search

kube-HPC / hkube

🐟 High Performance Computing over Kubernetes - Core Repo 🎣
http://hkube.io
MIT License
303 stars 20 forks source link
algorithm cluster hkube kubernetes pipeline

readme

HKube

HKube is a cloud-native open source framework to run distributed pipeline of algorithms built on Kubernetes.

HKube optimally utilizing pipeline's resources, based on user priorities and heuristics.

Features

User Guide

Installation

Dependencies

HKube runs on top of Kubernetes so in order to run HKube we have to install it's prerequisites.

Helm

  1. Add the HKube Helm repository to helm:

    helm repo add hkube http://hkube.io/helm/

  2. Configure a docker registry for builds
    Create a values.yaml file for custom helm values

    build_secret:
# pull secret is only needed if docker hub is not accessible
pull:
registry: ''
namespace: ''
username: ''
password: ''
# enter your docker hub / other registry credentials
push:
registry: '' # can be left empty for docker hub
namespace: '' # registry namespace - usually your username
username: ''
password: ''

  3. Install HKube chart

    helm install hkube/hkube  -f ./values.yaml --name my-release

This command installs HKube in a minimal configuration for development. Check production-deployment.

APIs

There are three ways to communicate with HKube: Dashboard, REST API and CLI.

UI Dashboard

Dashboard is a web-based HKube user interface. Dashboard supports every functionality HKube has to offer.

ui

REST API

HKube exposes it's functionality with REST API.

CLI

hkubectl is HKube command line tool.

hkubectl [type] [command] [name]

# More information
hkubectl --help

Download hkubectl latest version.

curl -Lo hkubectl https://github.com/kube-HPC/hkubectl/releases/latest/download/hkubectl-linux \
&& chmod +x hkubectl \
&& sudo mv hkubectl /usr/local/bin/

For mac replace with hkubectl-macos
For Windows download hkubectl-win.exe

Config hkubectl with your running Kubernetes.

# Config
hkubectl config set endpoint ${KUBERNETES-MASTER-IP}

hkubectl config set rejectUnauthorized false

Make sure kubectl is configured to your cluster.

HKube requires that certain pods will run in privileged security permissions, consult your Kubernetes installation to see how it's done.

API Usage Example

The Problem

We want to solve the next problem with given input and a desired output:

For example: N=5, k=2 will result:

Solution

We will solve the problem by running a distributed pipeline of three algorithms: Range, Multiply and Reduce.

Range Algorithm

Creates an array of length N.

 N = 5
 5 -> [1,2,3,4,5]

Multiply Algorithm

Multiples the received data from Range Algorithm by k.

k = 2
[1,2,3,4,5] * (2) -> [2,4,6,8,10]

Reduce Algorithm

The algorithm will wait until all the instances of the Multiply Algorithm will finish then it will summarize the received data together .

[2,4,6,8,10] -> 30

Building a Pipeline

We will implement the algorithms using various languages and construct a pipeline from them using HKube.

PipelineExample

Pipeline Descriptor

The pipeline descriptor is a JSON object which describes and defines the links between the nodes by defining the dependencies between them.

{
  "name": "numbers",
  "nodes": [
    {
      "nodeName": "Range",
      "algorithmName": "range",
      "input": ["@flowInput.data"]
    },
    {
      "nodeName": "Multiply",
      "algorithmName": "multiply",
      "input": ["#@Range", "@flowInput.mul"]
    },
    {
      "nodeName": "Reduce",
      "algorithmName": "reduce",
      "input": ["@Multiply"]
    }
  ],
  "flowInput": {
    "data": 5,
    "mul": 2
  }
}

Note the flowInput: data = N = 5, mul = k = 2

Node dependencies

HKube allows special signs in nodes input for defining the pipeline execution flow.

In our case we used:

(@)  —  References input parameters for the algorithm.

(#)  —  Execute nodes in parallel and reduce the results into single node.

(#@) — By combining # and @ we can create a batch processing on node results.

JSON

JSON Breakdown

We created a pipeline name numbers.

    "name":"numbers"

The pipeline is defined by three nodes.

"nodes":[
    {
            "nodeName":"Range",
            "algorithmName":"range",
            "input":["@flowInput.data"]
        },
        {
            "nodeName":"Multiply",
            "algorithmName":"multiply",
            "input":["#@Range","@flowInput.mul"]
        },
        {
            "nodeName":"Reduce",
            "algorithmName":"reduce",
            "input":["@Multiply"]
        },
    ]

In HKube, the linkage between the nodes is done by defining the algorithm inputs. Multiply will be run after Range algorithm because of the input dependency between them.

Keep in mind that HKube will transport the results between the nodes automatically for doing it HKube currently support two different types of transportation layers object storage and files system.

Group 4 (3)

The flowInput is the place to define the Pipeline inputs:

"flowInput":{
    "data":5,
    "mul":2
}

In our case we used Numeric Type but it can be any JSON type (Object, String etc).

Advance Options

There are more features that can be defined from the descriptor file.

"webhooks": {
    "progress": "http://my-url-to-progress",
      "result": "http://my-url-to-result"
    },
  "priority": 3,
  "triggers":
      {
      "pipelines":[],
        "cron":{}
      }
  "options":{
      "batchTolerance": 80,
      "concurrentPipelines": 2,
      "ttl": 3600,
      "progressVerbosityLevel":"info"
  }

Algorithm

The pipeline is built from algorithms which containerized with docker.

There are two ways to integrate your algorithm into HKube:

Implementing the Algorithms

We will create the algorithms to solve the problem, HKube currently support two languages for auto build Python and JavaScript.

Important notes:

  • Installing dependencies During the container build, HKube will search for the requirement.txt file and will try to install the packages from the pip package manager.
  • Advanced Operations HKube can build the algorithm only by implementing start function but for advanced operation such as one time initiation and gracefully stopping you have to implement two other functions init and stop.
Range (Python)
def start(args):
    print('algorithm: range start')
    input = args['input'][0]
    array = list(range(input))
    return array

The start method calls with the args parameter, the inputs to the algorithm will appear in the input property.

The input property is an array, so you would like to take the first argument ("input":["@flowInput.data"] as you can see we placed data as the first argument)

Multiply (Python)
def start(args):
    print('algorithm: multiply start')
    input = args['input'][0]
    mul = args['input'][1]
    return input * mul

We sent two parameters "input":["#@Range","@flowInput.mul"], the first one is the output from Range that sent an array of numbers, but because we using batch sign (#) each multiply algorithm will get one item from the array, the second parameter we passing is the mul parameter from the flowInput object.

Reduce (Javascript)
module.exports.start = args => {
  console.log('algorithm: reduce start');
  const input = args.input[0];
  return input.reduce((acc, cur) => acc + cur);
};

We placed ["@Multiply"] in the input parameter, HKube will collect all the data from the multiply algorithm and will sent it as an array in the first input parameter.

Integrate Algorithms

After we created the algorithms, we will integrate them with the CLI.

Can be done also through the Dashboard.

Create a yaml (or JSON) that defines the algorithm:

# range.yml
name: range
env: python # can be python or javascript
resources:
  cpu: 0.5
  gpu: 1 # if not needed just remove it from the file
  mem: 512Mi

code:
  path: /path-to-algorithm/range.tar.gz
  entryPoint: main.py

Add it with the CLI:

hkubectl algorithm apply --f range.yml

Keep in mind we have to do it for each one of the algorithms.

Integrate Pipeline

Create a yaml (or JSON) that defines the pipeline:

# number.yml
name: numbers
nodes:
  - nodeName: Range
    algorithmName: range
    input:
      - '@flowInput.data'
  - nodeName: Multiply
    algorithmName: multiply
    input:
      - '#@Range'
      - '@flowInput.mul'
  - nodeName: Reduce
    algorithmName: reduce
    input:
      - '@Multiply'
flowInput:
  data: 5
  mul: 2

Raw - Ad-hoc pipeline running

For running our pipeline as raw-data:

hkubectl exec raw --f numbers.yml

Stored - Storing the pipeline descriptor for next running

First we store the pipeline:

hkubectl pipeline store --f numbers.yml

Then you can execute it (if flowInput available)

# flowInput stored
hkubectl exec stored numbers

For executing the pipeline with other input, create yaml (or JSON) file with flowInput key:

# otherFlowInput.yml
flowInput:
  data: 500
  mul: 200

Then you can executed it by pipeline name:

# Executes pipeline "numbers" with data=500, mul=200
hkubectl exec stored numbers --f otherFlowInput.yml

Monitor Pipeline Results

As a result of executing pipeline, HKube will return a jobId.

# Job ID returned after execution.
result:
  jobId: numbers:a56c97cb-5d62-4990-817c-04a8b0448b7c.numbers

This is a unique identifier helps to query this specific pipeline execution.