Concepts

Edit This Page

Managing Compute Resources for Containers

When you specify a Pod, you can optionally specify how much CPU and memory (RAM) each Container needs. When Containers have resource requests specified, the scheduler can make better decisions about which nodes to place Pods on. And when Containers have their limits specified, contention for resources on a node can be handled in a specified manner. For more details about the difference between requests and limits, see Resource QoS.

Resource types

CPU and memory are each a resource type. A resource type has a base unit. CPU is specified in units of cores, and memory is specified in units of bytes. If you’re using Kubernetes v1.14 or newer, you can specify huge page resources. Huge pages are a Linux-specific feature where the node kernel allocates blocks of memory that are much larger than the default page size.

For example, on a system where the default page size is 4KiB, you could specify a limit, hugepages-2Mi: 80Mi. If the container tries allocating over 40 2MiB huge pages (a total of 80 MiB), that allocation fails.

Note: You cannot overcommit hugepages-* resources. This is different from the memory and cpu resources.

CPU and memory are collectively referred to as compute resources, or just resources. Compute resources are measurable quantities that can be requested, allocated, and consumed. They are distinct from API resources. API resources, such as Pods and Services are objects that can be read and modified through the Kubernetes API server.

Resource requests and limits of Pod and Container

Each Container of a Pod can specify one or more of the following:

  • spec.containers[].resources.limits.cpu
  • spec.containers[].resources.limits.memory
  • spec.containers[].resources.limits.hugepages-<size>
  • spec.containers[].resources.requests.cpu
  • spec.containers[].resources.requests.memory
  • spec.containers[].resources.requests.hugepages-<size>

Although requests and limits can only be specified on individual Containers, it is convenient to talk about Pod resource requests and limits. A Pod resource request/limit for a particular resource type is the sum of the resource requests/limits of that type for each Container in the Pod.

Meaning of CPU

Limits and requests for CPU resources are measured in cpu units. One cpu, in Kubernetes, is equivalent to 1 vCPU/Core for cloud providers and 1 hyperthread on bare-metal Intel processors.

Fractional requests are allowed. A Container with spec.containers[].resources.requests.cpu of 0.5 is guaranteed half as much CPU as one that asks for 1 CPU. The expression 0.1 is equivalent to the expression 100m, which can be read as “one hundred millicpu”. Some people say “one hundred millicores”, and this is understood to mean the same thing. A request with a decimal point, like 0.1, is converted to 100m by the API, and precision finer than 1m is not allowed. For this reason, the form 100m might be preferred.

CPU is always requested as an absolute quantity, never as a relative quantity; 0.1 is the same amount of CPU on a single-core, dual-core, or 48-core machine.

Meaning of memory

Limits and requests for memory are measured in bytes. You can express memory as a plain integer or as a fixed-point integer using one of these suffixes: E, P, T, G, M, K. You can also use the power-of-two equivalents: Ei, Pi, Ti, Gi, Mi, Ki. For example, the following represent roughly the same value:

128974848, 129e6, 129M, 123Mi

Here’s an example. The following Pod has two Containers. Each Container has a request of 0.25 cpu and 64MiB (226 bytes) of memory. Each Container has a limit of 0.5 cpu and 128MiB of memory. You can say the Pod has a request of 0.5 cpu and 128 MiB of memory, and a limit of 1 cpu and 256MiB of memory.

apiVersion: v1
kind: Pod
metadata:
  name: frontend
spec:
  containers:
  - name: db
    image: mysql
    env:
    - name: MYSQL_ROOT_PASSWORD
      value: "password"
    resources:
      requests:
        memory: "64Mi"
        cpu: "250m"
      limits:
        memory: "128Mi"
        cpu: "500m"
  - name: wp
    image: wordpress
    resources:
      requests:
        memory: "64Mi"
        cpu: "250m"
      limits:
        memory: "128Mi"
        cpu: "500m"

How Pods with resource requests are scheduled

When you create a Pod, the Kubernetes scheduler selects a node for the Pod to run on. Each node has a maximum capacity for each of the resource types: the amount of CPU and memory it can provide for Pods. The scheduler ensures that, for each resource type, the sum of the resource requests of the scheduled Containers is less than the capacity of the node. Note that although actual memory or CPU resource usage on nodes is very low, the scheduler still refuses to place a Pod on a node if the capacity check fails. This protects against a resource shortage on a node when resource usage later increases, for example, during a daily peak in request rate.

How Pods with resource limits are run

When the kubelet starts a Container of a Pod, it passes the CPU and memory limits to the container runtime.

When using Docker:

  • The spec.containers[].resources.requests.cpu is converted to its core value, which is potentially fractional, and multiplied by 1024. The greater of this number or 2 is used as the value of the --cpu-shares flag in the docker run command.

  • The spec.containers[].resources.limits.cpu is converted to its millicore value and multiplied by 100. The resulting value is the total amount of CPU time that a container can use every 100ms. A container cannot use more than its share of CPU time during this interval.

Note: The default quota period is 100ms. The minimum resolution of CPU quota is 1ms.

  • The spec.containers[].resources.limits.memory is converted to an integer, and used as the value of the --memory flag in the docker run command.

If a Container exceeds its memory limit, it might be terminated. If it is restartable, the kubelet will restart it, as with any other type of runtime failure.

If a Container exceeds its memory request, it is likely that its Pod will be evicted whenever the node runs out of memory.

A Container might or might not be allowed to exceed its CPU limit for extended periods of time. However, it will not be killed for excessive CPU usage.

To determine whether a Container cannot be scheduled or is being killed due to resource limits, see the Troubleshooting section.

Monitoring compute resource usage

The resource usage of a Pod is reported as part of the Pod status.

If optional tools for monitoring are available in your cluster, then Pod resource usage can be retrieved either from the Metrics API directly or from your monitoring tools.

Troubleshooting

My Pods are pending with event message failedScheduling

If the scheduler cannot find any node where a Pod can fit, the Pod remains unscheduled until a place can be found. An event is produced each time the scheduler fails to find a place for the Pod, like this:

kubectl describe pod frontend | grep -A 3 Events
Events:
  FirstSeen LastSeen   Count  From          Subobject   PathReason      Message
  36s   5s     6      {scheduler }              FailedScheduling  Failed for reason PodExceedsFreeCPU and possibly others

In the preceding example, the Pod named “frontend” fails to be scheduled due to insufficient CPU resource on the node. Similar error messages can also suggest failure due to insufficient memory (PodExceedsFreeMemory). In general, if a Pod is pending with a message of this type, there are several things to try:

  • Add more nodes to the cluster.
  • Terminate unneeded Pods to make room for pending Pods.
  • Check that the Pod is not larger than all the nodes. For example, if all the nodes have a capacity of cpu: 1, then a Pod with a request of cpu: 1.1 will never be scheduled.

You can check node capacities and amounts allocated with the kubectl describe nodes command. For example:

kubectl describe nodes e2e-test-node-pool-4lw4
Name:            e2e-test-node-pool-4lw4
[ ... lines removed for clarity ...]
Capacity:
 cpu:                               2
 memory:                            7679792Ki
 pods:                              110
Allocatable:
 cpu:                               1800m
 memory:                            7474992Ki
 pods:                              110
[ ... lines removed for clarity ...]
Non-terminated Pods:        (5 in total)
  Namespace    Name                                  CPU Requests  CPU Limits  Memory Requests  Memory Limits
  ---------    ----                                  ------------  ----------  ---------------  -------------
  kube-system  fluentd-gcp-v1.38-28bv1               100m (5%)     0 (0%)      200Mi (2%)       200Mi (2%)
  kube-system  kube-dns-3297075139-61lj3             260m (13%)    0 (0%)      100Mi (1%)       170Mi (2%)
  kube-system  kube-proxy-e2e-test-...               100m (5%)     0 (0%)      0 (0%)           0 (0%)
  kube-system  monitoring-influxdb-grafana-v4-z1m12  200m (10%)    200m (10%)  600Mi (8%)       600Mi (8%)
  kube-system  node-problem-detector-v0.1-fj7m3      20m (1%)      200m (10%)  20Mi (0%)        100Mi (1%)
Allocated resources:
  (Total limits may be over 100 percent, i.e., overcommitted.)
  CPU Requests    CPU Limits    Memory Requests    Memory Limits
  ------------    ----------    ---------------    -------------
  680m (34%)      400m (20%)    920Mi (12%)        1070Mi (14%)

In the preceding output, you can see that if a Pod requests more than 1120m CPUs or 6.23Gi of memory, it will not fit on the node.

By looking at the Pods section, you can see which Pods are taking up space on the node.

The amount of resources available to Pods is less than the node capacity, because system daemons use a portion of the available resources. The allocatable field NodeStatus gives the amount of resources that are available to Pods. For more information, see Node Allocatable Resources.

The resource quota feature can be configured to limit the total amount of resources that can be consumed. If used in conjunction with namespaces, it can prevent one team from hogging all the resources.

My Container is terminated

Your Container might get terminated because it is resource-starved. To check whether a Container is being killed because it is hitting a resource limit, call kubectl describe pod on the Pod of interest:

kubectl describe pod simmemleak-hra99
Name:                           simmemleak-hra99
Namespace:                      default
Image(s):                       saadali/simmemleak
Node:                           kubernetes-node-tf0f/10.240.216.66
Labels:                         name=simmemleak
Status:                         Running
Reason:
Message:
IP:                             10.244.2.75
Replication Controllers:        simmemleak (1/1 replicas created)
Containers:
  simmemleak:
    Image:  saadali/simmemleak
    Limits:
      cpu:                      100m
      memory:                   50Mi
    State:                      Running
      Started:                  Tue, 07 Jul 2015 12:54:41 -0700
    Last Termination State:     Terminated
      Exit Code:                1
      Started:                  Fri, 07 Jul 2015 12:54:30 -0700
      Finished:                 Fri, 07 Jul 2015 12:54:33 -0700
    Ready:                      False
    Restart Count:              5
Conditions:
  Type      Status
  Ready     False
Events:
  FirstSeen                         LastSeen                         Count  From                              SubobjectPath                       Reason      Message
  Tue, 07 Jul 2015 12:53:51 -0700   Tue, 07 Jul 2015 12:53:51 -0700  1      {scheduler }                                                          scheduled   Successfully assigned simmemleak-hra99 to kubernetes-node-tf0f
  Tue, 07 Jul 2015 12:53:51 -0700   Tue, 07 Jul 2015 12:53:51 -0700  1      {kubelet kubernetes-node-tf0f}    implicitly required container POD   pulled      Pod container image "k8s.gcr.io/pause:0.8.0" already present on machine
  Tue, 07 Jul 2015 12:53:51 -0700   Tue, 07 Jul 2015 12:53:51 -0700  1      {kubelet kubernetes-node-tf0f}    implicitly required container POD   created     Created with docker id 6a41280f516d
  Tue, 07 Jul 2015 12:53:51 -0700   Tue, 07 Jul 2015 12:53:51 -0700  1      {kubelet kubernetes-node-tf0f}    implicitly required container POD   started     Started with docker id 6a41280f516d
  Tue, 07 Jul 2015 12:53:51 -0700   Tue, 07 Jul 2015 12:53:51 -0700  1      {kubelet kubernetes-node-tf0f}    spec.containers{simmemleak}         created     Created with docker id 87348f12526a

In the preceding example, the Restart Count: 5 indicates that the simmemleak Container in the Pod was terminated and restarted five times.

You can call kubectl get pod with the -o go-template=... option to fetch the status of previously terminated Containers:

kubectl get pod -o go-template='{{range.status.containerStatuses}}{{"Container Name: "}}{{.name}}{{"\r\nLastState: "}}{{.lastState}}{{end}}'  simmemleak-hra99
Container Name: simmemleak
LastState: map[terminated:map[exitCode:137 reason:OOM Killed startedAt:2015-07-07T20:58:43Z finishedAt:2015-07-07T20:58:43Z containerID:docker://0e4095bba1feccdfe7ef9fb6ebffe972b4b14285d5acdec6f0d3ae8a22fad8b2]]

You can see that the Container was terminated because of reason:OOM Killed, where OOM stands for Out Of Memory.

Local ephemeral storage

FEATURE STATE: Kubernetes v1.18 beta
This feature is currently in a beta state, meaning:

  • The version names contain beta (e.g. v2beta3).
  • Code is well tested. Enabling the feature is considered safe. Enabled by default.
  • Support for the overall feature will not be dropped, though details may change.
  • The schema and/or semantics of objects may change in incompatible ways in a subsequent beta or stable release. When this happens, we will provide instructions for migrating to the next version. This may require deleting, editing, and re-creating API objects. The editing process may require some thought. This may require downtime for applications that rely on the feature.
  • Recommended for only non-business-critical uses because of potential for incompatible changes in subsequent releases. If you have multiple clusters that can be upgraded independently, you may be able to relax this restriction.
  • Please do try our beta features and give feedback on them! After they exit beta, it may not be practical for us to make more changes.

Kubernetes version 1.8 introduces a new resource, ephemeral-storage for managing local ephemeral storage. In each Kubernetes node, kubelet’s root directory (/var/lib/kubelet by default) and log directory (/var/log) are stored on the root partition of the node. This partition is also shared and consumed by Pods via emptyDir volumes, container logs, image layers and container writable layers.

This partition is “ephemeral” and applications cannot expect any performance SLAs (Disk IOPS for example) from this partition. Local ephemeral storage management only applies for the root partition; the optional partition for image layer and writable layer is out of scope.

Note: If an optional runtime partition is used, root partition will not hold any image layer or writable layers.

Requests and limits setting for local ephemeral storage

Each Container of a Pod can specify one or more of the following:

  • spec.containers[].resources.limits.ephemeral-storage
  • spec.containers[].resources.requests.ephemeral-storage

Limits and requests for ephemeral-storage are measured in bytes. You can express storage as a plain integer or as a fixed-point integer using one of these suffixes: E, P, T, G, M, K. You can also use the power-of-two equivalents: Ei, Pi, Ti, Gi, Mi, Ki. For example, the following represent roughly the same value:

128974848, 129e6, 129M, 123Mi

For example, the following Pod has two Containers. Each Container has a request of 2GiB of local ephemeral storage. Each Container has a limit of 4GiB of local ephemeral storage. Therefore, the Pod has a request of 4GiB of local ephemeral storage, and a limit of 8GiB of storage.

apiVersion: v1
kind: Pod
metadata:
  name: frontend
spec:
  containers:
  - name: db
    image: mysql
    env:
    - name: MYSQL_ROOT_PASSWORD
      value: "password"
    resources:
      requests:
        ephemeral-storage: "2Gi"
      limits:
        ephemeral-storage: "4Gi"
  - name: wp
    image: wordpress
    resources:
      requests:
        ephemeral-storage: "2Gi"
      limits:
        ephemeral-storage: "4Gi"

How Pods with ephemeral-storage requests are scheduled

When you create a Pod, the Kubernetes scheduler selects a node for the Pod to run on. Each node has a maximum amount of local ephemeral storage it can provide for Pods. For more information, see “Node Allocatable”.

The scheduler ensures that the sum of the resource requests of the scheduled Containers is less than the capacity of the node.

How Pods with ephemeral-storage limits run

For container-level isolation, if a Container’s writable layer and logs usage exceeds its storage limit, the Pod will be evicted. For pod-level isolation, if the sum of the local ephemeral storage usage from all containers and also the Pod’s emptyDir volumes exceeds the limit, the Pod will be evicted.

Monitoring ephemeral-storage consumption

When local ephemeral storage is used, it is monitored on an ongoing basis by the kubelet. The monitoring is performed by scanning each emptyDir volume, log directories, and writable layers on a periodic basis. Starting with Kubernetes 1.15, emptyDir volumes (but not log directories or writable layers) may, at the cluster operator’s option, be managed by use of project quotas. Project quotas were originally implemented in XFS, and have more recently been ported to ext4fs. Project quotas can be used for both monitoring and enforcement; as of Kubernetes 1.16, they are available as alpha functionality for monitoring only.

Quotas are faster and more accurate than directory scanning. When a directory is assigned to a project, all files created under a directory are created in that project, and the kernel merely has to keep track of how many blocks are in use by files in that project. If a file is created and deleted, but with an open file descriptor, it continues to consume space. This space will be tracked by the quota, but will not be seen by a directory scan.

Kubernetes uses project IDs starting from 1048576. The IDs in use are registered in /etc/projects and /etc/projid. If project IDs in this range are used for other purposes on the system, those project IDs must be registered in /etc/projects and /etc/projid to prevent Kubernetes from using them.

To enable use of project quotas, the cluster operator must do the following:

  • Enable the LocalStorageCapacityIsolationFSQuotaMonitoring=true feature gate in the kubelet configuration. This defaults to false in Kubernetes 1.16, so must be explicitly set to true.

  • Ensure that the root partition (or optional runtime partition) is built with project quotas enabled. All XFS filesystems support project quotas, but ext4 filesystems must be built specially.

  • Ensure that the root partition (or optional runtime partition) is mounted with project quotas enabled.

Building and mounting filesystems with project quotas enabled

XFS filesystems require no special action when building; they are automatically built with project quotas enabled.

Ext4fs filesystems must be built with quotas enabled, then they must be enabled in the filesystem:

% sudo mkfs.ext4 other_ext4fs_args... -E quotatype=prjquota /dev/block_device
% sudo tune2fs -O project -Q prjquota /dev/block_device

To mount the filesystem, both ext4fs and XFS require the prjquota option set in /etc/fstab:

/dev/block_device	/var/kubernetes_data	defaults,prjquota	0	0

Extended resources

Extended resources are fully-qualified resource names outside the kubernetes.io domain. They allow cluster operators to advertise and users to consume the non-Kubernetes-built-in resources.

There are two steps required to use Extended Resources. First, the cluster operator must advertise an Extended Resource. Second, users must request the Extended Resource in Pods.

Managing extended resources

Node-level extended resources

Node-level extended resources are tied to nodes.

Device plugin managed resources

See Device Plugin for how to advertise device plugin managed resources on each node.

Other resources

To advertise a new node-level extended resource, the cluster operator can submit a PATCH HTTP request to the API server to specify the available quantity in the status.capacity for a node in the cluster. After this operation, the node’s status.capacity will include a new resource. The status.allocatable field is updated automatically with the new resource asynchronously by the kubelet. Note that because the scheduler uses the node status.allocatable value when evaluating Pod fitness, there may be a short delay between patching the node capacity with a new resource and the first Pod that requests the resource to be scheduled on that node.

Example:

Here is an example showing how to use curl to form an HTTP request that advertises five “example.com/foo” resources on node k8s-node-1 whose master is k8s-master.

curl --header "Content-Type: application/json-patch+json" \
--request PATCH \
--data '[{"op": "add", "path": "/status/capacity/example.com~1foo", "value": "5"}]' \
http://k8s-master:8080/api/v1/nodes/k8s-node-1/status
Note: In the preceding request, ~1 is the encoding for the character / in the patch path. The operation path value in JSON-Patch is interpreted as a JSON-Pointer. For more details, see IETF RFC 6901, section 3.

Cluster-level extended resources

Cluster-level extended resources are not tied to nodes. They are usually managed by scheduler extenders, which handle the resource consumption and resource quota.

You can specify the extended resources that are handled by scheduler extenders in scheduler policy configuration.

Example:

The following configuration for a scheduler policy indicates that the cluster-level extended resource “example.com/foo” is handled by the scheduler extender.

  • The scheduler sends a Pod to the scheduler extender only if the Pod requests “example.com/foo”.
  • The ignoredByScheduler field specifies that the scheduler does not check the “example.com/foo” resource in its PodFitsResources predicate.

    {
    "kind": "Policy",
    "apiVersion": "v1",
    "extenders": [
    {
      "urlPrefix":"<extender-endpoint>",
      "bindVerb": "bind",
      "managedResources": [
        {
          "name": "example.com/foo",
          "ignoredByScheduler": true
        }
      ]
    }
    ]
    }

Consuming extended resources

Users can consume extended resources in Pod specs just like CPU and memory. The scheduler takes care of the resource accounting so that no more than the available amount is simultaneously allocated to Pods.

The API server restricts quantities of extended resources to whole numbers. Examples of valid quantities are 3, 3000m and 3Ki. Examples of invalid quantities are 0.5 and 1500m.

Note: Extended resources replace Opaque Integer Resources. Users can use any domain name prefix other than kubernetes.io which is reserved.

To consume an extended resource in a Pod, include the resource name as a key in the spec.containers[].resources.limits map in the container spec.

Note: Extended resources cannot be overcommitted, so request and limit must be equal if both are present in a container spec.

A Pod is scheduled only if all of the resource requests are satisfied, including CPU, memory and any extended resources. The Pod remains in the PENDING state as long as the resource request cannot be satisfied.

Example:

The Pod below requests 2 CPUs and 1 “example.com/foo” (an extended resource).

apiVersion: v1
kind: Pod
metadata:
  name: my-pod
spec:
  containers:
  - name: my-container
    image: myimage
    resources:
      requests:
        cpu: 2
        example.com/foo: 1
      limits:
        example.com/foo: 1

What's next

Feedback