You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

5.9 KiB

Raw Blame History

Resources, and how it impacts the scheduling

Introduction

When you declare a pod you can declare what resources it will use. Those resources are how much CPU and RAM your pods will use.

Specifying those will help Kubernetes to schedule more efficently your pods on the available nodes.

For each resource you can define the limits and the requests.

Resources definition

The CPU resource is measured in a number of CPU the pod will use for a given amount of time. It cannot be inferior to 0. Specifying 0.5 (or 500m, which means 500 millicpu), will give half of a CPU to your pod.

The RAM resource is measured in the number of bytes of RAM the pod will use. Specifying 123Mi (or 128974848, which means 128974848 bytes), will give that much RAM to your pod.

`requests` vs `limits`

Kubernetes lets you configure the requests and the limits for each resource.

They are put on at the container level:

apiVersion: v1
kind: Pod
metadata:
  name: my-application
spec:
  containers:
  - name: my-application
    image: my-application:my-tag
    resources:
      requests:
        memory: "64Mi"
        cpu: "250m"
      limits:
        memory: "128Mi"
        cpu: "500m"

Let's see in details each of those.

`requests`

The requests is the number that will help Kubernetes schedule your pod on a node where the resources are available.

Let's take an example. You have 3 nodes:

node 1 has 100m CPU available, and 1Gi RAM available
node 2 has 1000m CPU available, and 100Mi RAM available

You start a pod with only a CPU request of 10m. It can be scheduled on any node, so Kubernetes will take either of them.
You start a pod with only a CPU request of 500m. It can be scheduled only on node 2. Node 1 has only 100m CPU available.
You start a pod with a CPU request of 500m and a RAM request of 500Mi. Kubernetes will not be able to schedule your pod as no node has both requests at the same time. The pod will be in the state Unschedulable.

`limits`

The limits is the maximum utilization of a resource Kubernetes will allow for a given pod. If your pod goes above this limit it will be restarted.

Note on resources

On any computer, you can overuse CPU, but you cannot overuse RAM.

If running all your programs require more processing power than you have available your computer will just be slow.

On the other hand, if running all your program require more RAM than you have available, your computer will kill randomly processes that ask for too much RAM (at least on linux).

In Kubernetes it's a bit the same. You can ask for limits that are higher than the node CPU, but not the RAM.

Furthermore, you can ask for less than a CPU. But what does that mean? When you ask for 100 millicpu, Kubernetes will give you 100 milliseconds of CPU during a second of time. If your container wants to use more it'll be throttled during the remaining 900 milliseconds of the second.

Good practices

Only the requests is taken into account into the scheduling. So it's possible for a Kubernetes node to go on "overcommit", to have the sum of resources used above the physical resource limitation of a node. In this case Kubernetes will look for pods to terminate. Its algorithm is to look at pods that are above what they requested and terminates them. If a pod has no requests, so they are using more than what they requested, they will be the first to be terminated. Other candidates are the pods that have gone over their request but are still under their limit.

Unless your applications are designed to use multiple cores, it is usually a best practice to keep the CPU request at "1" or below.

Most of the time you could define request == limit. But careful as your pod will be terminated if it goes above the limit.

Exercices

First, you need to active the metric-server on minikube:

minikube addons enable metrics-server

We'll use the stress tool to simulate resource utilization. The used options will be:

--vm-bytes $MEM: Amount of RAM used by stress, for example "150M" for 150 megabytes of RAM allocated.
--cpu $CPU: Number of CPUs used by stress, an integer.

We will use other options, but ignore those. You can find more information on thos on the stress documentation.

CPU

First let's try to declare a CPU requests that is too high compared to what your cluster has available. Review and apply the file 01-cpu-requests.yml. Look at the pod created. What can you see?

Second, let's try to have a pod consuming more CPU resources than allowed. Review and apply the file 02-cpu-limits.yml. Look at the pod created. What can you see? Look at the logs of the pods:

kubectl logs -f cpu-limits

Note, you can see the usage of the pod with kubectl:

kubectl top pod cpu-limits

If you see the error Error from server (NotFound): podmetrics.metrics.k8s.io "default/cpu-limits" not found, retry in a few seconds. It only means the metric-server didn't get any metrics yet.

RAM

First let's try to declare a memory requests that is too high compared to what your cluster has available. Review and apply the file 03-ram-requests.yml. Look at the pod created. What can you see?

Second, let's try to have a pod consuming more RAM resources than allowed. Review and apply the file 04-ram-limits.yml. Look at the pod created. What can you see?

Clean up

kubectl delete service,deployment,pod --all

5.9 KiB Raw Blame History