xgenius

xgenius 🚀

xgenius is a command-line tool for managing remote jobs and containerized experiments across multiple clusters. It simplifies the process of building Docker images, converting them to Singularity format, and submitting jobs to clusters using SLURM.

Pre-requisites 🛠️

• You have a working Dockerfile.
• Singularity installed on your local machine.
• Docker installed on your local machine.
• docker login works.
• You have access to the clusters you want to run experiments on.
• Your project code is also cloned on the clusters.

Local Set-up 🧩

Installation 🔧

pip install xgenius

(Optional) Build Singularity Container from Dockerfile 🐳

xgenius-build-image --dockerfile=/path/to/Dockerfile \
--name=<output_image_name> \
--tag=<tag> \
--registry=<your_docker_username>

where --dockerfile is the ABSOLUTE path to your Dockerfile.

This command will build a Docker container, push it to your Docker registry, and then pull it to your local machine as a Singularity image. The Singularity image will be saved in the current directory under the name <output_image_name>.sif (the .sif extension will be added automatically).

Define Environment Variable 🌍

First, define the environment variable for the path where SLURM template files will be saved:

export XGENIUS_TEMPLATES_DIR=/path/to/your/templates

Recommendation: export XGENIUS_TEMPLATES_DIR=<your_project_path>/slurm_templates

Recommendation: Use a Conda environment and set:

conda env config vars set XGENIUS_TEMPLATES_DIR=/path/to/your/templates

This way you can have a different XGENIUS_TEMPLATES_DIR for each environment/project.

Otherwise, set the environment variable in your bashrc or ~/.zshrc to make it permanent.

Set-Up Cluster Configuration 🏗️

Run:

xgenius-setup-clusters

Follow the prompts to configure your cluster settings. You can add as many clusters as you want. Finish by answering 'done' at the end or the config file won’t be saved!

This creates cluster_config.json in the current directory.

Set-Up Run Configuration ⚙️

Pass the cluster_config.json file path to the following command to create run_config.json:

xgenius-setup-runs path/to/cluster_config.json

This creates run_config.json with placeholder values. The placeholder values are created according to the associated SLURM template for each cluster in cluster_config.json.

You are now all set up! Let’s run some experiments remotely!

Recommendation: If you leave cluster_config.json and run_config.json in your project directory, running commands will be super easy as you won't need to specify the paths ever again!

Running Experiments 🧪

• Push your Singularity image to the clusters you want:

  xgenius push-image \
  --image=path/to/singularity_image.sif \
  --clusters=cluster1,cluster2,cluster3
  

• Submit your jobs with:

  xgenius submit_jobs \
  --cluster=cluster1 \
  --run_command="python test.py" \
  --pull_repos
  

  Note: The --pull_repos flag is optional. It pulls changes from GitHub repositories before running the jobs. Always include it if your code is in a GitHub repository!

Done! Your jobs are now running on the cluster! 🎉

Batch jobs

You can also submit batch jobs using a JSON config file:

[
    {
        "command": "python test.py --test-arg1=1 --test-arg2=2",
        "cluster": "cluster1",
    },
    {
        "command": "python test.py --test-arg1=5 --test-arg2=10",
        "cluster": "cluster2",
    }
]

And running:

xgenius-batch-submit --batch-file=/path/to/batch_job.json --pull-repos

Utility Commands 🛠️

Check the status of your jobs in all clusters in cluster_config.json:

xgenius-check-jobs

Cancel all jobs in all clusters in cluster_config.json:

xgenius-cancel-jobs

Pull the results of your jobs from all clusters in cluster_config.json:

xgenius-pull-results

Remove the output folder in your clusters (useful before running a new batch of experiments)

xgenius-remove-results

Examples 📝

These files are created automatically with the commands above.

cluster_config.json

[
    {
        "cluster_name": "cluster1",
        "username": "<your_username>",
        "image_path": "<cluster1_scratch_folder>", # the path where the Singularity image will be saved in the cluster
        "project_path": "/path/to/project/code/in/cluster", # the path where your code is in the cluster. same as CODE_DIR_IN_CLUSTER in run_config.json
        "sbatch_template": "slurm_partition_template.sbatch" # the SLURM template file to use for this cluster. see the templates in the XGENIUS_TEMPLATES_DIR directory
    },
    {
        "cluster_name": "cluster2",
        "username": "<your_username>",
        "image_path": "<cluster2_scratch_folder>", 
        "project_path": "/path/to/project/code/in/cluster", 
        "sbatch_template": "slurm_partition_template.sbatch" 
    }
]

run_config.json

{
    "cluster1": {
        "SINGULARITY_COMMAND": "singularity", # or 'apptainer' depending on the cluster
        "NUM_GPUS": "1",
        "IMAGE_NAME": "<your_singularity_image_name>.sif",
        "PARTITION": "<partition_name>",
        "CODE_DIR_IN_CLUSTER": "/path/to/project/code/in/cluster",
        "OUTPUT_DIR_IN_CONTAINER": "/path/to/output/dir/in/container", # set this to the directory where your code writes output
        "TIME": "23:59:00", # for the time limit of the job
        "MODULES_TO_LOAD": "singularity", # or 'apptainer' depending on the cluster + any other modules
        "MEM": "12G", # example RAM memory per CPU
        "OUTPUT_DIR_IN_CLUSTER": "/path/to/cluster/scratch/runs", # your code outputs will be saved here. OUTPUT_DIR_IN_CLUSTER is binded to OUTPUT_DIR_IN_CONTAINER (see the slurm templates)
        "COMMAND": "python test.py", # the code you want to run
        "NUM_CPUS": "12", # example CPUs
        "OUTPUT_FILE": "/path/to/cluster/scratch/slurm-%j.out" # the logs file of the job
    }
}