5.4. Queueing Jobs (HPC, HTC)

Yet another interesting feature of junifer is the ability to queue jobs on computational clusters. This is done by adding the queue section in the Code-less Configuration file and executing the junifer queue command.

While junifer is meant to support HTCondor, SLURM and local queueing using GNU Parallel, only HTCondor and GNU Parallel are currently supported. This will be implemented in future releases of junifer. If you are in immediate need of any of these schedulers, please create an issue on the junifer github repository.

The queue section of the Code-less Configuration must start by defining the following general parameters:

  • jobname: Name of the job to be queued. This will be used to name the folder where the job files will be created, as well as any relevant file. Depending on the scheduler, it will also be listed in the queueing system with this name.

  • kind: The kind of scheduler to be used. Currently, only HTCondor and GNUParallelLocal are supported.

Example in YAML:

queue:
  jobname: TestHTCondorQueue
  kind: HTCondor

The rest of the parameters depend on the scheduler you are using.

5.4.1. HTCondor

When using HTCondor, junifer will use a DAG to queue one job per element (junifer run). As an option, the DAG can include a final job (junifer collect) to collect the results once all of the individual element jobs are finished.

The following parameters are available for HTCondor:

  • pre_run: Extra shell commands to run before junifer run.

  • pre_collect: Extra shell commands to run before junifer collect.

  • env: Definition of the Python environment. It has the following parameters:

    • kind: This is the kind of virtual environment to use:

      • conda

      • venv

      • local (no virtual environment)

    • name: This is the name of the environment to use in case a virtual environment is used. It should be the name when conda is used and the absolute or relative path to the virtualenv when venv is used. If relative path is used then it should be relative to the YAML.

    • shell: This is the shell to use. Only bash and zsh are supported as of now.

  • mem: Memory to be used by the job. It must be provided as a string with the units (e.g., "2GB").

  • cpus: Number of CPUs to be used by the job. It must be provided as an integer (e.g., 1).

  • disk: Disk space to be used by the job. It must be provided as a string with the units (e.g., "2GB"). Keep in mind that junifer uses a local working directory for each job, and datalad datasets might be cloned in this temporary directory.

  • extra_preamble: Extra lines to be added to the HTCondor submit file. This can be used to add extra parameters to the job, such as requirements.

  • collect: This parameter allows to include a collect to the DAG to collect the results once all of the individual element jobs are finished. This is useful if you want to run a junifer collect job only once all of the individual element jobs are finished. Valid options are:

    • yes: Include a collect job in the DAG that will be executed even if some of the individual element jobs fail.

    • on_success_only: Include a collect job to the DAG, but will only run if all of the individual element jobs are successful.

    • no: Do not include a collect job to the DAG.

Example in YAML:

queue:
  jobname: TestHTCondorQueue
  kind: HTCondor
  env:
    kind: conda
    name: junifer
    shell: zsh
  mem: 8G
  disk: 2G
  collect: "yes"  # wrap it in string to avoid boolean

Alternatively, if you use venv, the YAML would look like so:

queue:
  jobname: TestHTCondorQueue
  kind: HTCondor
  env:
    kind: venv
    name: /home/me/junifer-venv  # should be at the level above `bin/activate`
    shell: zsh
  mem: 8G
  disk: 2G
  collect: "yes"  # wrap it in string to avoid boolean

5.4.2. GNUParallelLocal

When using GNU Parallel, junifer will run only in local mode and leverage maximum computational power of whatever hardware it is run on.

The following parameters are available for GNUParallelLocal:

  • pre_run: Extra shell commands to run before junifer run.

  • pre_collect: Extra shell commands to run before junifer collect.

  • env: Definition of the Python environment. It has the following parameters:

    • kind: This is the kind of virtual environment to use:

      • conda

      • venv

      • local (no virtual environment)

    • name: This is the name of the environment to use in case a virtual environment is used. It should be the name when conda is used and the absolute or relative path to the virtualenv when venv is used. If relative path is used then it should be relative to the YAML.

    • shell: This is the shell to use. Only bash and zsh are supported as of now.

Example in YAML:

queue:
  jobname: TestGNUParallelLocalQueue
  kind: GNUParallelLocal
  env:
    kind: conda
    name: junifer
    shell: zsh

Alternatively, if you use venv, the YAML would look like so:

queue:
  jobname: TestGNUParallelLocalQueue
  kind: GNUParallelLocal
  env:
    kind: venv
    name: /home/me/junifer-venv  # should be at the level above `bin/activate`
    shell: zsh

Once the Code-less Configuration file is ready, including the queue section, you can queue the jobs by executing the junifer queue command.

The queue command will create a folder with the name of the job (jobname) under the junifer_jobs directory in the current working directory.

The queue command accepts the following arguments:

  • --help: Show a help message.

  • --verbose: Set the verbosity level. Options are warning, info, debug.

  • --submit: Submit the jobs to the queueing system. If not specified, the job submit files will be created but not submitted.

  • --overwrite: Overwrite the job folder if it already exists. If not specified, the command will fail if the job folder already exists.

  • --element: Queue only the specified element(s). If not specified, all elements will be queued.