hydra-lsf-launcher

This is a simple launcher for launching Hydra jobs on LSF clusters. It is a simple wrapper around the bsub command that allows you to launch Hydra jobs with a single command.

WARNING: expect bugs and run at your own risk. This is not an official Hydra project.

Installation

Install the package with pip:

pip install hydra-lsf-launcher

Usage

Here is an example using hydra-zen to launch a job on an LSF cluster:

from hydra_zen import store, zen
import time


def main_func(sentence):
    time.sleep(120)
    print(sentence)


store(
    main_func,
    hydra_defaults=[
            "_self_",
    ]
)


if __name__ == "__main__":
    store.add_to_hydra_store()
    zen(main_func).hydra_main(
        config_name="main_func", version_base="1.1", config_path=None
    )

Then you can run it with:

python myscript.py hydra/launcher=lsf sentence="Hello World!" -m 

Job groups

To limit the number of concurrent jobs, as for now job arrays are not supported, the user can set job groups:

bgadd /myJobGroup
bgmod -L 4 /myJobGroup  # limit to 4 concurrent jobs

Then you can run it with:

python myscript.py hydra/launcher=lsf hydra.launcher.bsub_args="-g /myJobGroup" sentence="Hello World!" -m 

Some useful commands. To change the limit of a job group:

bgmod -L <new_limit> /myJobGroup

To delete a job group:

bgdel /myJobGroup

More complex examples

Run 5 jobs (4 parallel), each with 2GB of memory, and 4 cores from 2 different nodes.

python myscript.py hydra/launcher=lsf hydra.launcher.verbose=False hydra.launcher.n=4 'hydra.launcher.bsub_args= -g /myJobGroup -R span\[ptile\=2\]' hydra.launcher.M="2GB" sentence="Hello World!,Hello World!,Hello World!,Hello World!,Hello World!" -m