Scheduling

When executing experiments, EOS uses scheduling to determine the execution order of tasks. EOS offers two types of scheduling policies:

1. Greedy: tasks are scheduled for execution as soon as their requirements (e.g., precedence and device constraints) are met.

2. CP-SAT: tasks are scheduled based on not only when their requirements are satisfied but also in a way that minimizes the total duration of all submitted experiments. To make this happen, the scheduler needs to know how long each task is expected to take.

Choosing a scheduler

The desired scheduler can be chosen by modifying the EOS config.yml file:

config.yml

# ...

scheduler:
    type: greedy # or cpsat

The greedy scheduler is lightweight and is the recommended scheduler for simpler experiments where task durations do not have a lot of variance.

The CP-SAT scheduler should be used for more complex experiments, such as experiments with parallel branches, where ordering tasks by taking into account their durations can make a big difference in the total duration of the experiment. CP-SAT is a state-of-the-art constraint programming solver, and thus may require significant computational resources to run, especially for complex experiment task graphs. CP-SAT greatly benefits from multiprocessing, so the more cores on the EOS computer, the better.

Task durations

When using the CP-SAT scheduler, every task in an experiment.yml file must be assigned an expected duration in seconds. For compatibility reasons, every task has a default duration of 10 seconds, unless specified otherwise. See below:

experiment.yml

# ...

tasks:
    # ...
    - id: analyze_color
        type: Analyze Color
        desc: Determine the RGB value of a solution in a container
        duration: 5
        devices:
          - lab_id: color_lab
            id: color_analyzer
        containers:
          beaker: move_container_to_analyzer.beaker
        dependencies: [move_container_to_analyzer]
    # ...