Skip to content

Latest commit

 

History

History
88 lines (62 loc) · 2.13 KB

README.md

File metadata and controls

88 lines (62 loc) · 2.13 KB
Raw

[WIP] Deploying a dask cluster on EC2 spot instances

This demonstrates how to set up a dask.distributed cluster on EC2 spot instances.

Leveraging existing devops tools like Terraform and Docker, we aim to provide easy to deploy and highly customizable dask cluster configurations while keeping scripts to a minimum.

For now, the recipes provided in this repository remains highly experimental, untested and not recommended for use in any settings. For a more thorough and mature solution see: dask/dask-ec2, an easy to use command-line tool to create dask clusters on EC2, part of the official dask project.

Prerequisites

  • Terraform and Packer binaries in PATH
  • AWS credentials set in ~/.aws/credentials

Quickstart

Create the base AMI with packer (install docker and build container):

packer build images/dask_base_cpu.json

Note: add argument --force to re-create an existing AMI.

Generate an ssh key to access dask-scheduler and worker nodes:

mkdir -p .ssh
ssh-keygen -t rsa -b 4096 -C "dask_node" -f .ssh/dask_node_key -P ""

Create the cluster with terraform:

terraform plan
terraform apply

Set up port redirection to access the dask scheduler:

ssh -N -i .ssh/dask_node_key \
    -L 8786:localhost:8786 \
    -L 9786:localhost:9786 \
    -L 8787:localhost:8787 \
    ubuntu@[scheduler.public_ip]

Run dask jobs:

from dask.distributed import Client

def square(x):
    return x ** 2

client = Client('127.0.0.1:8786')

# TODO: add a substantive example of dask job.
squares = client.map(square, range(100))

print(client.gather(squares))

Check the scheduler UI: localhost:8787/status

Release AWS resources when all jobs are finished:

terraform destroy

TODO and ideas

  • Use docker registry
  • Use Kubernetes or ECS
  • Command line tool?
  • GPU image/container with CUDA 8 support
  • Dask job example
  • Use terraform variables
  • Support for other cloud providers (openstack, GCE, etc.)