Quickly start YARN cluster on EC2
Switch branches/tags
Nothing to show
Clone or download
Pull request Compare This branch is 3 commits ahead, 4 commits behind tqchen:master.
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
data
.gitignore
LICENSE
README.md
bootstrap.py
ec2_util.py
yarn-ec2
yarn_ec2.py

README.md

YARN EC2

This is a script to help you quickly setup a dynamic YARN cluster on EC2. It adapts a cloud workflow to use S3 for distributed data storage and YARN for computing.

Features

  • Dynamically add or remove slave nodes from the cluster.
  • Customized installation of packages.

How to Use

To start a cluster, the script follows two steps: (1) start master machine; (2) add slaves. This two step procedure allows you to add and remove slaves on the fly.

  • Start your master machine
    • ./yarn-ec2 -k mykey -i mypem.pem launch cluster-name
  • Add slaves to the cluster
    • ./yarn-ec2 -k mykey -i mypem.pem -s nslave addslave cluster-name
  • Alternatively, you can add spot instance to the cluster
    • ./yarn-ec2 -k mykey -i mypem.pem -s nslave addspot cluster-name
    • On demand price is used by default, you can change it by --spot-price option.
  • Both addslave and addspot will send request to EC2 and may not be fullfilled immediately
    • They will connect to the master node after one bootstrap (which takes around 1 minimute).
    • You can browse the yarn resource manager for the status of the cluster.
  • Shutdown the machines manually in ec2 panel

Distributed Storage

Because the cluster is dynamic, all the nodes are only used as computing nodes. HDFS is only started on the master machine for temp code transfer. Normally S3 is used instead for distributed storage.

Customize Installation

You can modify custom_master_install and custom_all_nodes_install in bootstrap.py to add the packages you like to install on each machine.

Restart Master Machine

In case you stopped the master and restart it on the EC2. There is no need to do the launch step again. Instead, log into the master machine, and run startup.sh on the home folder. After the startup is finished, you can continue with the steps of adding slaves.

Acknowledgement

Part of yarn-ec2 is adopted from spark-ec2 script.

Note on Implementation

Most existing cluster launch script follows a start and deploy way.

  • First start all the nodes, copy the master's credentials to the slaves.
  • Deploy the slaves by using ssh or pssh command from master.

These scripts requires master to be aware of the slaves and are hard to dynamically add or remove nodes. yarn-ec2 uses another way, where the master does not need to be aware of slaves beforehand.

  • First start the master and listen requests from slaves.
  • When a slave get started, it runs the bootstrap script, install the dependencies and report to master.
    • The YARN will then dynamically add the slave to the cluster.
  • When a slave get removed, the master will detect the event with cluster health check, and remove it from the cluster.