🍃 Midori

Midori is a plugin-based orchestrator that allows researchers to run software engineering experiments on remote clusters. It is a byproduct of my thesis. It is designed to be automated for time-consuming experiments that require a large number of runs.

🧱 Architectural

• Why Midori? Because there are so many things to consider when it comes to running software engineering experiments, such as:
  • Cooling down before each trial.
  • SSH Connection.
  • Randomizing the order of the trials.
  • Switching between different treatments.
  • ...
• What is an orchestrator? Midori is meant to be a local tool that manipulates the remote cluster to run the experiments using SSH. In this way, the overhead of the experiment on the remote cluster is minimized.

🚀 Getting Started

Plugins and experiment subject are essential to conducting experiments with Midori. Plugins are used to interact with experiment subject or the remote cluster as a whole.

Installation

Midori is only compatible with Python 3.12. Install it via pip or other package managers:

pip install midori

Usage

Initialize the orchestrator with the following parameters and run:

from midori.core import Orchestrator

orchestrator = Orchestrator(
  hostname: str,
  username: str,
  password: str,
  repetitions: int,
  before_trial_cooling_time: int,
  trial_timespan: int,
  after_trial_cooling_time: int,
  variables: Dict[str, List[str]],
  subject_path: str,
  before_experiment_plugins: List[Type[PluginHelper]] = [],
  setup_plugins: List[Type[PluginHelper]] = [],
  end_trial_plugins: List[Type[PluginHelper]] = [],
  end_experiment_plugins: List[Type[PluginHelper]] = [],
)

orchestrator.run()

Remote Node and Experiment Subject Setup

• The subject_path is a directory on the remote node. It is a valid experiment subject if
  • It is a git repository.
  • It contains branches named after the combinations of the variables. Each branch represents a treatment.

For example, if variable is

variables = {'drink': ['orange', 'apple', 'watermelon'], 'size': [300, 600, 100], 'temperature': [20, 30]}

The subject_path should be a git repository that contains branches named orange-300-20, orange-300-30, orange-600-20, orange-600-30, etc. You do not have to include all the combinations, Midori will skip the missing branches.

Parameters

• hostname, username, password: Midori needs these parameters to connect to the remote node via SSH.
• repetitions: The number of repetitions for the experiment.
• before_trial_cooling_time: The cooling time before each trial, in seconds.
• trial_timespan: The time span of each trial, in seconds.
• after_trial_cooling_time: The cooling time after each trial, in seconds.
• variables: The variables to be manipulated in the experiment. For example, if you have two variables, x and y, and you want to run the experiment with the values of x as [1, 2, 3] and the values of y as [4, 5, 6], you can define the variables as {'x': [1, 2, 3], 'y': [4, 5, 6]}.
• subject_path: The path of the subject to be run on the remote cluster. For example, if you want to run the subject in the ~/subject directory, you can define the path as ~/subject.
• before_experiment_plugins: The plugins to be run before the experiment, such as setting environment variables, installing dependencies, close unnecessary services, etc.
• setup_plugins: The plugins to be run before each trial, such as setting up treatment-dependent environment variables, cleaning up the environment, etc.
• end_trial_plugins: The plugins to be run after each trial, such as collecting logs, cleaning up the environment, etc.
• end_experiment_plugins: The plugins to be run after the experiment.

To help you better understand this, the green boxes in the collapsed illustration represent the parameters.

What are these Plugins?

Clearly there are 4 parameters that invoke the plugins to be run at different stages of the experiment. The 4 parameters are lists of class definitions (not instances).

Plugins defined in before_experiment_plugins, setup_plugins, end_trial_plugins, end_experiment_plugins are run in the order. [A, B, C] means A is run first, then B, and finally C, with the output of A passed to B and the output of B passed to C.

The way we define these plugins is by inheriting from the PluginHelper class and overriding the action method. Here is an example:

from midori.plugins.helpers import PluginHelper

class SkaffoldSetup(PluginHelper):
    def action(self) -> str:
        return f"cd {self.subject_path} && skaffold delete && skaffold run"

class CheckRam(PluginHelper):
    def action(self) -> str:
        return "cat /proc/meminfo | grep MemTotal"

action() method should return either an str or None. If it returns an str, it will be executed on the remote cluster. If it returns None, it will be skipped. You may want to implement side effects in the action() method based on 3 accessible attributes:

• self.subject_path: The path of the subject to be run on the remote cluster, which is the same as the subject_path parameter of Orchestrator.
• self.previous_output: The output of the previous plugin. For example, if end_trial_plugins = [CheckRam, NewPlugin], where CheckRam is the plugin shown above, the NewPlugin will have access to the output of CheckRam in self.previous_output as MemTotal: 900000000 kB.
• self.treatment: The treatment of the current trial.

Escape Hatch

• If you really want access to the SSH client in the plugin action() method, you can access it via self._ssh (it's paramiko.SSHClient). Most of the time it's not necessary to use it. The only reason I kept it is to make my old plugins work with the new version of Midori.

🚧 Development and Contribution

1. Install poetry.
2. Navigate to the root of the project and run poetry install.
3. Enter the virtual environment by running poetry shell.
4. Run pre-commit install to install pre-commit hooks.
5. Develop on dev branch and create a pull request to dev branch.

📝 Citation

@misc{midori,
  author = {Xingwen Xiao},
  title = {Midori},
  year = {2024},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/imchell/midori}},
}