Resource requests in Autopilot

This page describes the maximum, minimum, and default resource requests that you can specify for your Google Kubernetes Engine (GKE) Autopilot workloads, and how Autopilot automatically modifies those requests to maintain workload stability.

Overview of resource requests in Autopilot

Autopilot uses the resource requests that you specify in your workload configuration to configure the nodes that run your workloads. Autopilot enforces minimum and maximum resource requests based on the compute class or the hardware configuration that your workloads use. If you don't specify requests for some containers, Autopilot assigns default values to let those containers run correctly.

When you deploy a workload in an Autopilot cluster, GKE validates the workload configuration against the allowed minimum and maximum values for the selected compute class or hardware configuration (such as GPUs). If your requests are less than the minimum, Autopilot automatically modifies your workload configuration to bring your requests within the allowed range. If your requests are greater than the maximum, Autopilot rejects your workload and displays an error message.

The following list summarizes the categories of resource requests:

• Default resource requests: Autopilot adds these if you don't specify your own requests for workloads
• Minimum and maximum resource requests: Autopilot validates your specified requests to ensure that they're within these limits. If your requests are outside the limits, Autopilot modifies your workload requests.
• Workload separation and extended duration requests: Autopilot has different default values and different minimum values for workloads that you separate from each other, or for Pods that get extended protection from GKE-initiated eviction.
• Resource requests for DaemonSets: Autopilot has different default, minimum, and maximum values for containers in DaemonSets.

How to request resources

In Autopilot, you request resources in your Pod specification. The supported minimum and maximum resources that you can request change based on the hardware configuration of the node on which the Pods run. To learn how to request specific hardware configurations, refer to the following pages:

• Choose compute classes for Autopilot Pods
• Deploy GPU workloads in Autopilot

Default resource requests

If you don't specify resource requests for some containers in a Pod, Autopilot applies default values. These defaults are suitable for many smaller workloads.

Additionally, Autopilot applies the following default resource requests regardless of the selected compute class or hardware configuration:

• Containers in DaemonSets

  • CPU: 50 mCPU
  • Memory: 100 MiB
  • Ephemeral storage: 100 MiB

• All other containers

  • Ephemeral storage: 1 GiB

For more information about Autopilot cluster limits, see Quotas and limits.

Default requests for compute classes

Autopilot applies the following default values to resources that are not defined in the Pod specification for Pods that run on compute classes. If you only set one of the requests and leave the other blank, GKE uses the CPU:memory ratio defined in the Minimum and maximum requests section to set the missing request to a value that complies with the ratio.

Default requests for accelerators

The following table describes the default values that GKE assigns to Pods that don't specify values in the requests field of the Pod specification. This table applies to Pods that use the Accelerator compute class, which is the recommended way to run accelerators in Autopilot clusters.

Supported GPUs without the Accelerator compute class

If you don't use the Accelerator compute class, only the following GPUs are supported. The default resource requests for these GPUs are the same as in the Accelerator compute class:

• NVIDIA A100 (40GB)
• NVIDIA A100 (80GB)
• NVIDIA L4
• NVIDIA Tesla T4

Minimum and maximum resource requests

The total resources requested by your deployment configuration should be within the supported minimum and maximum values that Autopilot allows. The following conditions apply:

• Ephemeral storage requests:

  • Ephemeral storage uses the VM boot disk unless your nodes have local SSDs attached.

    Compute hardware that includes local SSDs like A100 (80GB) GPUs, H100 (80GB) GPUs, or the Z3 machine series support a maximum request that's equal to the size of the local SSD minus any system overhead. For information about this system overhead, see Ephemeral storage backed by local SSDs.

  • In GKE version 1.29.3-gke.1038000 and later, Performance class Pods and hardware accelerator Pods support a maximum ephemeral storage request of 56 Ti unless the hardware includes local SSDs.

    In all other Autopilot Pods regardless of the GKE version, the total ephemeral storage request across all of the containers in the Pod must be between 10 MiB and 10 GiB unless otherwise specified.

  • For larger volumes, use generic ephemeral volumes, which provide equivalent functionality and performance to ephemeral storage but with significantly more flexibility as they can be used with any GKE storage option. For example, the maximum size for a generic ephemeral volume using pd-balanced is 64 TiB.

• For DaemonSet Pods, the minimum resource requests are as follows:

  • Clusters that support bursting: 1 mCPU per Pod, 2 MiB of memory per Pod, and 10 MiB of ephemeral storage per container in the Pod.
  • Clusters that don't support bursting: 10mCPU per Pod, 10 MiB of memory per Pod, and 10 MiB of ephemeral storage per container in the Pod.

  To check whether your cluster supports bursting, see Bursting availability in GKE.

• If your cluster supports bursting, Autopilot doesn't enforce 0.25 vCPU increments for your Pod CPU requests. If your cluster doesn't support bursting, Autopilot rounds up your CPU requests to the nearest 0.25 vCPU. To check whether your cluster supports bursting, see Bursting availability in GKE.

• The CPU:memory ratio must be within the allowed range for the selected compute class or hardware configuration. If your CPU:memory ratio is outside the allowed range, Autopilot automatically increases the smaller resource. For example, if you request 1 vCPU and 16 GiB of memory (1:16 ratio) for Pods running on the Scale-Out class, Autopilot increases the CPU request to 4 vCPUs, which changes the ratio to 1:4.

Minimums and maximums for compute classes

The following table describes the minimum, maximum, and allowed CPU-to-memory ratio for each compute class that Autopilot supports:

To learn how to request compute classes in your Autopilot Pods, refer to Choose compute classes for Autopilot Pods.

Minimums and maximums for accelerators

The following sections describe the minimum, maximum, and allowed CPU-to-memory ratio for Pods that use hardware accelerators like GPUs and TPUs.

Unless specified, the maximum ephemeral storage supported is 122 GiB in versions 1.28.6-gke.1369000 or later, and 1.29.1-gke.1575000 or later. For earlier versions, the maximum ephemeral storage supported is 10 GiB.

Minimums and maximums for the Accelerator compute class

The following table shows the minimum and maximum resource requests for Pods that use the Accelerator compute class, which is the recommended way to run accelerators with GKE Autopilot clusters. In the Accelerator compute class, GKE doesn't enforce CPU-to-memory request ratios.

To learn how to request GPUs in your Autopilot Pods, refer to Deploy GPU workloads in Autopilot.

Minimums and maximums for GPUs without a compute class

The following table shows the minimum and maximum resource requests for Pods that don't use the Accelerator compute class:

To learn how to request GPUs in your Autopilot Pods, refer to Deploy GPU workloads in Autopilot.

Resource requests for workload separation and extended duration

Autopilot lets you manipulate Kubernetes scheduling and eviction behavior using methods such as the following:

• Use taints and tolerations and node selectors to ensure that certain Pods only get placed on specific nodes. For details, see Configure workload separation in GKE.
• Use Pod anti-affinity to prevent Pods from co-locating on the same node. The default and minimum resource requests for workloads that use these methods to control scheduling behavior are higher than for workloads that don't.
• Use an annotation to protect Pods from eviction caused by node auto-upgrades and scale-down events for up to seven days. For details, see Extend the run time of Autopilot Pods.

If your specified requests are less than the minimums, the behavior of Autopilot changes based on the method that you used, as follows:

• Taints, tolerations, selectors, and extended duration Pods: Autopilot modifies your Pods to increase the requests when scheduling the Pods.
• Pod anti-affinity: Autopilot rejects the Pod and displays an error message.

The following table describes the default requests and the minimum resource requests that you can specify. If a configuration or compute class isn't in this table, Autopilot doesn't enforce special minimum or default values.

Init containers

Init containers run in serial and must complete before the application containers start. If you don't specify resource requests for your Autopilot init containers, GKE allocates the total resources available to the Pod to each init container. This behavior is different than in GKE Standard, where each init container can use any unallocated resources available on the node on which the Pod is scheduled.

Unlike application containers, GKE recommends that you don't specify resource requests for Autopilot init containers, so that each container gets the full resources available to the Pod. If you request less resources than the defaults, you constrain your init container. If you request more resources than the Autopilot defaults, you might increase your bill for the lifetime of the Pod.

Setting resource limits in Autopilot

Kubernetes lets you set both requests and limits for resources in your Pod specification. The behavior of your Pods changes depending on whether your limits are different than your requests, as described in the following table:

resources:
  limits:
    cpu: "400m"

resources:
  requests:
    cpu: "400m"
  limits:
    cpu: "400m"

resources:
  requests:
    cpu: "450m"
  limits:
    cpu: "400m"

resources:
  requests:
    cpu: "400m"
  limits:
    cpu: "400m"

In most situations, set adequate resource requests and equal limits for your workloads.

For workloads that temporarily need more resources than their steady-state, like during boot up or during higher traffic periods, set your limits higher than your requests to let the Pods burst. For details, see Configure Pod bursting in GKE.

Automatic resource management in Autopilot

If your specified resource requests for your workloads are outside of the allowed ranges, or if you don't request resources for some containers, Autopilot modifies your workload configuration to comply with the allowed limits. Autopilot calculates resource ratios and the resource scale up requirements after applying default values to containers with no request specified.

• Missing requests: If you don't request resources in some containers, Autopilot applies the default requests for the compute class or hardware configuration.
• CPU:memory ratio: Autopilot scales up the smaller resource to bring the ratio within the allowed range.
• Ephemeral storage: Autopilot modifies your ephemeral storage requests to meet the minimum amount required by each container. The cumulative value of storage requests across all containers cannot be more than the maximum allowed value. Prior to 1.28.6-gke.1317000, Autopilot scales down the requested ephemeral storage if the value exceeds the maximum. In version 1.28.6-gke.1317000 and later, Autopilot rejects your workload.
• Requests below minimums: If you request fewer resources than the allowed minimum for the selected hardware configuration, Autopilot automatically modifies the Pod to request at least the minimum resource value.

By default, when Autopilot automatically scales a resource up to meet a minimum or default resource value, GKE allocates the extra capacity to the first container in the Pod manifest. In GKE version 1.27.2-gke.2200 and later, you can tell GKE to allocate the extra resources to a specific container by adding the following to the annotations field in your Pod manifest:

autopilot.gke.io/primary-container: "CONTAINER_NAME"

Replace CONTAINER_NAME with the name of the container.

Resource modification examples

The following example scenario shows how Autopilot modifies your workload configuration to meet the requirements of your running Pods and containers.

Single container with < 0.05 vCPU

Multiple containers with total CPU < 0.05 vCPU

Single container with memory too low for requested CPU

In this example, the memory is too low for the amount of CPU (1 vCPU:1 GiB minimum). The minimum allowed ratio for CPU to memory is 1:1. If the ratio is lower than that, the memory request is increased.

What's next

• Learn how to select compute classes in your Autopilot workloads.
• Learn more about the supported Autopilot compute classes.
• Learn how to select GPUs in your Autopilot Pods.