Kuryr Kubernetes SR-IOV Integration

Kuryr Kubernetes SR-IOV Integration


This spec proposes an approach to allow kuryr-kubernetes manage pods that require SR-IOV interfaces.

Problem Description

SR-IOV (Single-root input/output virtualization) is a technique that allows a single physical PCIe device to be shared across several clients (VMs or otherwise). Each such network card would have a single PF (physical function) and multiple VFs (Virtual Functions), essentially appearing as multiple PCIe devices. These VFs can be then passed-through to VMs bypassing the hypervisor and virtual switch. This allows performance comparable to non-virtualized environments. SR-IOV support is present in nova and neutron, see docs [1].

It is possible to implement a similar approach within Kubernetes. Since Kubernetes uses separate network namespaces for Pods, it is possible to implement pass-through, simply by assigning a VF device to the desired Pod’s namespace.

There are several challenges that this task poses:

  • SR-IOV interfaces are limited and not every Pod would require them. This means that a Pod should be able to request 0(zero) or more VFs. Since not all Pods will require VFs, these interfaces should be optional.
  • For SR-IOV support to be practical the Pods should be able to request multiple VFs, possibly from multiple PFs. It’s important to note that Kubernetes only stores information about a single IP address per Pod, however it does not restrict configuring additional network interfaces and/or IP addresses for it.
  • Different PFs may map to different neutron physical networks(physnets). Pods need to be able to request VFs specific physnet and physnet information (vlan id, specifically) should be passed to the CNI for configuration.
  • Kubernetes does not have any knowledge about SR-IOV interfaces on the Node it runs. This can be mitigated by utilising Opaque Integer Resources [2] feature from 1.5.x and later series.
  • This feature would be limited to bare metal installations of Kubernetes, since it’s currently impossible to manage VFs of a PF inside a VM. (There is work to allow this in newer kernels, but latest stable kernels do not support it yet)

Proposed Change

Proposed solution consists of two major parts: add SR-IOV capabilities to VIF handler of kuryr-kubernetes controller and enhance CNI to allow it associate VFs to Pods.

Pod scheduling and resource management

Since Kubernetes is the one who actually schedules the Pods on a Node we need a way to tell it that a particular node is capable of handling a SR-IOV-enabled Pods. There are several techniques in Kubernetes, that allow limiting where a pod should be scheduled (i.e. Labels and NodeSelectors, Taints and Tolerations), but only Opaque Integer Resources [2] (OIR) allows exact bookkeeping of VFs. This spec proposes to use a predefined OIR pattern to track VFs on a node::


For example to request VFs for physnet2 it would be::


It will be deployer’s duty to set these resources, during node setup. kubectl does not support setting ORI as of yet, so it has to be done as a PATCH request to Kubernetes API. For example to add 7 VFs from physnet2 to k8s-node-1 one would issue the following request:

curl --header "Content-Type: application/json-patch+json" \
--request PATCH \
--data '[{"op": "add", "path":
"value": "7"}]' \

For more information please refer to OIR docs. [2] This process may be automated, using Node Feature Discovery [3] or a similar service, however these details are out of the scope of this spec.

Here’s how A Pod Spec might look like this:

  - name: vf-container
    image: vf-image
        pod.alpha.kubernetes.io/opaque-int-resource-sriov-vf-physnet2: 1
  - name: vf-other-container
    image: vf-other-image
        pod.alpha.kubernetes.io/opaque-int-resource-sriov-vf-physnet2: 1
        pod.alpha.kubernetes.io/opaque-int-resource-sriov-vf-physnet3: 1

These requests are per-container, and the total amount of VFs should be totalled for the Pod, the same way Kubernetes does it. The example above would require 2 VFs from physnet2 and 1 from physnet3.

An important note should be made about kubernetes Init Containers [4]. If we decide that it is important to support requests from Init Containers, they would have to be treated differently. Init Containers are designed to run sequentially, so we would need to scan them and get maximum request value across all of them.

Requesting SR-IOV ports

To implement SR-IOV capabilities current VIF handler will be modified to handle multiple VIFs. As a prerequisite of this the following changes have to be implemented:

Multi-VIF capabilities of generic handler

Instead of storing a single VIF in the annotation VIFHandler would store a dict, that maps desired interface name to a VIF object. As an alternative we can store VIFs in a list, but dict would give finer control over interface naming. Both handler and the CNI would have to be modified to understand this new format of the annotation. The CNI may also be kept backward-compatible, i.e. understand the old single-VIF format.

Even though this functionality is not a part of SR-IOV handling it acts as a prerequisite and would be implemented as part of this spec.

SR-IOV capabilities of generic handler

The handler would read OIR requests of a scheduled Pod and would see if the Pod has requested any SR-IOV VFs. (NOTE: at this point the Pod should already be scheduled to a node, meaning there are enough available VFs on that node). The handler would ask SR-IOV driver for sufficient number of direct ports from neutron and pass them on to the CNI via annotations. Network information should also include network’s VLAN info, to setup VF VLAN.

SR-IOV functionality requires additional knowledge of neutron subnets. The controller needs to know a subnet where it would allocate direct ports for certain physnet. This can be solved by adding a config setting that will map physnets to a default neutron subnet It might look like this:

default_physnet_subnets =  "physnet2:e603a1cc-57e5-40fe-9af1-9fbb30905b10,physnet3:0919e15a-b619-440c-a07e-bb5a28c11a75"

Alternatively we can request this information from neutron. However since there can be multiple networks within a single physnet and multiple subnets within a single network there is a lot of space for ambiguity. Finally we can combine the approaches: request info from neutron only if it’s not set in the config.


On the CNI side we will implement a CNI binding driver for SR-IOV ports. Since this work will be based on top of multi-vif support for both CNI and controller, no additional format changes would be implemented. The driver would configure the VF and pass it to the Pod’s namespace. It would scan /sys/class/net/<PF>/device directory for available virtual functions and pass the acquired device to Pods namespace.

The driver would need to know which devices map to which physnets. Therefore we would introduce a config setting physical_device_mappings. It will be identical to neutron-sriov-nic-agent’s setting. It might look like:

physical_device_mappings = "physnet2:enp1s0f0,physnet3:enp1s0f1"

As an alternative to storing this setting in kuryr.conf we may store it in /etc/cni/net.d/kuryr.conf file or in a kubernetes node annotation.


  • Current implementation does not concern itself with setting active status of the Port on the neutron side. It is not required for the feature to function properly, but may be undesired from operators standpoint. Doing so may require some additional integration with neutron-sriov-nic-agent and verification. There is a concern, that neutron-sriov-nic-agent does not detect port status correctly all the times.

Optional 2-Phase Approach

Initial implementation followed an alternative path, where SR-IOV functionality has been implemented as a separate handler/cni. This sparked several design discussions, where community agreed, that multi-VIF handler is preferred over multi-handler approach. However if implementing multi-vif handler would prove to be lengthy and difficult we may go with a 2-phase approach. First phase: polish and merge initial implementation. Second phase: Implement multi-vif approach and convert sriov-handler to use it.


  • It is possible to implement SR-IOV functionality as a separate handler. In this scenario both handlers would listen to Pod events and would handle them separately. They would have to use different annotation keys inside the Pod object. The CNI would have to be able to handle both annotation keys.
  • Since this feature is only practical for bare metal we can implement it entirely on the CNI side. (i.e. CNI would request ports from neutron). However this would introduce an alternative control flow.
  • It is also possible to implement a separate CNI, that would use static configuration, compatible with neutrons, much like [5]. This would eliminate the need to talk to neutron at all, but would put the burden of configuring multiple nodes network information on the deployer. This may be however desirable for some installations and may be considered as an option. At the same time in this scenario there would be little to no code shared between this CNI and regular kuryr-kubernetes. In this case it feels like the code will be more suited to a separate project, than kuryr-kubernetes.
  • As an alternative we may implement a separate kuryr-sriov-cni, that would only handle SR-IOV requests. This will allow a more granular approach and would decouple SR-IOV functionality from the main code. Implementing a kuryr-sriov-cni would mean, however, that operators would need to pick one of the implementations (kuryr-cni vs kuryr-sriov-cni) or use something like multus-cni [6] or CNI-Genie [7] to allow them work together.


Primary assignee: Zaitsev Kirill

Work Items

  • Implement Multi-VIF handler/CNI
  • Implement SR-IOV capabilities
  • Implement CNI SR-IOV handler
  • Active state monitoring for kuryr-sriov direct ports
  • Document deployment procedure for kuryr-sriov support

Possible Further Work

  • It may be desirable to be able to request specific ports from neutron subnet in the Pod Spec. This functionality may be extended to normal VIFs, beyond SR-IOV handler.
  • It may be desirable to add an option to assign network info to VFs statically
Creative Commons Attribution 3.0 License

Except where otherwise noted, this document is licensed under Creative Commons Attribution 3.0 License. See all OpenStack Legal Documents.