0g-chain/simulations

How To Run Sims In The Cloud

Sims run with AWS batch, with results uploaded to S3

AWS Batch

In AWS batch you define:

• a "compute environment"--just how many machines you want (and of what kind)
• a "job queue"--just a place to put jobs (pairs them with a compute environment)
• a "job definition"--a template for jobs

Then to run stuff you create "jobs" and submit them to a job queue.

The number of machines running auto-scales to match the number of jobs. When there are no jobs there are no machines, so you don't pay for anything.

Jobs are defined as a docker image (assumed hosted on dockerhub) and a command string.

e.g. kava/kava-sim:version1, go test ./app

This can run sims but doesn't collect the results. This is handled by a custom script.

Running sims and uploading to S3

The dockerfile in this repo defines the docker image to run sims. It's just a normal app, but with the aws cli included, and the custom script.

The custom script reads some input args, runs a sim and uploads the stdout and stderr to a S3 bucket.

AWS Batch allows for "array jobs" which are a way of specifying many duplicates of a job, each with a different index passed in as an env var.

Steps

• create and submit a new array job (based of the job definition) with
  • image kava/kava-sim:<some-version>
  • command run-then-upload.sh <starting-seed> <num-blocks> <block-size>
  • array size of how many sims you want to run
• any changes needed to the code or script necessitates a rebuild:
  • docker build -f simulations/Dockerfile -t kava/kava-sim:<some-version> .
  • docker push kava/kava-sim:<some-version>

Tips

• click on the compute environment name, to get details, then click the link ECS Cluster Name to get details on the actual machines running
• for array jobs, click the job name to get details of the individual jobs