Skip to main content

VM Performance

Palette Virtual Machine Orchestrator (VMO) allows administrators to oversubscribe the physical resources on a host to maximize the number of active workloads.

VM workloads typically have varying resource demands and peak utilization patterns. By oversubscribing resources, it is possible to allocate them flexibly and take advantage of the fact that not all VMs will require their maximum allocation simultaneously.

The hypervisor automatically overcommits CPU and memory. This means that more virtualized CPU and memory can be allocated to VMs than there are physical resources on the system.


CPU Overcommit​

Kubevirt offers the cpuAllocationRatio in its Custom Resource Definitions (CRD). This ratio is used to normalize the amount of CPU time the pod will request based on the number of virtual CPUs (vCPUs).

Using the following algorithm, when cpuAllocationRatio is set to 1, the full amount of vCPUs are requested for the pod: pod CPU request = number of vCPUs * 1/cpuAllocationRatio.

The cpuAllocationRatio is global, so setting it to greater than 1 has the effect of requesting less CPU from Kubernetes for each VM.

Certain workloads that require a predictable latency and enhanced performance would benefit from obtaining dedicated CPU resources. KubeVirt relies on the Kubernetes CPU manager to pin vCPUs to the physical host’s CPUs. To learn more, refer to Dedicated CPU Resources and Resources Requests and Limits Kubevirt documentation.


warning

  • We do not recommend overcommitting CPUs in a production environment without extensive testing. Applications that use 100 percent of processing resources may become unstable in overcommitted environments.

  • Ensure you don't overcommit guest VMs on more than the physical number of processing cores. For example, a guest VM with four vCPUs should only be deployed on a host physical machine with a quad-core processor instead of a dual-core processor.

    We recommend no more than 10 total allocated vCPUs per physical processor core.


Memory Overcommit​

KubeVirt allows you to assign more or less memory to a VM than a VM requests to Kubernetes. You may want to overcommit VM memory if you have a cluster or a few nodes that are dedicated to running VMs. In this case, overcommitting memory makes use of all the memory in the nodes regardless of reserved or requested memory from the system.

To learn about options for memory overcommitment, refer to Node Overcommit KubeVirt resource.

You can make several changes to reduce the memory footprint and overcommit the per-VMI memory overhead.


  • Enable guest overhead overcommit by setting spec.domain.resources.overcommitGuestOverhead to true.

  • Enable guest memory by setting spec.domain.memory.guest to a value higher than spec.domain.resources.requests.memory, as shown in the example.

apiVersion: kubevirt.io/v1alpha3
kind: VirtualMachineInstance
metadata:
  name: testvmi-nocloud
spec:
  terminationGracePeriodSeconds: 30
  domain:
    resources:
      overcommitGuestOverhead: true
      requests:
        memory: 1024M
    memory:
      guest: 2048M

  • Enable implicit memory overcommit by setting spec.configuration.developerConfiguration.memoryOvercommit in the KubeVirt CRD to a percentage of the desired memory overcommit.

Resources​