Typical workflow

Step 0: Setup AWS credentials (one time)

SpotML uses your AWS credentials to manage the runs for you. You can verify that you have the credentials setup if you see the contents of ~/.aws/credentials to be something like below.

[my-aws-profile]
aws_access_key_id = your_aws_access_key_id
aws_secret_access_key = your_aws_secret_access_key

Secondly, the above access key IAM user needs to have the permissions to create all the aws resources.

If you are new to AWS, configure your aws cli by going through this setup guide.

Step 1: Install spotml cli (one time)

pip install spotml --upgrade

Step 2: Configure spotml.yaml

First, copy-paste below file in your code root folder.

project:
  name: mnist
  syncFilters:
    - exclude:
        - .git/*
        - .idea/*
        - '*/__pycache__/*'

containers:
  - &DEFAULT_CONTAINER
    projectDir: /workspace/project
    image: tensorflow/tensorflow:latest-py3
    volumeMounts:
      - name: workspace
        mountPath: /workspace
    env:
      PYTHONPATH: /workspace/project
    ports:
      # tensorboard
      - containerPort: 6006
        hostPort: 6006
      # jupyter
      - containerPort: 8888
        hostPort: 8888

instances:
  - name: aws-1
    provider: aws
    parameters:
      region: us-east-1
      instanceType: t2.large
      spotStrategy: on-demand
      ports: [6006, 6007, 8888]
      rootVolumeSize: 125
      volumes:
        - name: workspace
          parameters:
            size: 50

scripts:
  train: |
    python train.py

  tensorboard: |
    tensorboard --port 6006 --logdir results/

  jupyter: |
    CUDA_VISIBLE_DEVICES="" jupyter notebook --allow-root --ip 0.0.0.0

2. Change the instanceType to the aws instance you want to run the code on.

      instanceType: t2.large
      spotStrategy: on-demand

Step 3: Create a Docker file (optional)

In the above spotml.yaml notice that we've used a docker image directly.

  image: tensorflow/tensorflow:latest-py3

But if you want to configure a custom docker container, you can write your own Dockerfile. See example setup here.

Step 4: Start an instance

spotml start

This will automatically do the following for you:

Syncs your source code folder to S3
Launch your instance (on-demand or spot instance)
Create a persistent EBS volume and attach it to the above instance.
Setup your docker environment inside the AWS instance
Pulls the code/data from S3 to the EBS volume attached to the instance

The resources created in your AWS account(EBS Volume, EC2 instance, S3 Bucket, Security Groups, etc.) will all have the prefix spotml.

You should see an output like below once the instance is launched.

Step 5: SSH into the instance

spotml sh

You should see a screen like below. Notice that your source code is already synced into the instance.

You can now run your training command here to kick start a manual training.

Step 6: Schedule a managed run (optional)

If you have a long running training jobs that you only want to run on spot instances, it's a pain to manage them manually.

Configure your script section in spotml.yaml file to the command you want

scripts:
  train: |
    python train.py

2. Change the spotStrategy to on-demand

spotStrategy: on-demand

3. Then type below command to let SpotML automatically manage the run.

spotml run train

If the instance is not already running, spotML tries to spawn a new spot instance and runs the above script once the instance is ready.

Note that if a spot instance is not available, spotML backend service keeps trying every 15 mins, until it can spawn the instance. So you can turn off your laptop and do other things, while SpotML tries to schedule the run.

If you intend to cancel the scheduled run, type