Skip to content

Storage Overview

Configuring storage for KubeVirt Hosted Control Planes falls within three categories, each of which have differing requirements.

ETCD Storage

ETCD requires usage of high performance persistent storage on the management cluster hosting the ETCD pods. Due to the performance requirements, usage of a local storage csi driver such as LVM Storage is prefered. When a guest cluster is created in HighAvailability mode, ETCD is replicated in pods across three separate management cluster nodes. This replication ensures data resiliency even when a local storage csi driver is in use.

More information about ETCD storage configuration can be found in the ETCD Storage Configuration section.

Node Root Volume Storage

In a KubeVirt Hosted Control Plane, the worker nodes are hosted in KubeVirt VMs. It is recommended to use a csi driver capable of providing ReadWriteMany access mode and Block volume mode for the VM root volume storage. This allows the KubeVirt VMs to live migrate and remain available even when the underlying infra cluster nodes are disrupted. Ceph is an example of a csi driver that would meet these requirements.

More information about configuring root volume storage can be found in the KubeVirt VM Root Volume Configuration section.

KubeVirt CSI Storage

The KubeVirt CSI driver allows storage classes present on the infra cluster (the cluster the KubeVirt VMs run on) to be mapped directly to storage classes within the KubeVirt guest cluster. This lets the guest cluster utilize the same storage that is available on the infra cluster.

It is recommended to use an underlying infra storage class capable of ReadWriteMany access mode and Block volume mode for KubeVirt CSI. This allows KubeVirt CSI to pass storage to the VMs in a way that still allows for the VMs to live migrate and be portable across infra nodes.

Below is a chart that outlines the current features of KubeVirt CSI as they map to the infra cluster's storage class.

Infra CSI Capability Guest CSI Capability VM Live Migration Support Notes
RWX Block RWO (Block/Filesystem) RWX (Block) Supported
RWO Block RWO (Block/Filesystem) Not Supported
RWO Filesystem RWO (Block/Filesystem) Not Supported suboptimal guest block volume mode performance.

More information about configuring KubeVirt CSI can be found in the KubeVirt CSI StorageClass Mapping section.

Configuring Storage

By default, if no advanced configuration is provided, the default storageclass is used for the the KubeVirt VM images, the kubevirt csi mapping, and the etcd volumes.

KubeVirt CSI StorageClass Mapping

KubeVirt CSI permits any infra storage class with the ReadWriteMany access mode to be exposed to the guest cluster. This mapping of infra cluster storage class to guest cluster storage class can be configured during cluster creation using the hcp cli tool and the --infra-storage-class-mapping cli argument.

Below is an example of how to map two infra storage classes called infra-sc1 and infra-sc2 to guest storage classes called guest-sc1 and guest-sc2. Note that --infra-storage-class-mapping argument can be used multiple times within the create command.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
export CLUSTER_NAME=example
export PULL_SECRET="$HOME/pull-secret"
export MEM="6Gi"
export CPU="2"
export WORKER_COUNT="2"

hcp create cluster kubevirt \
--name $CLUSTER_NAME \
--node-pool-replicas $WORKER_COUNT \
--pull-secret $PULL_SECRET \
--memory $MEM \
--cores $CPU \
--infra-storage-class-mapping=infra-sc1/guest-sc1 \
--infra-storage-class-mapping=infra-sc2/guest-sc2

Once the guest cluster is created, the guest-sc1 and guest-sc2 storage classes will be visible within the guest cluster. When users create a PVC within the guest cluster that utilizes one of these storage classes, KubeVirt CSI will facilitate provisioning that volume using the infra storage class mapping that was configured during cluster creation.

Note

KubeVirt CSI only supports mapping an infrastructure storage class that is capable of ReadWriteMany (RWX) access.

KubeVirt CSI VolumeSnapshotClass Mapping

KubeVirt CSI permits infra VolumeSnapshotClasses to be exposed to the guest cluster. Since VolumeSnapshotClasses are tied to a particular provisioner the mapping between VolumeSnapshotClasses and StorageClasses needs to expressed to hypershift during guest cluster creation. using the hcp cli tool and the --infra-volumesnapshot-class-mapping cli argument.

Below is an example of a simple setup with a single infra storage class and a single matching volume snapshot class in the infra cluster being mapped to a single storage class and single volume snapshot class in the guest cluster.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
export CLUSTER_NAME=example
export PULL_SECRET="$HOME/pull-secret"
export MEM="6Gi"
export CPU="2"
export WORKER_COUNT="2"

hcp create cluster kubevirt \
--name $CLUSTER_NAME \
--node-pool-replicas $WORKER_COUNT \
--pull-secret $PULL_SECRET \
--memory $MEM \
--cores $CPU \
--infra-storage-class-mapping=infra-sc1/guest-sc1 \
--infra-volumesnapshot-class-mapping=infra-vsc1/guest-vsc1

If you omit the --infra-storage-class-mapping and the --infra-volumesnapshot-class-mapping. The system will use the default storage class and the default volume snapshot class in the infra cluster. If the default is not set, then the snapshot functionality will not work and the snapshot request will never reach ready state. This is because it not possible to create a correct snapshot in the infra cluster.

A more complex setup could contain multiple storage classes with multiple volume snapshot classes. In particular in this setup one volume snapshot class is only compatible with certain storage classes but not all. So we have infra storage class a and b, and infra snapshot volume class a and b. Only the a's are compatible with each other and only the b's are compatible with each other.

To properly group them together use the 'group' option of the --infra-volumesnapshot-class-mapping and group option of the --infra-storage-class-mapping. Below is an example of this setup

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
export CLUSTER_NAME=example
export PULL_SECRET="$HOME/pull-secret"
export MEM="6Gi"
export CPU="2"
export WORKER_COUNT="2"

hcp create cluster kubevirt \
--name $CLUSTER_NAME \
--node-pool-replicas $WORKER_COUNT \
--pull-secret $PULL_SECRET \
--memory $MEM \
--cores $CPU \
--infra-storage-class-mapping=infra-sca/guest-sca,group=a \
--infra-storage-class-mapping=infra-scb/guest-scb,group=b \
--infra-storage-class-mapping=infra-scc/guest-scc,group=a \
--infra-volumesnapshot-class-mapping=infra-vsca/guest-vsca,group=a \
--infra-volumesnapshot-class-mapping=infra-vscb/guest-vscb,group=b

Since both storage class infra-sca and volume snapshot class infra-vsca are in the same group, this indicates to KubeVirt CSI that they are compatible and be used to create snapshots of volumes from storage class guest-sca using the guest volume snapshot class guest-vsca. The same is true with with the b grouping as well. Since infra-scc is also in the a group, creating snapshots of volumes from storage class guest-scc will use the same volume snapshot class in the infra cluster as making a snapshot of volumes that use storage class guest-sca

Note

KubeVirt CSI passes snapshot requests to the underlying infra. This means that snapshots will only work for compatible volumes. Please ensure the proper mapping is configured before attempting to create a snapshot in the guest cluster.

KubeVirt VM Root Volume Configuration

The storage class used to host the KubeVirt Virtual Machine root volumes can be configured at cluster creation time using the hcp cli tool and the --root-volume-storage-class argument. Likewise, the size of the volume can be configured using the --root-volume-size cli argument.

Below is an example of setting a custom storage class and volume size for the KubeVirt VMs. The result will be a guest cluster with VMs hosted on 64Gi PVCs hosted by the ocs-storagecluster-ceph-rdb storage class

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
export CLUSTER_NAME=example
export PULL_SECRET="$HOME/pull-secret"
export MEM="6Gi"
export CPU="2"
export WORKER_COUNT="2"

hcp create cluster kubevirt \
--name $CLUSTER_NAME \
--node-pool-replicas $WORKER_COUNT \
--pull-secret $PULL_SECRET \
--memory $MEM \
--cores $CPU \
--root-volume-storage-class ocs-storagecluster-ceph-rbd \
--root-volume-size 64

KubeVirt VM Image Caching

KubeVirt VM image caching is an advanced feature that can be used to optimize both cluster startup time and storage utilization. This feature requires the usage of a storage class capable of smart cloning and the ReadWriteMany access mode.

Image caching works by importing the VM image once to a single PVC associated with the HostedCluster, and then making a unique clone of that PVC for every KubeVirt VM added as a worker node to the cluster. This reduces VM startup time by only requiring a single image import, and can further reduce overall cluster storage usage when the storage class supports copy on write cloning.

Image caching can be enabled during cluster creation using the hcp cli tool with the --root-volume-cache-strategy=PVC argument. Below is an example.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
export CLUSTER_NAME=example
export PULL_SECRET="$HOME/pull-secret"
export MEM="6Gi"
export CPU="2"
export WORKER_COUNT="2"

hcp create cluster kubevirt \
--name $CLUSTER_NAME \
--node-pool-replicas $WORKER_COUNT \
--pull-secret $PULL_SECRET \
--memory $MEM \
--cores $CPU \
--root-volume-cache-strategy=PVC

ETCD Storage Configuration

The storage class used to host the etcd data can be customized at cluster creation time using the hcp cli and the --etcd-storage-class cli argument. When no --etcd-storage-class argument is provided, the default storage class will be used.

Below is an example of how to configure usage of a storage class for etcd.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
export CLUSTER_NAME=example
export PULL_SECRET="$HOME/pull-secret"
export MEM="6Gi"
export CPU="2"
export WORKER_COUNT="2"

hcp create cluster kubevirt \
--name $CLUSTER_NAME \
--node-pool-replicas $WORKER_COUNT \
--pull-secret $PULL_SECRET \
--memory $MEM \
--cores $CPU \
--etcd-storage-class="local"