Skip to content

Run many functions (adaptively) on many cores (>10k-100k) using mpi4py.futures, ipyparallel, loky, or dask-mpi. πŸŽ‰

License

Notifications You must be signed in to change notification settings

basnijholt/adaptive-scheduler

Repository files navigation

Asynchronous Job Scheduler for Adaptive πŸš€

PyPI Conda Downloads Build Status Documentation Status CodeCov

This is an asynchronous job scheduler for Adaptive, designed to run many adaptive.Learners on many cores (>10k-100k) using mpi4py.futures, ipyparallel, loky, concurrent.futures.ProcessPoolExecutor, or dask.distributed.

πŸ“š Table of Contents

πŸ€” What is this?

Adaptive Scheduler is designed to address the challenge of executing a large number of adaptive.Learners in parallel, even when using more than 1k-100k cores. Traditional engines like ipyparallel and distributed can struggle with such high core counts because there is a central process that communicates with each worker.

This library schedules a separate job for each adaptive.Learner, and manages the creation and execution of these jobs. This ensures that your calculations will run even if the cluster is currently fully occupied (because job will just be put in the queue). The approach allows for nearly limitless core usage, whether you allocate 10 nodes for a single job or 1 core for a single job while scheduling hundreds of jobs.

The computation is designed for maximum locality. If a job crashes, it will automatically reschedule a new one and continue the calculation from where it left off, thanks to Adaptive's periodic saving functionality. Even if the central "job manager" fails, the jobs will continue to run, although no new jobs will be scheduled.

🎯 Design Goals

  1. Needs to be able to run efficiently on >30k cores.
  2. Works seamlessly with the Adaptive package.
  3. Minimal load on the file system.
  4. Removes all boilerplate of working with a scheduler:
    • Writes job script.
    • (Re)submits job scripts.
  5. Handles random crashes (or node evictions) with minimal data loss.
  6. Preserves Python kernel and variables inside a job (in contrast to submitting jobs for every parameter).
  7. Separates the simulation definition code from the code that runs the simulation.
  8. Maximizes computation locality, jobs continue to run when the main process dies.

πŸ§ͺ How does it work?

You create a bunch of learners and corresponding fnames so they can be loaded, like:

import adaptive
from functools import partial

def h(x, pow, a):
    return a * x**pow

combos = adaptive.utils.named_product(
    pow=[0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
    a=[0.1, 0.5],
)  # returns list of dicts, cartesian product of all values

learners = [adaptive.Learner1D(partial(h, **combo),
            bounds=(-1, 1)) for combo in combos]
fnames = [f"data/{combo}" for combo in combos]

Then you start a process that creates and submits as many job-scripts as there are learners, like:

import adaptive_scheduler

def goal(learner):
    return learner.npoints > 200

scheduler = adaptive_scheduler.scheduler.SLURM(cores=10)  # every learner gets this many cores

run_manager = adaptive_scheduler.server_support.RunManager(
    scheduler,
    learners,
    fnames,
    goal=goal,
    log_interval=30,  # write info such as npoints, cpu_usage, time, etc. to the job log file
    save_interval=300,  # save the data every 300 seconds
)
run_manager.start()

That's it! You can run run_manager.info() which will display an interactive ipywidget that shows the amount of running, pending, and finished jobs, buttons to cancel your job, and other useful information.

Widget demo

πŸ” But how does it really work?

The adaptive_scheduler.server_support.RunManager basically does the following:

  • You need to create N learners and fnames (like in the section above).
  • Then a "job manager" writes and submits max(N, max_simultaneous_jobs) job scripts but doesn't know which learner it is going to run!
  • This is the responsibility of the "database manager", which keeps a database of job_id <--> learner.
  • The job script starts a Python file run_learner.py in which the learner is run.

In a Jupyter notebook, you can start the "job manager" and the "database manager", and create the run_learner.py like:

import adaptive_scheduler
from adaptive_scheduler import server_support

# create a scheduler
scheduler = adaptive_scheduler.scheduler.SLURM(cores=10)

# create a new database that keeps track of job <-> learner
db_fname = "running.json"
url = (
   server_support.get_allowed_url()
)  # get a url where we can run the database_manager
database_manager = server_support.DatabaseManager(
   url, scheduler, db_fname, learners, fnames
)
database_manager.start()

# create unique names for the jobs
n_jobs = len(learners)
job_names = [f"test-job-{i}" for i in range(n_jobs)]

job_manager = server_support.JobManager(
    job_names,
    database_manager,
    scheduler,
    save_interval=300,
    log_interval=30,
    goal=0.01,
)
job_manager.start()

Then, when the jobs have been running for a while, you can check server_support.parse_log_files(database_manager, scheduler).

And use scheduler.cancel(job_names) to cancel the jobs.

You don't actually ever have to leave the Jupyter notebook; take a look at the example notebook.

πŸ““ Jupyter Notebook Example

See example.ipynb.

πŸ’» Installation

Install the latest stable version from conda (recommended):

conda install adaptive-scheduler

or from PyPI:

pip install adaptive_scheduler

or install main with:

pip install -U https://github.com/basnijholt/adaptive-scheduler/archive/main.zip

or clone the repository and do a dev install (recommended for dev):

git clone [email protected]:basnijholt/adaptive-scheduler.git
cd adaptive-scheduler
pip install -e .

πŸ› οΈ Development

In order not to pollute the history with the output of the notebooks, please set up the git filter by executing:

python ipynb_filter.py

in the repository.

We also use pre-commit, so pip install pre_commit and run:

pre-commit install

in the repository.

⚠️ Limitations

Currently, adaptive_scheduler only works for SLURM and PBS. However, only a class like adaptive_scheduler/scheduler.py would have to be implemented for another type of scheduler.