README.md

condor-tune

Warning

This project is not actively maintained. It was written with Ray Tune 1.11.0 and HTCondor 9.5.0. Since then, several vulnerabilities have been discovered (and patched) in Ray. requirements.txt has been updated to grab the latest version of Ray and HTCondor, though Ray's Tune API is likely to have changed.

This example project demonstrates the condor-tune framework for hyperparameter optimization. It comes from an application in ML for Power Systems research, where we first train a neural network using a GPU, then run parallel power flows using the trained network on every example in its respective validation set. We then must compute a weighted metric incorporating results from all power flows as well as raw model performance.

See condor_tune.py, Ray Tune, and HTCondor for relevant documentation.

Ray Tune restricts us to run all jobs on a single machine, so we need to spawn a condor job for each trial and automatically collect results from the trials. We then have access to Ray Tune to automatically optimize hyperparameters. Due to this structure, we can spawn multiple condor jobs per trial, even spawning job arrays if needed. This allows us to leverage the resources of the entire cluster rather than being limited to a single machine with a limited number of CPUs/GPUs. We can specify hardware requirements for each stage of the trial, freeing up resources for other trials and other cluster users.

The submitting process for a job array is a bit more involved. We have to prepare the necessary submit file and data files during the first (training) job:

1. At the end of the training job, we write an HTCondor submit file defining a job array to THIS_TRIAL_DIR/run_flow.cmd (i.e. queue args from file.txt)
2. Suppose every job in this array reads a json file like THIS_TRIAL_DIR/flows/42.json and runs a job with the required parameters. We can then create the flows directory and write the json files at the end of the training job.
3. Once we are done preparing the run_flow.cmd file and the flows directory, we create the flag file THIS_TRIAL_DIR/training_done and exit the training job.
4. Finally, this script submits the job array via the bash interface.

The scripts train.sh and metric.sh are used to run the neural network training and metric calculations respectively. They are located in TUNE_DIR. They look something like:

eval "$(conda shell.bash hook)"
conda activate project_env
python /home/klamike/condor-tune-project/train.py "$@"

The full folder structure is outlined at the bottom of misc.py. We omit the actual training, flow, and metric calculation scripts as they are out of scope for this project.

---

Michael Klamkin 2022
Georgia Institute of Technology
H. Milton Stewart School of Industrial Engineering
Pascal Van Hentenryck Lab, Risk-Aware Market Clearing (RAMC) Group