Cluster Futures

This module provides a Python concurrent.futures executor that lets you run functions on remote systems in your HTCondor or Slurm cluster. Stop worrying about writing job files, scattering/gathering, and serialization---this module does it all for you.

It uses the cloudpickle library to allow (most) closures to be used transparently, so you're not limited to "pure" functions.

Installation:

pip install clusterfutures

Usage:

import cfut
def square(n):
    return n * n

with cfut.SlurmExecutor() as executor:
    for result in executor.map(square, [5, 7, 11]):
        print(result)

See slurm_example.py and condor_example.py for further examples. The easiest way to get started is to ignore the fact that futures are being used at all and just use the provided map function, which behaves like itertools.imap but transparently distributes your work across the cluster.

Goals & design

clusterfutures is a simple wrapper to run Python functions in batch jobs on an HPC cluster. Each future corresponds to one batch job. The functions that you run through clusterfutures should normally run for at least a few seconds each: running smaller functions will be inefficient because of the overhead of launching jobs and moving data.

Functions, parameters and return values are sent by creating files; this assumes that the control process and the worker nodes have a shared filesystem. This mechanism is convenient for relatively small amounts of data; it's probably not the best way to transfer large amounts of data to & from workers.