This is code to test in the sense of benchmarking the Hail copy code. It is definitely rough but I think it points in a good direction. I found it quite useful.

It does four things:

• It creates VMs in GCP and AWS to run the tests.

• It creates data in a variety of configurations for the tests.

• It runs the tests and (over)writes results.json with the results.

• It cleans everything up.

You can do these all at once, or to support a development workflow, you can do each of these independently. In the course of use, you might want to add additional commands for smaller steps you want to isolate (e.g. a single test case instead of all of them).

Resources get prefixed with the profile ID so you can have multiple benchmarks running at once. You can set the profile with --profile.

To run this you will need:

• gcloud to be authorized to account that has access to the 1-day bucket and the broad-ctsa project. Your user account should work for this.

• Default AWS credentials should be for the AWS broad-hail-bench account test user. I find the AWS named profiles a useful feature when working with multiple accounts (like our account and the ODP account). See: https://docs.aws.amazon.com/cli/latest/userguide/cli-configure-profiles.html.

• An ssh keypair for broad-hail-bench-default in ~/.ssh/aws-broad-hail-bench-default AWS account to be able to log into instances in that account. You can manage those in the AWS console.

• A Hail package must be installed for the cleanup command that uses AsyncFS.rmtree.

The setup for the tests is specified in config.yaml. It has has extensive tests: 5GB spread across 1, 20, or 40K files or 40GB in one file, copied 6 ways: two and from GCS, S3 and the local file system.

config.yaml contains commands to create or delete VMs on GCP and AWS. On GCP, it uses a n2-standard-8 with an SSD, on Amazon the closest instance type I could find, a m5d.2xlarge with an SSD.

A data configuration is a total size, the number of files to spread the data across, and the number of directory levels to include (directory names are hex digits so the branching factor is 16).

test-copy.py is the main entrypoint.

The create-vms command creates the VMs.

The create-data command creates data in the locations given in the config file, which also specifies which VM to run them on.

The test-copy runs the tests given by the cases section of the config file. A case is a source, destination, and place to run the test. It writes the results.json file (and gives lots of output). test-copy does not verify the integrity of the copied data.

The cleanup command deletes all the resources.

The tool pulls the Hail repo on the VMs when running the tests. Which repo/version to choose is determined by the get section of the config file.