Virtual-kubelet-cmd

From epsciwiki
Jump to navigation Jump to search

Configuring Virtual-Kubelet-Cmd

Using Shell Scripts to Start Virtual-Kubelet-Cmd

The test-run/start.sh script provides an example of how to initiate the VK. It does this by setting up specific environment variables.

#!/bin/bash
export MAIN="/workspaces/virtual-kubelet-cmd"
export VK_PATH="$MAIN/test-run/apiserver"
export VK_BIN="$MAIN/bin"

export APISERVER_CERT_LOCATION="$VK_PATH/client.crt"
export APISERVER_KEY_LOCATION="$VK_PATH/client.key"
export KUBECONFIG="$HOME/.kube/config"

export NODENAME="vk"
export VKUBELET_POD_IP="172.17.0.1"
export KUBELET_PORT="10255" 

export JIRIAF_WALLTIME="60" 
export JIRIAF_NODETYPE="cpu"
export JIRIAF_SITE="Local"

"$VK_BIN/virtual-kubelet" --nodename $NODENAME --provider mock --klog.v 3 > ./$NODENAME.log 2>&1

Environment Variables

Environment Variable Description
MAIN Main workspace directory
VK_PATH Path to the directory containing the apiserver files
VK_BIN Path to the binary files
APISERVER_CERT_LOCATION Location of the apiserver certificate
APISERVER_KEY_LOCATION Location of the apiserver key
KUBECONFIG Points to the location of the Kubernetes configuration file, which is used to connect to the Kubernetes API server. By default, it’s located at $HOME/.kube/config.
NODENAME The name of the node in the Kubernetes cluster.
VKUBELET_POD_IP The IP address of the VK that metrics server talks to. If the metrics server is running in a Docker container and VK is running on the same host, this is typically the IP address of the docker0 interface.
KUBELET_PORT The port on which the Kubelet service is running. The default port for Kubelet is 10250. This is for the metrics server and should be unique for each node.
JIRIAF_WALLTIME Sets a limit on the total time that a node can run. It should be a multiple of 60 and is measured in seconds. If it’s set to 0, there is no time limit.
JIRIAF_NODETYPE Specifies the type of node that the job will run on. This is just for labeling purposes and doesn’t affect the actual job.
JIRIAF_SITE Used to specify the site where the job will run. This is just for labeling purposes and doesn’t affect the actual job.

Running Pods on Virtual-Kubelet-Cmd Nodes

Pods, along with their associated containers, can be deployed on Virtual-Kubelet-Cmd (VK) nodes. The following table contrasts the capabilities of a VK node with those of a standard kubelet:

Feature Virtual-Kubelet-CMD Regular Kubelet
Container Executes as a series of Linux processes Runs as a Docker container
Image Defined as a shell script Defined as a Docker container image

Enhanced Features for Script Storage and Execution in Pods

Feature Description
configMap/ secret These are used as volume types for storing scripts during the pod launch process
volumes This feature is implemented within the pod to manage the use of configMap and secret
volumeMounts This feature is used to relocate scripts to the specified mountPath. The mountPath is defined as a relative path. Its root is structured as $HOME/$podName/containers/$containerName
command and args These are utilized to execute scripts
env This feature is supported for passing environment variables to the scripts running within a container
image The image corresponds to a volumeMount in the container and shares the same name

Process Group Management in Containers Using pgid files

The pgid file is a feature used to manage the process group of a shell script running within a container. Each container has a unique pgid file to ensure process management. The pgid can be found at the following location: $HOME/$podName/containers/$containerName/pgid.

Lifecycle of containers and pods

Description of container states

The following tables provide a description of the container states and their associated methods.

CreatePod method called, the following states are used:

UID Stage State StartAt FinishedAt ExitCode Reason Message IsError Description
create-cont-readDefaultVolDirError CreatePod Terminated Start of pod Now 1 readDefaultVolDirError fmt.Sprintf(“Failed to read default volume directory %s; error: %v”, defaultVolumeDirectory, err) Y Scan the default volume directory for files
create-cont-copyFileError CreatePod Terminated Start of pod Now 1 copyFileError fmt.Sprintf(“Failed to copy file %s to %s; error: %v”, path.Join(defaultVolumeDirectory, file.Name()), path.Join(mountDirectory, file.Name()), err) Y Copy the file to the mount directory
create–cont-cmdStartError CreatePod Terminated Start of pod Now 1 cmdStartError cmd.Start() failed Y The command is initiated with cmd.Start().
create-cont-getPgidError CreatePod Terminated Start of pod Now 1 getPgidError failed to get pgid Y The process group id is retrieved using syscall.Getpgid(cmd.Process.Pid).
create-cont-createStdoutFileError CreatePod Terminated Start of pod Now 1 createStdoutFileError failed to create stdout file Y The stdout file is created using os.Create(path.Join(stdoutPath, “stdout”)).
create-cont-createStderrFileError CreatePod Terminated Start of pod Now 1 createStderrFileError failed to create stderr file Y The stderr file is created using os.Create(path.Join(stdoutPath, “stderr”)).
create-cont-cmdWaitError CreatePod Terminated Start of pod Now 1 cmdWaitError cmd.Wait() failed Y A goroutine is initiated to wait for the command to complete with cmd.Wait()
create-cont-writePgidError CreatePod Terminated Start of pod Now 1 writePgidError fmt.Sprintf(“failed to write pgid to file %s; error: %v”, pgidFile, err) Y Write the process group ID to a file
create-cont-containerStarted CreatePod Running Start of pod nan nan nan nan N No error; init container state

GetPods method called, the following states are used:

UID Stage State StartAt FinishedAt ExitCode Reason Message IsError Description
get-cont-create GetPods Terminated Prev Prev 1 from those with ExitCode 1 from those with ExitCode 1 Y Container failed to start
get-cont-getPidsError GetPods Terminated Prev Prev 2 getPidsError Error getting pids Y Failed to get system PIDs
get-cont-getStderrFileInfoError GetPods Terminated Prev Prev 2 getStderrFileInfoError Error getting stderr file info Y Failed to get info about stderr file of container
get-cont-stderrNotEmpty GetPods Terminated Prev Prev 3 stderrNotEmpty The stderr file is not empty. N All processes are in Z. Stderr is not empty. Container is done with errors.
get-cont-completed GetPods Terminated Prev Prev 0 completed Remaining processes are zombies N All processes are in Z. Stderr is empty. Container is done without errors.
get-cont-running GetPods Running Prev nan nan nan nan N Not all processes are in Z. Container is running.

Field Descriptions

Field Description
UID A unique identifier for container state.
Stage Method that container state is associated with.
State State of container.
StartAt Get time container started. Prev means time of previous state. Now means current time.
FinishedAt Get time container finished. Prev means time of previous state. Now means current time.
ExitCode Exit code of container.
Reason Reason for container’s state. 1: Errors when createPod is called. 2: Errors when getPods is called. 3: stderr file is not empty. 0: Container is completed.
Message Message associated with container’s state.
IsError Boolean value that indicates whether container state is an error.
Description Description of container’s state.

Note: The method GetPods is called every 5 seconds to check the state of the container. The method CreatePod is called when the pod is created.

The flowchart for creating and monitoring lifecycle of the containers in a pod

The following points describe the process of creating and monitoring containers and pods in the virtual-kubelet-cmd:

  1. The 🔄 all containers block indicates a loop that iterates over all containers in the pod.
  2. The blocks in blue represent the process of creating container state instances.
  3. The blocks in purple illustrate the process of creating and updating the pod status instances. This is based on the created container states and the pod phase.
  4. The blocks in red depict the process of redirecting flows under various conditions.

Note: The Unique Identifier (UID) assigned to each container state is derived from the tables provided in the preceding section.

Flowchart1
Flowchart2
Flowchart3

Procedure to Deploy a Pod Executing a Shell Script

  • The image field is defined as a shell script. This means that the image field corresponds to the name of volumeMounts.
  • Use a configMap to store the shell script.
  • Use volumeMounts to mount the script into the container.
  • The command and args fields are used to execute the script.

Here’s an example of how to create a pod that runs a shell script:

kind: ConfigMap
apiVersion: v1
metadata:
  name: direct-stress
data:
  stress.sh: |
    #!/bin/bash
    stress --timeout $1 --cpu $2 # test memory
---
apiVersion: v1
kind: Pod
metadata:
  name: p1
  labels:
    app: new-test-pod
spec:
  containers:
    - name: c1
      image: direct-stress # this name should be the same as the name in the volumeMounts
      command: ["bash"]
      args: ["300", "2"] # the first argument is the timeout, and the second argument is the cpu number as defined in the stress.sh
      volumeMounts:
        - name: direct-stress
          mountPath: stress/job1 # the root path of the mountPath is $HOME/p1/containers/c1
  volumes:
    - name: direct-stress 
      configMap:
        name: direct-stress

Running Pods on Virtual Kubelet Nodes

To schedule pods on Virtual Kubelet (VK) nodes, it’s necessary to include specific labels in both nodeSelector and tolerations.

nodeSelector:
    kubernetes.io/role: agent
tolerations:
  - key: "virtual-kubelet.io/provider"
    value: "mock"
    effect: "NoSchedule"

Setting Affinity for Pods on Virtual Kubelet Nodes

  • The affinity of pods for Virtual Kubelet (VK) nodes is determined by three labels: jiriaf.nodetype, jiriaf.site, and jiriaf.alivetime. These labels correspond to the environment variables JIRIAF_NODETYPE, JIRIAF_SITE, and JIRIAF_WALLTIME in the start.sh script.
  • Note that if JIRIAF_WALLTIME is set to 0, the jiriaf.alivetime label will not be defined, and therefore, the affinity will not be applied.
  • JRM's status changes from Ready to NotReady after jiriaf.alivetime turns zero. Notice that the process of JRM will NOT be killed after the alivetime limit is reached.
  • To add more labels to the VK nodes, modify ConfigureNode in internal/provider/mock/mock.go.
  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchExpressions:
          - key: jiriaf.nodetype
            operator: In
            values:
            - "cpu"
          - key: jiriaf.site
            operator: In
            values:
            - "mylin"
          - key: jiriaf.alivetime # if JIRIAF_WALLTIME is set to 0, this label should not be defined.
            operator: Gt
            values:
            - "10"

Metrics Server Deployment

The Metrics Server is a tool that collects and provides resource usage data for nodes and pods within a Kubernetes cluster. The necessary deployment configuration is located in the metrics-server/components.yaml file.

To deploy the Metrics Server, execute the following command:

kubectl apply -f metrics-server/components.yaml

Note: The flag --kubelet-use-node-status-port is added to the metrics-server container in the metrics-server deployment to allow the Metrics Server to communicate with the Virtual Kubelet nodes.

Essential Scripts

The primary control mechanisms for the Virtual Kubelet (VK) are contained within the following files: - internal/provider/mock/mock.go - internal/provider/mock/command.go - internal/provider/mock/volume.go

References