Memory Footprint Test

Why measure memory footprint?

  • This platform is designed for a light-weight edge computing deployment, capable of running on devices with few resources (for example, 256MB RAM)
  • It is important to know when deploying many pods that it showcases as little memory footprint as possible

KPI’s measured

  • %CPU
  • %Memory
  • Resident Set Size (RSS)

How to test

After deployment and provisioning of KubeEdge cloud and edge components in 2 VM’s (supported and tested over Ubuntu 16.04) respectively, start deploying pods from 0 to 100 in steps of 5. Keep capturing above KPI’s using standard linux ps commands, after each step.

Test setup

KubeEdge Test Setup

Fig 1: KubeEdge Test Setup

Creating a setup

Requirements

  • Host machine’s or VM’s resource requirements can mirror the edge device of your choice
  • Resources used for above setup are 4 CPU, 8GB RAM and 200 GB Disk space. OS is Ubuntu 16.04.
  • Docker image used to deploy the pods in edge, needs to be created. The steps are:
    1. Go to github.com/kubeedge/kubeedge/edge/hack/memfootprint-test/
    2. Using the Dockerfile available here and create docker image (perftestimg:v1).
    3. Execute the docker command sudo docker build --tag "perftestimg:v1" ., to get the image.

Installation

  • For KubeEdge Cloud and Edge:

    Please follow steps mentioned in KubeEdge README.md

  • For docker image:

  • Deploy docker registry to either edge on any VM or host which is reachable to edge. Follow the steps mentioned here: https://docs.docker.com/registry/deploying/

  • Create perftestimg:v1 docker image on the above mentioned host

  • Then push this image to docker registry using docker tag and docker push commands (Refer: Same docker registry url mentioned above) [Use this image’s metadata in pod deployment yaml]

Steps

  1. Check edge node is connected to cloud. In cloud console/terminal, execute the below command
  1. root@ubuntu:~/edge/pod_yamls# kubectl get nodes
  2. NAME STATUS ROLES AGE VERSION
  3. 192.168.20.31 Unknown <none> 11s
  4. ubuntu NotReady master 5m22s v1.14.0
  1. On cloud, modify deployment yaml (github.com/kubeedge/kubeedge/edge/hack/memfootprint-test/perftestimg.yaml), set the image name and set spec.replica as 5
  2. Execute sudo kubectl create -f ./perftestimg.yaml to deploy the first of 5 pods in edge node
  3. Execute sudo kubectl get pods | grep Running | wc to check if all the pods come to Running state. Once all pods come to running state, go to edge VM
  4. On Edge console, execute ps -aux | grep edgecore. The output shall be something like:
  1. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND
  2. root 102452 1.0 0.5 871704 42784 pts/0 Sl+ 17:56 0:00 ./edgecore
  3. root 102779 0.0 0.0 14224 936 pts/2 S+ 17:56 0:00 grep --color=auto edge
  1. Collect %CPU, %MEM and RSS from respective columns and record
  2. Repeat step 2 and this time increase the replica by 5
  3. This time execute sudo kubectl apply -f <PATH>/perftestimg.yaml
  4. Repeat steps from 4 to 6.
  5. Now repeat steps from 7 to 9, till the replica count reaches 100

最近更新于 Mar 16, 2021