Configure simultaneous multi-threading (SMT)


This page shows you how to configure simultaneous multi-threading in your Google Kubernetes Engine (GKE) clusters.

Overview

GKE nodes are Compute Engine virtual machine (VM) instances that GKE creates on your behalf. On these VMs, it's common to have simultaneous multi-threading (SMT) enabled by default. SMT is the ability to execute multiple independent threads on a single physical CPU core. With SMT enabled, a single physical core can run two virtual CPUs (vCPUs) as independent, separate threads. For example, an n2-standard-32 GKE node is, in fact, running 32 vCPUs on 16 physical CPU cores.

With the introduction of SMT configurability on GKE nodes, you can now configure SMT so that the number of effective vCPUs is equal to the number of physical cores, which means that your GKE nodes can use the full physical core per vCPU.

Benefits

Configuring SMT on GKE provides the following benefits.

Improved performance

Some High Performance Computing (HPC) workloads might see a performance improvement by disabling SMT. Typically, for many general computing tasks or tasks that require lots of I/O, SMT can increase application throughput significantly. In contrast, for compute-bound jobs in which both virtual cores are compute bound, SMT can hinder overall application performance and add nondeterministic variance to jobs. Therefore, with SMT disabled, some HPC applications might get better and more predictable performance.

Reduced licensing costs

Some software licensing agreements have billing models that are tied to the number of vCPUs that are consumed. Disabling SMT reduces the number of vCPUs for each core by half, which might reduce your licensing costs. For more information, consult your licensing agreement.

See the pricing section to understand how Google Cloud bills when SMT is configured.

Limitations

  • Machine types: SMT is not configurable on machine types that have fewer than 2 vCPUs (for example n1-standard-1) or on the Tau T2D machine type.
  • Node auto-provisioning: SMT configurability is not supported for clusters that have node auto-provisioning enabled.
  • Cluster autoscaler: Scaling from zero nodes with SMT configurability is only supported on the cluster autoscaler for clusters running GKE version 1.21 or later.
  • GKE Sandbox: You can configure SMT settings in GKE Sandbox nodes in GKE Standard clusters running version 1.24.2-gke.300 and later using the --threads-per-core flag, and in prior versions using a standalone DaemonSet. For instructions, refer to GKE Sandbox. Since multithreading is disabled by default in GKE Sandbox before version 1.24.2-gke.300, configuring SMT using threads-per-core is not supported. If you use GKE Sandbox and would like to enable SMT, see SMT may be disabled.
  • Cluster mode: SMT configurability is not supported on GKE Autopilot clusters.

Pricing

SMT for GKE is billed based on the underlying Compute Engine pricing for SMT. You are billed for the number of vCPUs defined by a VM's machine type, not the number of threads that run on each core. Even if you configure SMT, you will continue to be billed for the default number of vCPUs defined by the VM machine type. For more information about how you are billed for VMs, see VM instance pricing.

For example, a GKE node created using the n2-standard-8 machine type runs with SMT enabled by default, and can run up to 8 vCPUs, which is 2 vCPUs for each of the four physical CPU cores. If you configure SMT to use only one thread per core (effectively 4 vCPUs), you are still billed for 8 vCPUs.

Configure SMT

You can configure SMT on new clusters or node pools by specifying the number of threads per core.

The following table shows the possible values for threads per core and the corresponding SMT behavior for the node:

Threads per core SMT behavior
null (not specified) Compute Engine VM default
0 Invalid
1 SMT off
2 SMT on
> 2 Invalid

Configure SMT on a new cluster

You can configure SMT on new clusters using the gcloud CLI or the Google Cloud console.

gcloud

To configure SMT, run the following command:

gcloud container clusters create CLUSTER_NAME \
    --machine-type=MACHINE_TYPE \
    --threads-per-core=THREADS_PER_CORE

Replace the following:

  • CLUSTER_NAME: the name for the new cluster.
  • MACHINE_TYPE: the machine type to use for your GKE nodes. This must be a machine type with 2 or more vCPUs.
  • THREADS_PER_CORE: the number of threads per physical core. The previous table shows the values you can specify.

For a full list of options, refer to the gcloud container clusters create documentation.

Console

  1. Go to the Google Kubernetes Engine page in the Google Cloud console.

    Go to Google Kubernetes Engine

  2. Click Create.

  3. In the Create cluster dialog, next to GKE Standard, click Configure.

  4. In the navigation menu, in the Node pools section, click the name of a node pool, and then click Nodes.

  5. In the Machine Configuration section, select a supported Machine type, such as e2-standard-2.

  6. Click CPU Platform and GPU.

  7. From the vCPUs to core ratio drop-down menu, select one of the following:

    • 1 vCPU per core: turn SMT off.
    • 2 vCPUs per core: keep SMT on.
  8. Configure the cluster and node pools as required, and then click Create.

Configure SMT on a new node pool

You can configure SMT on new node pools using the gcloud CLI or the Google Cloud console.

gcloud

To configure SMT, run the following command:

gcloud container node-pools create POOL_NAME \
    --cluster=CLUSTER_NAME \
    --machine-type=MACHINE_TYPE \
    --threads-per-core=THREADS_PER_CORE

Replace the following:

  • POOL_NAME: the name for the new node pool.
  • CLUSTER_NAME: the name of the existing cluster.
  • MACHINE_TYPE: the machine type to use for your GKE nodes. This must be a machine type with 2 or more vCPUs.
  • THREADS_PER_CORE: the number of threads per physical core. The previous table shows the values you can specify.

For a full list of options, refer to the gcloud container clusters create documentation.

Console

  1. Go to the Google Kubernetes Engine page in the Google Cloud console.

    Go to Google Kubernetes Engine

  2. Click the name of the cluster that you want to modify.

  3. Click Add node pool.

  4. In the navigation menu, click Nodes.

  5. In the Machine Configuration section, select a Machine family and a Machine type that support SMT, such as n2-standard-2.

  6. Click CPU Platform and GPU.

  7. From the vCPUs to core ratio drop-down menu, select one of the following options:

    • 1 vCPU per core: turn SMT off.
    • 2 vCPUs per core: keep SMT on.
  8. Configure the node pool as needed, and then click Create.

Verify the SMT configuration

Linux

To verify the SMT configuration on Linux nodes, perform the following steps:

  1. SSH into the node:

    gcloud compute ssh NODE_NAME
    

    Replace NODE_NAME with the name of your node.

  2. Check the number of vCPUs:

    cat /proc/cpuinfo | grep processor | wc -l
    

    The output is similar to the following:

    16
    

    If you set the number of threads per core to 1, the output should show half the Compute Engine VM default number of vCPUs. For example, the output for the n2-standard-32 with SMT off should be 16 instead of the default value of 32.

Windows

To verify the SMT configuration on Windows nodes, perform the following steps:

  1. Set a username and password to RDP to the node:

    gcloud compute reset-windows-password NODE_NAME
    

    Replace NODE_NAME with the name of your node.

  2. Establish a connection to the EXTERNAL_IP:3389 for the VM.

  3. Run the following powershell command:

    Get-WmiObject -Class Win32_Processor -ComputerName. | Select-Object -Property [a-z]* | Select NumberOfCores,NumberOfLogicalProcessors
    

    If you set the number of threads per core to 1, the NumberOfCores value should be equal to the NumberOfLogicalProcessors value, such as in the following example output:

    Output (example n1-standard-2 with SMT off)
    
    NumberOfCores NumberOfLogicalProcessors
    ------------- -------------------------
                1                         1
    

What's next