Skip to content
/ maestrowf Public

A tool to easily orchestrate general computational workflows both locally and on supercomputers

License

Notifications You must be signed in to change notification settings

LLNL/maestrowf

Repository files navigation

Maestro Workflow Conductor (maestrowf)

Build Status PyPI Spack Issues Forks Stars License

Downloads Downloads

Maestro can be installed via pip:

pip install maestrowf

Documentation

Getting Started is Quick and Easy

Create a YAML file named study.yaml and paste the following content into the file:

description:
    name: hello_world
    description: A simple 'Hello World' study.

study:
    - name: say-hello
      description: Say hello to the world!
      run:
          cmd: |
            echo "Hello, World!" > hello_world.txt

PHILOSOPHY: Maestro believes in the principle of a clearly defined process, specified as a list of tasks, that are self-documenting and clear in their intent.

Running the hello_world study is as simple as...

maestro run study.yaml

Creating a Parameter Study is just as Easy

With the addition of the global.parameters block, and a few simple tweaks to your study block, the complete specification should look like this:

description:
    name: hello_planet
    description: A simple study to say hello to planets (and Pluto)

study:
    - name: say-hello
      description: Say hello to a planet!
      run:
          cmd: |
            echo "Hello, $(PLANET)!" > hello_$(PLANET).txt

global.parameters:
    PLANET:
        values: [Mercury, Venus, Earth, Mars, Jupiter, Saturn, Uranus, Neptune, Pluto]
        label: PLANET.%%

PHILOSOPHY: Maestro believes that a workflow should be easily parameterized with minimal modifications to the core process.

Maestro will automatically expand each parameter into its own isolated workspace, generate a script for each parameter, and automatically monitor execution of each task.

And, running the study is still as simple as:

    maestro run study.yaml

Scheduling Made Simple

But wait there's more! If you want to schedule a study, it's just as simple. With some minor modifications, you are able to run on an HPC system.

description:
    name: hello_planet
    description: A simple study to say hello to planets (and Pluto)

batch:
    type:  slurm
    queue: pbatch
    host:  quartz
    bank:  science

study:
    - name: say-hello
      description: Say hello to a planet!
      run:
          cmd: |
            echo "Hello, $(PLANET)!" > hello_$(PLANET).txt
          nodes: 1
          procs: 1
          walltime: "00:02:00"

global.parameters:
    PLANET:
        values: [Mercury, Venus, Earth, Mars, Jupiter, Saturn, Uranus, Neptune, Pluto]
        label: PLANET.%%

NOTE: This specification is configured to run on LLNL's quartz cluster. Under the batch header, you will need to make the necessary changes to schedule onto other HPC resources.

PHILOSOPHY: Maestro believes that how a workflow is defined should be decoupled from how it's run. We achieve this capability by providing a seamless interface to multiple schedulers that allows Maestro to readily port workflows to multiple platforms.

For other samples, see the samples subfolder. To continue with our Hello World example, see the Basics of Study Construction in our documentation.

An Example Study using LULESH

Maestro comes packed with a basic example using LULESH, a proxy application provided by LLNL. You can find the example here.

What is Maestro?

Maestro is an open-source HPC software tool that defines a YAML-based study specification for defining multistep workflows and automates execution of software flows on HPC resources. The core design tenants of Maestro focus on encouraging clear workflow communication and documentation, while making consistent execution easier to allow users to focus on science. Maestro's study specification helps users think about complex workflows in a step-wise, intent-oriented, manner that encourages modularity and tool reuse. These principles are becoming increasingly important as computational science is continuously more present in scientific fields and has started to require a similar rigor to physical experiment. Maestro is currently in use for multiple projects at Lawrence Livermore National Laboratory and has been used to run existing codes including MFEM, and other simulation codes. It has also been used in other areas including in the training of machine-learned models and more.

Maestro's Foundation and Core Concepts

There are many definitions of workflow, so we try to keep it simple and define the term as follows:

A set of high level tasks to be executed in some order, with or without dependencies on each other.

We have designed Maestro around the core concept of what we call a "study". A study is defined as a set of steps that are executed (a workflow) over a set of parameters. A study in Maestro's context is analogous to an actual tangible scientific experiment, which has a set of clearly defined and repeatable steps which are repeated over multiple specimen.

Maestro's core tenets are defined as follows:

Repeatability

A study should be easily repeatable. Like any well-planned and implemented science experiment, the steps themselves should be executed the exact same way each time a study is run over each set of parameters or over different runs of the study itself.

Consistent

Studies should be consistently documented and able to be run in a consistent fashion. The removal of variation in the process means less mistakes when executing studies, ease of picking up studies created by others, and uniformity in defining new studies.

Self-documenting

Documentation is important in computational studies as much as it is in physical science. The YAML specification defined by Maestro provides a few required key encouraging human-readable documentation. Even further, the specification itself is a documentation of a complete workflow.


Setting up your Python Environment

To get started, we recommend using virtual environments. If you do not have the Python virtualenv package installed, take a look at their official documentation to get started.

To create a new virtual environment:

python -m virtualenv maestro_venv
source maestro_venv/bin/activate

Getting Started for Contributors

If you plan to develop on Maestro, install the repository directly using:

pip install poetry
poetry install

Once set up, test the environment. The paths should point to a virtual environment folder.

which python
which pip

Using Maestro Dockerfiles

Maestro comes packaged with a set of Docker files for testing things out. The two primary files are:

  • A standard Dockerfile in the root of the Maestro repository. This file is a standard install of Maestro meant to try out Maestro on the demo samples provided with this repository. In order to try Maestro locally, with Docker installed run:

    docker build -t maestrowf .
    docker run -ti maestrowf
    

    From within the container run the following:

    maestro run ./maestrowf/samples/lulesh/lulesh_sample1_unix.yaml
    
  • In order to try out Flux 0.19.0 integration, from the root of the Maestro repository run the following:

    docker build -t flux_0190 -f ./docker/flux/0.19.0/Dockerfile .
    docker run -ti flux_0190
    

    From within the container run the following:

    maestro run ./maestrowf/samples/lulesh/lulesh_sample1_unix_flux.yaml
    

Contributors

Many thanks go to MaestroWF's contributors.

If you have any questions or to submit feature requests please open a ticket.


Release

MaestroWF is released under an MIT license. For more details see the NOTICE and LICENSE files.

LLNL-CODE-734340