation. However, several considerations are common for all of them:

  • There has to be a provisioning network, which is used by nodes during the deployment process. If allowed by the architecture, this network should not be accessible by end users, and should not have access to the internet.

  • There has to be a cleaning network, which is used by nodes during the cleaning process.

  • There should be a rescuing network, which is used by nodes during the rescue process. It can be skipped if the rescue process is not supported.


In the majority of cases, the same network should be used for cleaning, provisioning and rescue for simplicity.

Unless noted otherwise, everything in these sections apply to all three networks.

  • The baremetal nodes must have access to the Bare Metal API while connected to the provisioning/cleaning/rescuing network.


    Only two endpoints need to be exposed there:

    GET /v1/lookup
    POST /v1/heartbeat/[a-z0-9\-]+

    You may want to limit access from this network to only these endpoints, and make these endpoint not accessible from other networks.

  • If the pxe boot interface (or any boot interface based on it) is used, then the baremetal nodes should have untagged (access mode) connectivity to the provisioning/cleaning/rescuing networks. It allows PXE firmware, which does not support VLANs, to communicate with the services required for provisioning.


    It depends on the network interface whether the Bare Metal service will handle it automatically. Check the networking documentation for the specific architecture.

    Sometimes it may be necessary to disable the spanning tree protocol delay on the switch - see DHCP during PXE or iPXE is inconsistent or unreliable.

  • The Baremetal nodes need to have access to any services required for provisioning/cleaning/rescue, while connected to the provisioning/cleaning/rescuing network. This may include:

    • a TFTP server for PXE boot and also an HTTP server when iPXE is enabled

    • either an HTTP server or the Object Storage service in case of the direct deploy interface and some virtual media boot interfaces

  • The Baremetal Conductors need to have access to the booted baremetal nodes during provisioning/cleaning/rescue. A conductor communicates with an internal API, provided by ironic-python-agent, to conduct actions on nodes.

HA and Scalability


The Bare Metal API service is stateless, and thus can be easily scaled horizontally. It is recommended to deploy it as a WSGI application behind e.g. Apache or another WSGI container.


This service accesses the ironic database for reading entities (e.g. in response to GET /v1/nodes request) and in rare cases for writing.


High availability

The Bare Metal conductor service utilizes the active/active HA model. Every conductor manages a certain subset of nodes. The nodes are organized in a hash ring that tries to keep the load spread more or less uniformly across the conductors. When a conductor is considered offline, its nodes are taken over by other conductors. As a result of this, you need at least 2 conductor hosts for an HA deployment.


Conductors can be resource intensive, so it is recommended (but not required) to keep all conductors separate from other services in the cloud. The minimum required number of conductors in a deployment depends on several factors:

  • the performance of the hardware where the conductors will be running,

  • the speed and reliability of the management controller of the bare metal nodes (for example, handling slower controllers may require having less nodes per conductor),

  • the frequency, at which the management controllers are polled by the Bare Metal service (see the sync_power_state_interval option),

  • the bare metal driver used for nodes (see Hardware and drivers above),

  • the network performance,

  • the maximum number of bare metal nodes that are provisioned simultaneously (see the max_concurrent_builds option for the Compute service).

We recommend a target of 100 bare metal nodes per conductor for maximum reliability and performance. There is some tolerance for a larger number per conductor. However, it was reported [1] [2] that reliability degrades when handling approximately 300 bare metal nodes per conductor.

Disk space

Each conductor needs enough free disk space to cache images it uses. Depending on the combination of the deploy interface and the boot option, the space requirements are different:

  • The deployment kernel and ramdisk are always cached during the deployment.

  • When [agent]image_download_source is set to http and Glance is used, the conductor will download instances images locally to serve them from its HTTP server. Use swift to publish images using temporary URLs and convert them on the node’s side.

    When [agent]image_download_source is set to local, it will happen even for HTTP(s) URLs. For standalone case use http to avoid unnecessary caching of images.

    In both cases a cached image is converted to raw if force_raw_images is True (the default).

    See Deploy with custom HTTP servers and Streaming raw images for more details.

  • When network boot is used, the instance image kernel and ramdisk are cached locally while the instance is active.


All images may be stored for some time after they are no longer needed. This is done to speed up simultaneous deployments of many similar images. The caching can be configured via the image_cache_size and image_cache_ttl configuration options in the pxe group.

Other services

When integrating with other OpenStack services, more considerations may need to be applied. This is covered in other parts of this guide.