CMMAi/KubeflowSetupGuide

Kubeflow Setup Guide

Try it on Brandy Test Server

Single Node Deployment

Linux

Prerequisites

• Ubuntu 18 machine with
  • min 8 cores,
  • 16GB RAM,
  • 250GB storage
• Root privileges on a Ubuntu machine

Kubernetes

• Docker CE

  apt-get update
  apt-get install -y apt-transport-https ca-certificates curl software-properties-common
  curl -fsSL https://download.docker.com/linux/ubuntu/gpg | apt-key add -
  add-apt-repository \
     "deb [arch=amd64] https://download.docker.com/linux/ubuntu \
     $(lsb_release -cs) \
     stable"
  apt-get update
  apt-get install docker-ce docker-ce-cli containerd.io
  

• kubectl

  curl -LO https://storage.googleapis.com/kubernetes-release/release/v1.15.0/bin/linux/amd64/kubectl
  chmod +x ./kubectl
  mv ./kubectl /usr/local/bin/kubectl
  

• Minikube

  curl -Lo minikube https://storage.googleapis.com/minikube/releases/v1.2.0/minikube-linux-amd64
  chmod +x minikube
  cp minikube /usr/local/bin/
  rm minikube
  

  • Start Minikube

  minikube start --vm-driver=none --cpus 6 --memory 12288 --disk-size=120g --extra-config=apiserver.authorization-mode=RBAC --extra-config=kubelet.resolv-conf=/run/systemd/resolve/resolv.conf --extra-config kubeadm.ignore-preflight-errors=SystemVerification
  

Installation of Kubeflow

1. Download the kfctl here
2. Unpack the tar ball: tar -xvf kfctl_v1.0.2_<platform>.tar.gz
3. Edit your .bashrc or .zshrc or any others shell confige file like this:

export PATH=$PATH:<path to where kfctl was unpacked>
export KF_NAME=<your choice of name for the Kubeflow deployment>
export BASE_DIR=<path to a base directory>
export KF_DIR=${BASE_DIR}/${KF_NAME}
export CONFIG_URI="https://raw.githubusercontent.com/kubeflow/manifests/v1.0-branch/kfdef/kfctl_k8s_istio.v1.0.2.yaml"

mkdir -p ${KF_DIR}
cd ${KF_DIR}
kfctl apply -V -f ${CONFIG_URI}

Launch of Kubeflow central dashboard

• Set the environment variables in your .bashrc or .zshrc or any others shell confige file like this:

export INGRESS_HOST=$(minikube ip)
export INGRESS_PORT=$(kubectl -n istio-system get service istio-ingressgateway -o jsonpath='{.spec.ports[?(@.name=="http2")].nodePort}')

• then you could get the deployment url by echo http://$INGRESS_HOST:$INGRESS_PORT

For more detail, please checkout https://www.kubeflow.org/docs/started/workstation/minikube-linux/

MacOS

Docker, kubectl, Minikube

brew cask install docker
brew install minikube

then

minikube start --cpus 6 --memory 12288 --disk-size=120g --extra-config="apiserver.authorization-mode=Node,RBAC" --extra-config=kubelet.resolv-conf=/run/systemd/resolve/resolv.conf --extra-config kubeadm.ignore-preflight-errors=SystemVerification

• --extra-config="apiserver.authorization-mode=Node,RBAC", please refer to this issue

Others are the same as Linux part

My Configuration File Setup

# kubeflow
export PATH=$PATH:$HOME/.kubeflow/
export KF_NAME=rainforest
export BASE_DIR=$HOME/.kubeflow
export KF_DIR=$BASE_DIR/$KF_NAME
export CONFIG_URI="https://raw.githubusercontent.com/kubeflow/manifests/v1.0-branch/kfdef/kfctl_k8s_istio.v1.0.2.yaml"

export INGRESS_HOST=$(minikube ip)
export INGRESS_PORT=$(kubectl -n istio-system get service istio-ingressgateway -o jsonpath='{.spec.ports[?(@.name=="http2")].nodePort}')

• For more detail about kubernetes control, please refer to Minikube Website

Zinfandel 可行性評估

Single Node

• Kubeflow appliance
  • MiniKF by Vagrant -> 跟 vmware 結合
  • MicroK8s -> 部署極為方便，彈性較低
  • Minikube -> Single Node Kubernetes 彈性大

Kubernetes Installation -> Server 需支援 Kubernetes，能發揮 Containerized 帶來的所有優勢

Kubeflow Components

• Jupyter Server
  • GUI 分配硬體支援（類似 CoLab）
  • 支援 CPU, GPU 版 TF 1.5, 2.0
    • Pytorch 支援待確認，但套件基本可自行安裝
• Katib for Hyperparameter Tuning
• Pipelines -> 感覺很酷炫，訓練過程可視化，修改方便！？
  • Display Results like TensorBoard
• KFServing
• Frameworks for Training