Colmena simplifies building autonomous applications that steer large campaigns of simulations on supercomputers.
Such "high-throughput" campaigns were, historically, guided by humans identifying the which tasks to run next — a time-consuming process with a high latency between "new data" and "decisions."
Colmena was created to build applications which augment or replace human steering with Artificial Intelligence (AI).
Colmena is available via PyPI: pip install colmena
Consult our Installation Guide for further details.
Colmena applications describe a computational campaign in two components: a "Thinker" that picks computations and a "Doer" which executes them.
Thinkers encode the logic for how to run new calculations as "agents." Complex strategies are simple to express when you decompose them into simple steps. For example, a distributed optimizer:
from random import random
from colmena.thinker import BaseThinker, result_processor, task_submitter, ResourceCounter
from colmena.queue import PipeQueues
from colmena.models import Result
# Build queues to connect Thinker and Doer
queues = PipeQueues()
class Thinker(BaseThinker):
def __init__(self, queues, num_workers: int, num_guesses=100):
super().__init__(queues, ResourceCounter(num_workers))
self.best_result = None
self.answer = -10 # A (bad) starting guess
self.num_guesses = num_guesses
@task_submitter()
def submit_task(self):
"""Submit a new guess close to the current best whenever a node is free"""
self.queues.send_inputs(self.answer - 1 + 2 * random(), method='simulate')
@result_processor()
def store_result(self, result: Result):
"""Update best guess whenever a simulation finishes"""
assert result.success, result.failure_info
# Update the best result
if self.best_result is None or result.value > self.best_result:
self.answer = result.args[0]
self.best_result = result.value
self.rec.release() # Mark that a node is now free
# Determine if we are done
self.num_guesses -= 1
if self.num_guesses <= 0:
self.done.set()
thinker = Thinker(queues, 8)Doers describe the types of computations and available compute resources. Colmena provides Task Servers backed by several workflow engines, such as those from the ExaWorks project. Building one using Parsl requires only that your computations are expressed as Python functions:
from parsl.configs.htex_local import config # Configuration to run locally
from colmena.task_server import ParslTaskServer
# Define your function
def simulate(x: float) -> float:
return - x ** 2 + 4
# Make the Doer
doer = ParslTaskServer([simulate], queues, config)Once these are defined, launching the application involves starting both
# Launch the Thinker and doer
doer.start()
thinker.start()
# Wait until it finishes
thinker.join()
queues.send_kill_signal() # Stop the doer
# Done!
print(f'Answer: f({thinker.answer:.2f}) = {thinker.best_result:.2f}')Visit the Quickstart to learn to build a full application.
See the demo_apps to see a variety of ways to use Colmena.
Our Read-the-Docs provides the most up-to-date information about Colmena.
You can also learn more about Colmena in the papers we published about it:
- Ward et al. "Colmena: Scalable Machine-Learning-Based Steering of Ensemble Simulations for High Performance Computing". 2021 IEEE/ACM Workshop on Machine Learning in High Performance Computing Environments (MLHPC) [doi] [ArXiv] [slides] [YouTube]
This project was supported in part by the Exascale Computing Project (17-SC-20-SC) of the U.S. Department of Energy (DOE) and by DOE’s Advanced Scientific Research Office (ASCR) under contract DE-AC02-06CH11357.