parallelic 0.1.2

Hyper-parallel multi-node task execution engine

Parallelic

Parallelic is a hyperparallel multi-node task execution engine with shared data and wokspace capabilities.

Note of warning

Parallelic is not a containerization/sandboxing engine. It does not constitute a full task isolation, and provides no guarantee of such. That may change in the future, and feel free to contribute your code towards that goal, but in the mean time, keep this in consideration when giving access to a Parallelic system to third parties.

Installation

From git

1. Clone the git repo locally.
2. Download python3(.7) and corresponding pip
3. Install Poetry
4. Run poetry install to create a virtualenv and install dependencies
   At this point, you can use parallelic through
   poetry run python -m parallelic
5. Run poetry build to build a wheel
6. Run pip install dist/parallelic-[version]-py3-none-any.whl
   Now you can use parallelic without poetry:
   python -m parallelic

From pip

1. Run pip install parallelic

Usage

Running a task

To run an already defined task, you upload the task package (a zipped up task root directory) via the Parallelic WebUI, or Parallelic CLI client, to the Parallelic manager instance.
You may need to provide access credentials before being allowed to upload the task package, as per your Parallelic system configuration.
From there, the Parallelic manager instance will take care of everything else.

Defining a task

Task definitions follow a particullar directory tree.
The task root contains a task.yml file, that contains metadata required for the manager to set up and prepare resources for the compute nodes in order to run the particullar task.

The source directory gets distributed to all compute nodes, it should contain only the source files required to run the task.
This directory is set up with Read+eXecute permissions. Every task has to contain this directory.

The data directory gets distributed to all compute nodes, if "shared_data" is requested in the task's configuration. It is set up with Read+Write permissions.
It is suitable for datasets that need to be shared across all nodes, but whose modifications by the nodes shouldn't propagate to other nodes. A task should contain this directory if "shared_data" is requested.
It can be accessd by the task in the task's working directory.

The workspace directory gets created automatically by the manager, if "shared_workspace" is requested in the task's configuration.
It is set up with Read+Write permissions, but task.yml can request the manager to create a scripts subdirectory, with Read+Write+eXecute permissions.
It should never be in the task's root directory, and any content from the task's root directory won't get sent to the compute nodes. Compute nodes should populate this directory themselves, at runtime. Any changes to this directory will propagate across all compute nodes, and it can be access in the task's working directory.