Configure NUMA-aware VM scheduling

This document describes how to configure clusters and VMs to support high performance and low latency workloads with the computing efficiencies of non-uniform memory access (NUMA). There are instructions for tuning Kubernetes settings for cluster nodes. This document also includes instructions for configuring Virtual Machines (VMs) with NUMA affinity so that they get scheduled on and take advantage of the NUMA nodes.

With a NUMA-aware VM, all the communication within the VM is local to the NUMA node. The NUMA-aware VM avoids data transactions to and from remote resources that can degrade VM performance.

Configure nodes to use NUMA

The following sections describe how to configure the critical Kubernetes components to tune the node and make sure it can schedule NUMA-aware containers. These NUMA nodes are tuned to optimize CPU and memory performance. Follow the instructions for each node that you want to run NUMA-aware VMs.

Update the kubelet configuration

As part of the node configuration to support NUMA node affinity, you need to make the following changes in the kubelet configuration:

  • Enable the CPU Manager with a static policy
  • Enable the Memory Manager with a Static policy
  • Enable the Topology manager with restricted topology

To configure kubelet on your worker node:

  1. Locate the kubelet file on your worker node and open it for editing:

    edit /etc/default/kubelet

    If you don't see the kubelet file, create it with the following command:

    echo "KUBELET_EXTRA_ARGS=\"\"" >> /etc/default/kubelet

    This command creates the kubelet file with an empty KUBELET_EXTRA_ARGS="" section.

  2. To enable the CPU Manager with a static policy, add the --cpu-manager-policy=static flag to the KUBELET_EXTRA_ARGS="" section of the file:

  3. To enable the Memory Manager with a Static policy, add the --memory-manager-policy=Static flag to the KUBELET_EXTRA_ARGS="" section of the file:

    KUBELET_EXTRA_ARGS="--cpu-manager-policy=static --memory-manager-policy=Static"
  4. To enable the Topology Manager with a restricted policy, add the --topology-manager-policy=restricted flag to the KUBELET_EXTRA_ARGS="" section of the file:

    KUBELET_EXTRA_ARGS="--cpu-manager-policy=static --memory-manager-policy=Static --topology-manager-policy=restricted"
  5. Check the current amount of memory reserved by Anthos clusters on bare metal:

    cat /var/lib/kubelet/kubeadm-flags.env

    The output should look like the following:

    KUBELET_KUBEADM_ARGS="--anonymous-auth=false --authentication-token-webhook=true --authorization-mode=Webhook --container-runtime=remote --container-runtime-endpoint=unix:///run/containerd/containerd.sock --feature-gates=SeccompDefault=true --kube-reserved=cpu=100m,memory=3470Mi --max-pods=110 --node-ip=,,, --provider-id=baremetal:// --read-only-port=10255 --rotate-server-certificates=true --seccomp-default=true"

    The --kube-reserved=cpu=100m,memory=3470Mi setting indicates that Anthos clusters on bare metal has reserved 3,470 mebibytes of memory on the node.

  6. Set the --reserved-memory flag in the KUBELET_EXTRA_ARGS section of the kubelet file to 100 mebibytes more than the current reserved memory to account for the eviction threshold. If there is no reserved memory, you can skip this step.

    For example, with the reserved memory of 3470Mi from the example in the preceding step, you reserve 3570Mi of memory in the kubelet file:

    KUBELET_EXTRA_ARGS="--cpu-manager-policy=static --memory-manager-policy=Static --topology-manager-policy=restricted --reserved-memory=0:memory=3570Mi"
  7. Remove CPU and memory state files from the /var/lib directory:

    rm /var/lib/cpu_manager_state
    rm /var/lib/memory_manager_state
  8. Restart kubelet:

    systemctl start kubelet

For more information about these policy settings, see the following Kubernetes documentation:

Configure the node to use hugepages

Once you have enabled Memory Manager with the Static policy, you can add hugepages to further improve container workload performance on your NUMA nodes. Hugepages, as the name suggests, let you specify memory pages that are larger than the standard 4 kibibyte (KiB). Anthos VM Runtime supports 2 mebibyte (MiB) and 1 gibibyte (GiB) hugepages. You can set hugepages for a node at runtime, or for when the node machine boots. We recommend that you configure hugepages on each node that you want to run NUMA-aware VMs.

  1. To configure the number of hugepages of a specific size on your NUMA node at runtime, use the following command:

    echo HUGEPAGE_QTY > \

    Replace the following:

    • HUGEPAGE_QTY: the number of hugepages to allocate of the specified size.

    • NUMA_NODE: the NUMA node, such as node0, to which you're allocating hugepages.

    • HUGEPAGE_SIZE: the size of the hugepages in kibibytes, 2048 (2 MiB) or1048576 (1 GiB).

Configure a VM to use the NUMA node

Once your cluster nodes are tuned for NUMA, you can create NUMA-aware VMs. NUMA-aware VMs are scheduled on NUMA nodes.

To create a NUMA-aware VM:

  1. Follow the instructions to create a VM from a manifest.

    Use the following compute settings to configure your VM to be NUMA-aware:

    • spec.compute.guaranteed: Set guaranteed to true. With this setting, the virt-launcher Pod is configured to be placed in the Kubernetes Guaranteed Quality of Service (QoS) class.

    • spec.compute.advancedCompute:

      • dedicatedCPUPlacement: Set dedicatedCPUPlacement to true. This setting pins virtual CPUs to the physical CPUs of the node.
      • hugePageSize: Set hugePageSize to either 2Mi or 1Gi to specify the hugepages size for your VM to use, 2 mebibyte or 1 gibibyte.
      • numaGuestMappingPassthrough: Include an empty structure ({}) for this setting. This setting establishes NUMA affinity so that your VM is scheduled on NUMA nodes only.

    The following example VirtualMachine manifest shows how a NUMA-aware VM configuration might look:

    kind: VirtualMachine
      name: vm1
          vcpus: 2
        guaranteed: true
          dedicatedCPUPlacement: true
          hugePageSize: 2Mi
          numaGuestMappingPassthrough: {}
          capacity: 256Mi
      - name: eth0
        networkName: pod-network
        default: true
      - virtualMachineDiskName: disk-from-gcs
        boot: true
        readOnly: true