CODES Experiments

This repository contains scripts for network simulation using the CODES discrete-event simulation framework. Specifically, all scripts here run the binary model-net-mpi-replay which simulates the behaviour of one or multiple jobs running on an HPC network.

Structure

• individual-scripts/ - scripts to run one experiment at the time
• mpi_replay/ - Python 3.13 script to run a battery of tests
• visualizing_jobs/ - Visualizes each iteration time for all jobs in an experiment

Each experiment directory contains:

• conf/ - Configuration files for network topology and simulation parameters
• results/ - Output from simulation runs (logs, statistics, performance data)
• A python or shell script to run experiments with specific parameters

Feel free to copy individual scripts (or their entire subdirectory) and modify them to run new scenarios. Within mpi_replay/, you should be able to make a copy of run_mpi_surrogacy_experiments.py to run a series of experiments.

Individual scripts structure

• individual-scripts/dfly-1056/ - Dragonfly topology experiments with 1,056 nodes
• individual-scripts/dfly-72/ - Dragonfly topology experiments with 72 nodes
• individual-scripts/dfly-8448/ - Dragonfly topology experiments with 8,448 nodes
• individual-scripts/torus-64/ - Torus topology experiments with 64 nodes

Visualize iterations by multiple jobs

Once you have got some results, you can visualize how long each iteration took. Simply run:

python3 visualizing_jobs/print-iterations.py path-to/results/exp-XXX/experiment-name/iteration-logs/

For command line help, run python3 visualizing_jobs/print-iterations.py --help.

Usage

Run experiments using the provided script (it assumes you have compiled CODES using the CODES-compile-instructions.sh script and have downloaded this repo under the same directory that script resides. Please check the script CODES-compile-instructions.sh at https://github.com/codes-org/codes):

bash run-experiment.sh path-to-experiment/script.sh
# or in the case of mpi_replay
bash run-experiment.sh mpi_replay/run_mpi_surrogacy_experiments.py
# or in case you want to pass arguments to your experiment script, you can simply
bash run-experiment.sh path-to-experiment/script.sh --argument some-file.txt --other-arg

Results are automatically stored in path-to-experiment/results/.

In case you want to run an experiment with sbatch, you can use the script run-sbatch.sh instead of run-experiment.sh. The run-sbatch.sh script will run the experiments in a different folder to that of the script. This is because in systems where sbatch is needed, one often stores data in a different folder than the folder one is running the script. Under this new folder, the script will create a results/ subfolder just as run-experiment.sh does.

Dependencies

Requires the CODES simulation framework to be built and configured. It currently works with commit version @73cdbd5 of CODES.