Run many `adaptive.Learner`s on many cores (>10k) using `mpi4py.futures`, `ipyparallel`, `dask-mpi`, or `process-pool`.
Project description
Run many adaptive.Learners on many cores (>10k) using mpi4py.futures, ipyparallel, or distributed.
What is this?
The Adaptive scheduler solves the following problem, you need to run more learners than you can run with a single runner and/or can use >1k cores.
ipyparallel and distributed provide very powerful engines for interactive sessions. However, when you want to connect to >1k cores it starts to struggle. Besides that, on a shared cluster there is often the problem of starting an interactive session with ample space available.
Our approach is to schedule a different job for each adaptive.Learner. The creation and running of these jobs are managed by adaptive-scheduler. This means that your calculation will definitely run, even though the cluster might be fully occupied at the moment. Because of this approach, there is almost no limit to how many cores you want to use. You can either use 10 nodes for 1 job (learner) or 1 core for 1 job (learner) while scheduling hundreds of jobs.
Everything is written such that the computation is maximally local. This means that is one of the jobs crashes, there is no problem and it will automatically schedule a new one and continue the calculation where it left off (because of Adaptive’s periodic saving functionality). Even if the central “job manager” dies, the jobs will continue to run (although no new jobs will be scheduled.)
Design goals
Needs to be able to run on efficiently >30k cores
Works seamlessly with the Adaptive package
Minimal load on the file system
Removes all boilerplate of working with a scheduler
writes job script
(re)submits job scripts
Handles random crashes (or node evictions) with minimal data loss
Preserves Python kernel and variables inside a job (in contrast to submitting jobs for every parameter)
Separates the simulation definition code from the code that runs the simulation
Maximizes computation locality, jobs continue to run when the main process dies
How does it work?
You create a bunch of learners and corresponding fnames such that 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 get 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.
But how does it really work?
The ~adaptive_scheduler.server_support.RunManager basically does the following. So, 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 we 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, run_script="run_learner.py",)
# 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 the Python script that runs a learner (run_learner.py)
server_support._make_default_run_script(
url=url,
save_interval=300,
log_interval=30,
goal=None,
executor_type=scheduler.executor_type,
run_script_fname=scheduler.run_script,
)
# 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)
job_manager.start()Then when the job have been running for a while you can check server_support.parse_log_files(job_names, database_manager, scheduler).
And use scheduler.cancel(job_names) to cancel the jobs.
You don’t actually ever have to leave the Jupter notebook, take a look at the example notebook.
Jupyter notebook example
See example.ipynb.
Installation
WARNING: This is still the pre-alpha development stage.
Install the latest stable version from conda with (recommended)
conda install adaptive-scheduleror from PyPI with
pip install adaptive_scheduleror install master with
pip install -U https://github.com/basnijholt/adaptive-scheduler/archive/master.zipor clone the repository and do a dev install (recommended for dev)
git clone git@github.com:basnijholt/adaptive-scheduler.git
cd adaptive-scheduler
pip install -e .Development
In order to not pollute the history with the output of the notebooks, please setup the git filter by executing
python ipynb_filter.pyin the repository.
We also use pre-commit, so pip install pre_commit and run
pre-commit installin the repository.
Limitations
Right now adaptive_scheduler is only working for SLURM and PBS, however only a class like adaptive_scheduler/scheduler.py would have to be implemented for another type of scheduler. Also there are no tests at all!
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
File details
Details for the file adaptive_scheduler-0.9.15.tar.gz.
File metadata
- Download URL: adaptive_scheduler-0.9.15.tar.gz
- Upload date:
- Size: 42.1 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/3.4.1 importlib_metadata/4.0.1 pkginfo/1.7.0 requests/2.25.1 requests-toolbelt/0.9.1 tqdm/4.60.0 CPython/3.9.4
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 | e4a3fe47fe9d8d54133bf18fb895786ceaa27f99cd5349a05694de0f1c982e67 |
|
| MD5 | 4347ed84ae40b55774b534dab302dfa4 |
|
| BLAKE2b-256 | 048976af0872f9b82a655695b75337a124fcfe36ec4a9d6bfcea389416346f77 |
File details
Details for the file adaptive_scheduler-0.9.15-py3-none-any.whl.
File metadata
- Download URL: adaptive_scheduler-0.9.15-py3-none-any.whl
- Upload date:
- Size: 42.5 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/3.4.1 importlib_metadata/4.0.1 pkginfo/1.7.0 requests/2.25.1 requests-toolbelt/0.9.1 tqdm/4.60.0 CPython/3.9.4
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 | 7250949181871452ef5db0b07e306a2307648d2262da6e4ecfefb5316e14cdcf |
|
| MD5 | 8c7c48cd081e7b1a3406af7b272af466 |
|
| BLAKE2b-256 | de2bdb56500f05b1df4f12ec2a5e580ec263ebd2653a9873f9791425f69f8cab |
