MarioUhrik / k8s-cpu-contention-test

A scripted experiment to observe the effects of CPU limits/requests between two CPU-hungry pods

Intro

Many iffy opinions about how CPU requests & limits in Kubernetes work circulate the Internet. For example, some people believe that if a pod requests 300mCPU, it is guaranteed to have the 300mCPU available during runtime. This project is a simple experiment aimed at disproving that.

Experiment methodology

• A minikube cluster with 1 CPU is spawned.
• We create two simultaneous containers, both designed to try to eat 100% CPU, to compete with each other for CPU.
• The containers are deployed throughout several scenarios in various combinations of CPU requests/limits.
• At the end of each scenario, the CPU usage of both pods is examined and compared.

The entire experiment is automated within run.sh. Simply run it, wait, and observe the results.

Requirements:

• bash
• minikube
• kubectl

Expected results & commentary

-----Initiating test scenarios-----
-----Scenario 1-----
deployment.apps/no-request-no-limit-1 created
deployment.apps/no-request-no-limit-2 created
NAME                                     CPU(cores)   MEMORY(bytes)   
no-request-no-limit-1-769c86549d-v5rnl   364m         1Mi             
no-request-no-limit-2-79cbf77884-w69rf   393m         1Mi             
deployment.apps "no-request-no-limit-1" deleted
deployment.apps "no-request-no-limit-2" deleted
-----Scenario 2-----
deployment.apps/750-request-no-limit-1 created
deployment.apps/no-request-no-limit-2 created
NAME                                      CPU(cores)   MEMORY(bytes)   
750-request-no-limit-1-86c5947c45-hpwxl   405m         1Mi             
no-request-no-limit-2-79cbf77884-szncj    295m         1Mi             
deployment.apps "750-request-no-limit-1" deleted
deployment.apps "no-request-no-limit-2" deleted
-----Scenario 3-----
deployment.apps/750-request-no-limit-1 created
deployment.apps/100-request-no-limit-2 created
NAME                                      CPU(cores)   MEMORY(bytes)   
100-request-no-limit-2-5575b7cdf9-ncd4g   307m         1Mi             
750-request-no-limit-1-86c5947c45-8jrww   365m         1Mi             
deployment.apps "750-request-no-limit-1" deleted
deployment.apps "100-request-no-limit-2" deleted
-----Scenario 4-----
deployment.apps/750-request-no-limit-1 created
deployment.apps/100-request-100-limit-2 created
NAME                                       CPU(cores)   MEMORY(bytes)   
100-request-100-limit-2-7cf7bfd956-x5qlp   100m         1Mi             
750-request-no-limit-1-86c5947c45-tkxvn    780m         1Mi             
deployment.apps "750-request-no-limit-1" deleted
deployment.apps "100-request-100-limit-2" deleted
-----Scenario 5-----
deployment.apps/750-request-750-limit-1 created
deployment.apps/100-request-100-limit-2 created
NAME                                       CPU(cores)   MEMORY(bytes)   
100-request-100-limit-2-7cf7bfd956-2v2hz   100m         1Mi             
750-request-750-limit-1-8b94b7d8b-l7vg6    733m         1Mi             
deployment.apps "750-request-750-limit-1" deleted
deployment.apps "100-request-100-limit-2" deleted

• Scenario 1 results imply that with no CPU requests/limits in place, both pods get approximately the same amount of CPU
• Scenario 2 results imply that both pods get approximately the same amount of CPU, despite one requesting more
• Scenario 3 results imply that both pods get approximately the same amount of CPU, despite one requesting more, and the other requesting less
• Scenario 4 results imply that limiting the CPU usage of the pod with the lower request ensures that the pod with the higher request gets what it requested