The evolution of the Smart NICs to DPUs

The ideas behind “Smart NICs” have evolved as time has progressed.

And honestly it feels like we in the Ironic community helped drive some of that improvement in better, more secure directions. Hey! We said we we’re changing the world!

What started out as highly advanced network cards which infrastructure operators desired to offload some traffic, have morphed into a more generic workload, yet still in oriented in the direction of offload. Except now these newer generations of cards have their own BMC modules, and a limited subset of hardware management can occur.

But the access model and use/interaction also means that a server can have a primary BMC, and then N number of subset BMCs, some of which may, or may not be able to communicate with the overall BMC and operating system.

And in order to really support this, and the varying workflows, we need to consider some major changes to the overall model of interaction and support. This is not because the device is just a subset, but a generalized computer inside of a computer with it’s own unique needs for management protocols, boot capabilities/devices, architecture, and has it’s own console, internal state, credentials, et cetra.

Problem description

Context Required

In order to navigate this topic, we need to ensure we have context of various terms in use and as they relate.

Smart NIC

These are best viewed as a “first generation” of DPU cards where an offload workload is able to be executed on a card, such as a Neutron agent connected to the message bus in order to bind ports to the physical machine.

Some initial community and vendor discussions also centered around futher potential use cases of providing storage attachments through, similar to the behavior of a Fibre Channel HBA.

Composible Hardware

The phrase “Composible Hardware” is unfortunately overloaded. This is best described as use of a centralized service to “compose” hardware for use by a workload. A good way to view this, at least in a classical sense is through an API or application constructing a cohesive functioning computer resource with user definable CPU, Memory, Storage, and Networking. Essentially to virtualize the hardware interaction/modeling like we have with Virtual Machines.

Aside from some limited hardware offerings from specific vendors, Composible Hardware largely hasn’t been realized as initally pitched by the hardware manufacters.


A DPU or Data Processing Unit is best viewed as a more generalized, “second generation” of the Smart NIC which is designed to run more generalized workloads, however this is not exclusively a network or adapter to network attached resources. For example, one may want to operate a BGP daemon inside of the card, which is entirely out of scope ironic to operate with and manage, but they could run the service there in order to offload the need to run it on the main CPUs of the system. A popular, further idea, is to utilize the card as a “root of trust”

A similarity between DPU’s and Composible hardware in modeling is the concept of providing potentially remote resources to the running operating system.

Given the overall general purpose capabilities of DPUs and increased focus of specific computing workload offloads, we need to be careful to specifically delineate which use cases we’re attempting to support, and also not try to assume one implies the other. In other words, DPUs do offer some interesting capabilities towards Composible Hardware, however it is inherently not full configuration as the underlying host is still a static entity.


DPUs are also sometimes expressed as xPUs, because classical graphics cards are Graphics Processing Units. While there does not appear to be any explicit movement into supporting that specific offload, some vendors are working on highly specific processing cards such those as performing protocol/signal translation. They may, or may not be able to have an operating system or provisioned application.

The problem

Today, In Ironic, we have a concept of a baremetal node. Inside of that baremetal node, there may be various hardware which can be centrally managed, and interacted with. It has a single BMC which controls basically all aspects.

We also have a concept of a “smart nic” in the form of is_smartnic on port objects. However this only impacts Ironic’s power and boot mode management.

Except, not all of these cards are “network” cards, or at least “network” cards in any traditional computing model. Think Radios!

And the intertwined challenge is this nested access model.

For the purpose of this example, I’m going to refer to Nvidia Bluefield2 cards with a BMC. It should be noted we have support in Antelope IPA to update the BMC firmware of these cards. At least, that is our understanding of the feature.

But to do so:

  1. From the host, the access restrictions need to be dropped by requesting the BMC on the card to permit the overall host OS to Access the card’s BMC. This is achievable with an IPMI raw command, but against the card’s BMC.

  2. Then you would apply BMC firmware updates, to the card’s BMC Today this would boot IPA, and perform it from the host OS, which also means that we’re going to need to interact with the overall host BMC, and boot the baremetal machine overall.

  3. Then you would naturally want to drop these rights, which requires calling out to the card’s BMC to change the access permissions.

  4. Then if you wanted to update the OS on the card itself, you would rinse and repeat the process, with a different set of commands to open the access between the OS on the card, and the OS on the BMC.


We know the Bluefield2 cards can both be network booted, and updated by SSH’ing into the BMC and streaming the new OS image to the installer command over SSH. That, itself, would be a separate RFE or feature but overall modeling changes would still be needed which this specification seeks to resolve.

Hopefully this illustrates the complexity, begins to underscore the need as to why we need to begin to support a parent/child device model, and permit the articulation of steps upon a parent node which applies to the one or more children nodes.

What complicates this further is is ultimately we’re just dealing with many different Linux systems, which have different models of access. For example, A more recent Dell Server running IPA, with two of these specalized cards, henceforth referred to as Data Processing Units (DPUs), would have Linux running on the host BMC, On the host processors, inside the BMCs of each DPU card, and inside of the processor on the DPU card.

This specification does inherently exclude configuration of the operating state configuration of the DPU card. There are other projects which are working on that, and we look forward to integrating with them as they evolve.


The other project in mind is the OPI project, which is working on quite a lot of capabilities in this space, however they explicitly call out automated/manual deployment via outsize of zero touch provisioning is out of scope for their project. Such is sensible to stand-up a general purpose workload, but operating lifecycle and on-going management is an aspect where Ironic can help both operators who run a variety of workloads and configurations, or need to perform more specific lifecyle operations.

Proposed change

The overall idea with this specification is to introduce the building blocks to enable the orchusrtration and articulation of actions between parent and child devices.

  • Introduction of parent_node field on the node object with an API version increase.

  • Introduction of a sub-node resource view of /v1/nodes/<node>/children which allows the enumeration of child nodes.

  • Default the /v1/nodes list to only list nodes without a parent, and add a query filter to return nodes with parents as well.

  • Introduction of a new step field value, execute_on_child_nodes which can be submitted. The default value is False. Steps which return CLEANWAIT, i.e. steps which expect asynchronous return will not be permitted under normal conditions, however this will be avaiable via a configuration option.

  • Introduction of a new step field value, limit_child_node_execution, which accepts a list of node UUIDs to allow filtering and constraint of steps on some nodes. Sepcifically, this is largely separate from the execute_on_child_nodes field due to JSON Schema restrictions.

  • Introduction of the ability to call a vendor passthrough interface as a step. In the case of some smartnics, they need the ability to call IPMI raw commands across child nodes.

  • Introduction of the ability to call set_boot_device as a step. In this case, we may want to set the DPU cards to PXE boot en-mass to allow for software deployment in an IPA ramdisk, or other mechanism.

  • Introduction of the ability to call power_on, power_off management interface methods through the conductor set_power_state helpers (which includes guarding logic for aspects like fast track).

  • Possibly: Consider “physical” network interfaces optional for some classes of nodes. We won’t know this until we are into the process of implementation of the capabilities.

  • Possibly: Consider the machine UUID reported by the BMC as an identifier to match for agent operations. This has long been passively desired inside of the Ironic community as a “nice to have”.

  • Possibly: We may need to continue to represent a parent before child or child before parent power management modeling like we did with the Port object is_smartnic field. This is relatively minor, and like other possible changes, we won’t have a good idea of this until we are further along or some community partners are able to provide specific feedback based upon their experiences.

With these high level and workflow changes, it will be much easier for an operator to orchestrate management actions across an single node to extend further into distinct devices within the whole of the system.

In this model, the same basic rules for child nodes would apply, they may have their own power supplies and their own power control, and thus have inherent “on” and “off” states, so deletion of a parent should cause all child nodes to be deleted. For the purpose of state tracking, the individual cards if managed with a specific OS via Ironic, may be moved into a deployed state, however they may just also forever be in a manageable state independent of the parent node. This is because of the overall embedded nature, and it being less of less of a general purpose compute resource compute resource while still being a general computing device. This also sort of reflects the inherent model of it being more like “firmware” management to update these devices.

Outstanding Questions

  • Do we deprecate the port object field is_smartnic? This is documented as a field to be used in the wild for changing the power/boot configuration flow on first generation smartnics which is still applicable on newer generations of cards should the operator have something like Neutron OVS agent connected on the message bus to allow termination of VXLAN connections to the underlying hardware within the card.

Out of Scope, for this Specification

Ideally, we do eventually want to have DPU specific hardware types, but the idea of this specification is to build the substrate needed to build upon to enable DPU specific hardware types and enable advanced infrastructure operators to do the needful.


Three alternatives exist. Technically four.

Do nothing

The first option is to do nothing, and force administrators to manage their nested hardware in a piecemeal fashion. This will create a barrier to Ironic usage, and we already know from some hardware vendors who are utilizing these cards along side Ironic, that the existing friction is a problem point in relation to just power management. Which really means, this is not a viable option for Ironic’s use in more complex environments.

Limit scope and articulate specific workflows

A second option is to potentially limit the “scope of support” to just power or boot operations. However, we have had similar discussions, in relation to supporting xPU’s in servers with external power supplies in the past, and have largely been unable to navigate a workable model, in large part because this model would generally require a single task.node to be able to execute with levels of interfaces with specific parameters. For example, to the system BMC for base line power management, and then to a SNMP PDU for the auxillary power. This model also doesn’t necessarilly work because then we would inherently have blocked ourselves from more general managmeent capabilities and access to on DPU card features such as “serial consoles” through it’s own embedded BMC without substantial refactoring and re-doing the data model.

There is also the possibility that nesting access controls/modeling may not be appropriate. You don’t necessarilly want to offer an baremetal tenant in a BMaaS who has lessee access to Ironic, the ability to get to a serial console which kind of points us to the proposed solution in order to provide capabilities to handle the inherently complex nature of modeling which can result. Or eat least provide organic capabilities based upon existing code.

Use Chassis

The third possibility is to use the existing Chassis resource. The idea of a parent/child relationship does sound similar to the modeling of Chasssis and a Node.

Chassis was originally intended to allow the articulation of entire Racks or Blade Chassis in Ironic’s data model, in part to allow relationship and resoruce tracking more in lines with a Configuration Management Data Base (CMDB) or Asset Inventory. However, Chassis never gained much traction because those systems are often required and decoupled in enterprise environments.

Chassis has been proposed to be removed several times in Ironic, and does allow the creation of a one to many relationship which cannot presently be updated after it is set. Which inherently is problematic and creates a maintenanance burden should a card need to be moved or a chassis replaced but the DPU is just moved to the new chassis.

But the inherent one to many modeling which can exist with DPUs ultimately means that the modeling is in reverse from what is implemented for usage. Nodes would need to be Chassises, but then how do users schedule/deploy “instances”, much less perform targetted lifecycle operations against part of the machine which is independent, and can be moved to another chassis.

Overall, this could result in an area where we may make less progress because we would essentially need to re-model the entire API, which might be an interesting challenge, but that ultimately means the work required is substantially larger, and we would potentially be attempting to remodel interactions and change the user experience, which means the new model would also be harder to adopt with inherently more risk if we do not attempt to carry the entire feature set to a DPU as well. If we look at solving the overall problem from a “reuse” standpoint, the proposed of this specification document solution seems like a lighter weight solution which also leaves the door open to leverage the existing capabilities and provide a solid foundation for future capabilities.

Realistically, The ideal use case for chassiss is fully composible hardware where some sort of periodic works to pre-populate “available” nodes to be scheduled upon by services like Nova from a pool of physical resources, as well as works to reconcile overall differences. The blocker though to that is ultimately availability of the hardware and related documentation to make a realistic Chassis driver happen in Open Source.

Create a new interface or hardware type

We could create a new interface on a node, or a new hardware type.

We do eventually want some DPU specific items to better facilitate and enable end operators, however there is an underling issue of multiple devices, a one to many relationship. Further complicated by a single machine may have a number of different types of cards or devices, which kind of points us back to the original idea proposed.

Data model impact

A parent_node field will be added, and the field will be indexed. A possibility exists that the DB index added may be a multi-field compound index as well, but that is considered an implementation detail.

State Machine Impact

No State Machine impact is expected.

REST API impact

GET /v1/nodes/?include_children=True

Returns a list of base nodes with all child nodes child nodes, useful for a big picture view of all things Ironic is responsible for.

GET /v1/nodes/

The view will by default return only nodes where the parent_node field is null. Older API clients will still recieve this default behavior change.

GET /v1/nodes/<node_ident>/children

Will return the list of nodes, with the pre-existing access list constraints and modeling of all defined nodes where parent_node matches node_ident. In alignment with existing node list behavior, if access rights do not allow the nodes to be viewed, or there are no nodes, an empty list will be returned to the API client.

Additional parameters may also be appropriate with this field, but at present they are best left to be implementation details leaning towards the need to not support additional parameters.


We would likely need to validate the submitted node_ident is also a UUID, otherwise resolve the name to a node, and then lookup the UUID.

A links field will refer to each node, back to the underlying node which may require some minor tuning of the logic behind node listing and link generation.

All of the noted changes should be expected to be merged together with a microversion increase. The only non-version controlled change, being the presence/match of the parent_node field.

Corresponding API client changes will be needed to interact with this area of the code.

Client (CLI) impact

“openstack baremetal” CLI

The baremetal command line interface will need to recieve parameters to query child nodes, and query the child nodes of a specific node.


An SDK change may not be needed, or may be better suited to occur organically as someone identifies a case where they need cross-service support.

RPC API impact

No additional RPC API calls are anticipated.

Driver API impact

No direct driver API changes are anticipated as part of this aside from ensureing the management interface set_boot_device as well as the IPMI interface send_raw commands can be called via the steps framework.

Nova driver impact

None are anticipated, this is intended to be invisible to Nova.

Ramdisk impact

The execution of our ramdisk inside of a DPU is presently considered out of scope at this time.

Some of the existing smartnics might not be advisable to have operations like “cleaning” as well, for example Bluefield2 cards with more traditional SPI flash as opposed to NVMe in Bluefield3 cards. Given some of the speciality in methods of interacting with such hardware, we anticipate we may eventually want to offer specific deployment or boot interfaces which may bypass some of the inherent “agent” capabilities.

Security impact

No additional security impact is anticipated as part of this change.

Other end user impact


Scalability impact

This change does propose an overall relationship and ability which may result far more nodes to be managed in ironic’s database. It may also be that for child devices, a power synchronization loop may not be needed, or can be far less frequent. These are ultimately items we need to discuss furhter, and consider some additional controls if we determine the need so operators may not feel any need nor impact to their deployments due to the increase in rows int the “nodes” table.


It should be noted that the way the hash ring works in Ironic, is that the ring consists of the conductors, which are then mapped to based upon node properties. It may be that a child node’s mapping should be the parent node. These are questions to be determined.

Performance Impact

No direct negative impact is anticipated. The most direct impact will be the database and some periodics which we have already covered in the preceeding section. Some overall performance may be avoided by also updating some of the periodics to not possibly match any child node, the logical case is going to be things like RAID periodics, which would just never apply and should be never configured for such a device, which may itself make the need to make such a periodic change moot.

Other deployer impact

No negative impact is anticipated, but it might be that operators may rapidly identify need for a “BMC SSH Command” interface, as the increasing trend of BMCs being linux powered offers increased capabilities and possibilities, along with potential needs if logical mappings do not map out.

Developer impact




Primary assignee:

Julia (TheJulia) Kreger <>

Other contributors:

<IRC handle, email address, None>

Work Items

  • Addition of parent_node db field and node object.

  • Addition of node query functionality.

  • Introduction of the /v1/nodes/<node>/children API resource and the resulting API microversion increase.

  • Add step support to iterate through step definitions which has mixed step commands for parent nodes and child node.

  • Introduction of generalized power interface steps: * power_on * power_off

  • Add an IPMI management interface raw command step method.

  • Examples added for new step commands and invocation of child node objects.




Basic tempest API contract testing is expected, however a full tempest scenario test is not expected.

Upgrades and Backwards Compatibility

No negative impact is anticipated.

Documentation Impact

Documentation and examples are expected as part of the work items.