Dedicated CPU resources

Certain workloads, requiring a predictable latency and enhanced performance during its execution would benefit from obtaining dedicated CPU resources. KubeVirt, relying on the Kubernetes CPU manager, is able to pin guest’s vCPUs to the host’s pCPUs.

Kubernetes CPU manager

Kubernetes CPU manager is a mechanism that affects the scheduling of workloads, placing it on a host which can allocate Guaranteed resources and pin certain Pod’s containers to host pCPUs, if the following requirements are met:

• Pod’s QoS is Guaranteed
  • resources requests and limits are equal
  • all containers in the Pod express CPU and memory requirements
• Requested number of CPUs is an Integer

Additional information:

• Enabling the CPU manager on Kubernetes
• Enabling the CPU manager on OKD
• Kubernetes blog explaining the feature

Requesting dedicated CPU resources

Setting spec.domain.cpu.dedicatedCpuPlacement to true in a VMI spec will indicate the desire to allocate dedicated CPU resource to the VMI

Kubevirt will verify that all the necessary conditions are met, for the Kubernetes CPU manager to pin the virt-launcher container to dedicated host CPUs. Once, virt-launcher is running, the VMI’s vCPUs will be pinned to the pCPUS that has been dedicated for the virt-launcher container.

Expressing the desired amount of VMI’s vCPUs can be done by either setting the guest topology in spec.domain.cpu (sockets, cores, threads) or spec.domain.resources.[requests/limits].cpu to a whole number integer ([1-9]+) indicating the number of vCPUs requested for the VMI. Number of vCPUs is counted as sockets * cores * threads or if spec.domain.cpu is empty then it takes value from spec.domain.resources.requests.cpu or spec.domain.resources.limits.cpu.

Note: Users should not specify both spec.domain.cpu and spec.domain.resources.[requests/limits].cpu

Note: spec.domain.resources.requests.cpu must be equal to spec.domain.resources.limits.cpu

Note: Multiple cpu-bound microbenchmarks show a significant performance advantage when using spec.domain.cpu.sockets instead of spec.domain.cpu.cores.

All inconsistent requirements will be rejected.

apiVersion: kubevirt.io/v1
kind: VirtualMachineInstance
spec:
  domain:
    cpu:
      sockets: 2
      cores: 1
      threads: 1
      dedicatedCpuPlacement: true
    resources:
      limits:
        memory: 2Gi
[...]

OR

apiVersion: kubevirt.io/v1
kind: VirtualMachineInstance
spec:
  domain:
    cpu:
      dedicatedCpuPlacement: true
    resources:
      limits:
        cpu: 2
        memory: 2Gi
[...]

Requesting dedicated CPU for QEMU emulator

A number of QEMU threads, such as QEMU main event loop, async I/O operation completion, etc., also execute on the same physical CPUs as the VMI’s vCPUs. This may affect the expected latency of a vCPU. In order to enhance the real-time support in KubeVirt and provide improved latency, KubeVirt will allocate an additional dedicated CPU, exclusively for the emulator thread, to which it will be pinned. This will effectively “isolate” the emulator thread from the vCPUs of the VMI. In case ioThreadsPolicy is set to auto IOThreads will also be “isolated” and placed on the same physical CPU as the QEMU emulator thread.

This functionality can be enabled by specifying isolateEmulatorThread: true inside VMI spec’s Spec.Domain.CPU section. Naturally, this setting has to be specified in a combination with a dedicatedCpuPlacement: true.

Example:

apiVersion: kubevirt.io/v1
kind: VirtualMachineInstance
spec:
  domain:
    cpu:
      dedicatedCpuPlacement: true
      isolateEmulatorThread: true
    resources:
      limits:
        cpu: 2
        memory: 2Gi

Compute Nodes with SMT Enabled

When the following conditions are met:

• The compute node has SMT enabled
• Kubelet’s CPUManager policy is set to static - full-pcpus-only
• The VM is configured to have an even number of CPUs
• dedicatedCpuPlacement and isolateEmulatorThread are enabled

The VM is scheduled, but rejected by the kubelet with the following event:

SMT Alignment Error: requested 3 cpus not multiple cpus per core = 2

In order to address this issue:

1. Enable the AlignCPUs feature gate in the KubeVirt CR.
2. Add the following annotation to the Kubevirt CR:

alpha.kubevirt.io/EmulatorThreadCompleteToEvenParity:

KubeVirt will then add one or two dedicated CPUs for the emulator threads, in a way that completes the total CPU count to be even.

Identifying nodes with a running CPU manager

At this time, Kubernetes doesn’t label the nodes that has CPU manager running on it.

KubeVirt has a mechanism to identify which nodes has the CPU manager running and manually add a cpumanager=true label. This label will be removed when KubeVirt will identify that CPU manager is no longer running on the node. This automatic identification should be viewed as a temporary workaround until Kubernetes will provide the required functionality. Therefore, this feature should be manually enabled by activating the CPUManager feature gate to the KubeVirt CR.

When automatic identification is disabled, cluster administrator may manually add the above label to all the nodes when CPU Manager is running.

• Nodes’ labels are view-able: kubectl describe nodes

• Administrators may manually label a missing node: kubectl label node [node_name] cpumanager=true

Sidecar containers and CPU allocation overhead

Note: In order to run sidecar containers, KubeVirt requires the Sidecar feature gate to be enabled in KubeVirt’s CR.

According to the Kubernetes CPU manager model, in order the POD would reach the required QOS level Guaranteed, all containers in the POD must express CPU and memory requirements. At this time, Kubevirt often uses a sidecar container to mount VMI’s registry disk. It also uses a sidecar container of it’s hooking mechanism. These additional resources can be viewed as an overhead and should be taken into account when calculating a node capacity.

Note: The current defaults for sidecar’s resources: CPU: 200m Memory: 64M As the CPU resource is not expressed as a whole number, CPU manager will not attempt to pin the sidecar container to a host CPU.