This repository contains the deployment code that sets up the necessary AWS resources to utilize the AWS Batch runner in Raster Vision. Using Batch is advantageous because it starts and stops instances automatically and runs jobs sequentially or in parallel according to the dependencies between them. In addition, this deployment sets up distinct CPU and GPU resources and utilizes spot instances, which is more cost-effective than always using a GPU on-demand instance. Deployment is driven via the AWS console using a CloudFormation template. This AWS Batch setup is an "advanced" option that assumes some familiarity with Docker, AWS IAM, named profiles, availability zones, EC2, ECR, CloudFormation, and Batch.

• AWS Account Setup
• AWS Credentials
• AMI Creation
• Deploying Batch resources
• Update Raster Vision configuration

In order to setup Batch using this repo, you will need to setup your AWS account so that:

• you have either root access to your AWS account, or an IAM user with admin permissions. It may be possible with less permissions, but we haven't figured out how to do this yet after some experimentation.
• you have the ability to launch P2 or P3 instances which have GPUs. In the past, it was necessary to open a support ticket to request access to these instances. You will know if this is the case if the Packer job fails when trying to launch the instance.
• you have requested permission from AWS to use availability zones outside the USA if you would like to use them. (New AWS accounts can't launch EC2 instances in other AZs by default.) If you are in doubt, just use us-east-1.

Using the AWS CLI, create an AWS profile for the target AWS environment. An example, naming the profile raster-vision:

$ aws --profile raster-vision configure
AWS Access Key ID [****************F2DQ]:
AWS Secret Access Key [****************TLJ/]:
Default region name [us-east-1]: us-east-1
Default output format [None]:

You will be prompted to enter your AWS credentials, along with a default region. The Access Key ID and Secret Access Key can be retrieved from the IAM console. These credentials will be used to authenticate calls to the AWS API when using Packer and the AWS CLI.

This step uses packer to install nvidia-docker on the base ECS AMI in order to run GPU jobs on AWS Batch.

Copy the settings.mk.template file to settings.mk, and fill out the options shown in the table below. Remaining variables in the settings file will be filled in later.

To find the latest Deep Learning Base AMI, search in the AMI section of your EC2 AWS console for Deep Learning Base AMI (Amazon Linux). The steps for doing that are as follows.

First, go into the ECS section of the interface as depicted in the image below.

Second, click on the Launch Instance button within the ECS interface. You will not launch an instance from the web interface as part of these instructions, we are merely using the interface behind that button to find the desired AMI.

Finally, type Deep Learning Base AMI (Amazon Linux) into the search box on that page. The AMI identification number that you need can (at time of writing) be found in the place highlighted in the screenshot. Please note that the AMI that you need varies with time and region, so you should go through the three steps rather than copying from the screenshot.

Ensure that the AWS profile for the account you want to create the AMI in is set in your AWS_PROFILE environment variable setting. If you skip this step, Packer will freeze.

Then run:

> make create-ami

This will run Packer, which will spin up an EC2 instance, install the necessary resources, create an AMI off of the instance, and shut the instance down. Be sure to record the AMI ID, which will be given in the last line of the output.

To deploy AWS Batch resources using AWS CloudFormation, start by logging into your AWS console. Then, follow the steps below:

• Navigate to CloudFormation > Create Stack
• In the Choose a template field, select Upload a template to Amazon S3 and upload the template in cloudformation/template.yml
• Prefix: If you are setting up multiple RV stacks within an AWS account, you need to set a prefix for namespacing resources. Otherwise, there will be name collisions with any resources that were created as part of another stack.
• Specify the following required parameters:
  • Stack Name: The name of your CloudFormation stack
  • VPC: The ID of the Virtual Private Cloud in which to deploy your resource. Your account should have at least one by default.
  • Subnets: The ID of any subnets that you want to deploy your resources into. Your account should have at least two by default; make sure that the subnets you select are in the VPC that you chose by using the AWS VPC console, or else CloudFormation will throw an error. (Subnets are tied to availability zones, and so affect spot prices.) In addition, you need to choose subnets that are available for the instance type you have chosen. To find which subnets are available, go to Spot Pricing History in the EC2 console and select the instance type. Then look up the availability zones that are present in the VPC console to find the corresponding subnets.
  • SSH Key Name: The name of the SSH key pair you want to be able to use to shell into your Batch instances. If you've created an EC2 instance before, you should already have one you can use; otherwise, you can create one in the EC2 console. Note: If you decide to create a new one, you will need to log out and then back in to the console before creating a Cloudformation stack using this key.
  • AMI: For the GPU AMI, provide the ID of the AMI that you created above. For the CPU AMI, you need to use the ECS-optimized AMI. You can find the AMI ID for your availability zone here. If you use the same AMI for both, CPU jobs will fail with the following error:

  Container messageCannotStartContainerError: Error response from daemon: OCI runtime create failed: container_linux.go:348: starting container process caused "process_linux.go:402: container init caused \"process_linux.go:385: running prestart hook 0 caused \\\"error runni )
  

  • Instance Types: Provide the instance types you would like to use. (For GPUs, p3.2xlarge is approximately 4 times the speed for 4 times the price.)
• Adjust any preset parameters that you want to change (the defaults should be fine for most users) and click Next.
  • Advanced users: If you plan on modifying Raster Vision and would like to publish a custom image to run on Batch, you will need to specify (CPU and GPU) ECR repo names and a tag name to use for both. Note that the repo names cannot be the same as the Stack name (the first field in the UI) and cannot be the same as any existing ECR repo names. If you are in a team environment where you are sharing the AWS account, the repo names should contain an identifier such as your username.
• Accept all default options on the Options page and click Next
• Accept I acknowledge that AWS CloudFormation might create IAM resources with custom names on the Review page and click Create
• Watch your resources get deployed!

If you setup ECR repositories during the CloudFormation setup (the "advanced user" option), then you will need to follow this step, which publishes local Raster Vision images to those ECR repositories. Every time you make a change to your local Raster Vision images and want to use those on Batch, you will need to run this step.

Run ./docker/build in the main Raster Vision repo to build local copies of the CPU and GPU images.

In settings.mk, fill out the options shown in the table below.

Run make publish-container to publish the CPU and GPU images to your ECR repositories.

Finally, make sure to update your Raster Vision configuration with the Batch resources that were created.