ictasks 0.2.3

A utility to run multiple small tasks in a single HPC job submission.

ictasks

ictasks is a library to support running collections of small tasks in parallel, from a laptop through to a HPC cluster.

It is being developed to support projects and workflow standarisation at the Irish Centre for High End Computing (ICHEC).

There are many libraries like this, this one focuses on:

• being small and simple - easy for ICHEC researchers to understand all elements of it
• having low dependencies - can be installed with pip
• having an easy to use configuration format with yaml
• having verbose and structured reporting of task related metadata for analysis

Features

The library allows a few different ways to specify tasks.

1. As a Python package

You can create a collection of Task instances and launch them with the session.run() function.

2. Using a tasklist.dat file and the command line interface (CLI):

ictasks taskfarm --tasklist $PATH_TO_TASKLIST

See the test/data/tasklist.dat file for an example input with two small tasks.

3. Using a yaml config file, giving task and environment info.

job_id: my_job
task_distribution:
    cores_per_node: 1

tasks:
    items:
    - id: 0
      launch_cmd: "echo 'hello from task 0'"
    - id: 1
      launch_cmd: "echo 'hello from task 1'"

we can run this with:

ictasks taskfarm --config my_config.yaml

Launching the run will launch the task in that directory and output a status file task.json. By default all processors on the machine (or compute node) will be assigned tasks.

task_distribution

There are 5 variables available in task_distribution

• cores_per_node=0
• threads_per_core=0
• cores_per_task=1
• gpus_per_node=0
• gpus_per_task=0

If left unspecified, both cores_per_task and gpus_per_task will be automatically detected and set.

The number of active workers will be set depending on these resource limitations, with gpu limits being hit first if needed. These workers can then execute tasks in parallel.

gpus in your tasks

When taskfarming with gpus, there will be a gpu_ids.txt file in each task directory which contains a gpu_id per line, the user's program must then use this file to specify which gpu(s) to use. The function ictasks.task.get_gpu_ids can be used to retrieve the ids in this file as a Python list. See an example here.

Installing

The package can be installed from PyPI:

pip install ictasks

Contact

Our Gitlab hosting arrangement doesn't allow us to easily accept external contributions or feedback, however they are still welcome.

In future if there is interest

If you have any feedback on this library

License

This package is Coypright of the Irish Centre for High End Computing. It can be used under the terms of the GNU Public License (GPL v3+). See the included LICENSE.txt file for details.

If you are an ICHEC collaborator or user of the National Service different licensing terms can be offered - please get in touch to discuss.