Description of problem:
virt-launcher pod consumes more memory than it should be allowed.
- The VM was configured with requests of 8G, and that is the total mem visible when we exec free on the guest os. The actual used mem observed from inside the VM is approx 2Gi. However, when inspecting the pod stats with crictl we can see mem usage by the pod is in dozens of Gi RAM.
Version-Release number of selected component (if applicable):
OpenShift 4.10.16, CNV in version 4.10.1, Kubevirt v0.49.0-155-g32b905704
How reproducible:
- VM has a single network interface attached to a cnv-bridge NAD, used for ingress/egress
- A pvc of 3.2 TB was attached to the VM, which downloaded 2.5 TB of data from an s3 bucket (ceph object storage). It was indeed saved to the machine.
- The workload was supposed to be CPU intensive, not memory intensive.
- The VM was configured with requests of 8G, and that is the total mem visible when we exec free on the guest os. The actual used mem observed from inside the VM is approx 2Gi. However, when inspecting the pod stats with crictl we can see mem usage by the pod is in dozens of Gi RAM.
- The attached PVC is an additional disk, not an extension of the root disk, and it is of type block with virtio-scsi driver. The customer used mkfs.xfs on the entire block with no partitions nor LVMs. We see the host cach
Actual results:
Expected results:
VM pod stays within limit of memory assigned to it.
Additional info: