esi-acme 2025.6

Asynchronous Computing Made ESI

ACME: Asynchronous Computing Made ESI

main:

dev:

Table of Contents

1. Summary
2. Installation
3. Usage
   • Basic Examples
   • Intermediate Examples
   • Advanced Use
4. Handling Results
   • Load Results From Files
   • Collect Results in Single HDF5 Dataset
   • Collect Results in Local Memory
5. Debugging
6. Documentation and Contact
   • Resources

Summary

The objective of ACME (pronounced "ak-mee") is to provide easy-to-use wrappers for calling Python functions concurrently ("embarassingly parallel workloads"). ACME is developed at the Ernst Strüngmann Institute (ESI) gGmbH for Neuroscience in Cooperation with Max Planck Society and released free of charge under the BSD 3-Clause "New" or "Revised" License. ACME relies heavily on the concurrent processing library dask and was primarily designed to facilitate the use of SLURM on the ESI HPC cluster (although other HPC infrastructure running SLURM can be leveraged as well). Local multi-processing hardware (i.e., multi-core CPUs) is fully supported too. ACME is itself used as the parallelization engine of SyNCoPy.

Installation

ACME can be installed with pip

pip install esi-acme

or via conda

conda install -c conda-forge esi-acme

To get the latest development version, simply clone our GitHub repository:

git clone https://github.com/esi-neuroscience/acme.git
cd acme/
pip install -e .

Usage

Basic Examples

Simplest use, everything is done automatically.

from acme import ParallelMap

def f(x, y, z=3):
  return (x + y) * z

with ParallelMap(f, [2, 4, 6, 8], 4) as pmap:
  pmap.compute()

See also our Quickstart Guide.

Intermediate Examples

Set number of function calls via n_inputs

import numpy as np
from acme import ParallelMap

def f(x, y, z=3, w=np.zeros((3, 1)), **kwargs):
    return (sum(x) + y) * z * w.max()

pmap = ParallelMap(f, [2, 4, 6, 8], [2, 2], z=np.array([1, 2]), w=np.ones((8, 1)), n_inputs=2)

with pmap as p:
  p.compute()

More details in Override Automatic Input Argument Distribution

Advanced Use

Allocate custom client object and recycle it for several computations (use slurm_cluster_setup on non-ESI HPC infrastructure or local_cluster_setup when working on your local machine)

import numpy as np
from acme import ParallelMap, esi_cluster_setup

def f(x, y, z=3, w=np.zeros((3, 1)), **kwargs):
    return (sum(x) + y) * z * w.max()

def g(x, y, z=3, w=np.zeros((3, 1)), **kwargs):
    return (max(x) + y) * z * w.sum()

n_workers = 200
client = esi_cluster_setup(partition="8GBXS", n_workers=n_workers)

x = [2, 4, 6, 8]
z = range(n_workers)
w = np.ones((8, 1))

pmap = ParallelMap(f, x, np.random.rand(n_workers), z=z, w=w, n_inputs=n_workers)
with pmap as p:
    p.compute()

pmap = ParallelMap(g, x, np.random.rand(n_workers), z=z, w=w, n_inputs=n_workers)
with pmap as p:
    p.compute()

For more information see Reuse Worker Clients

Handling Results

Load Results From Files

By default, results are saved to disk in HDF5 format and can be accessed using the results_container attribute of ParallelMap:

def f(x, y, z=3):
  return (x + y) * z

with ParallelMap(f, [2, 4, 6, 8], 4) as pmap:
  filenames = pmap.compute()

Example loading code:

import h5py
import numpy as np
out = np.zeros((4,))

with h5py.File(pmap.results_container, "r") as h5f:
  for k, key in enumerate(h5f.keys()):
    out[k] = h5f[key]["result_0"][()]

See also Where Are My Results?

Collect Results in Single HDF5 Dataset

If possible, results can be slotted into a single HDF5 dataset using the result_shape keyword (None denotes the dimension for stacking results):

def f(x, y, z=3):
  return (x + y) * z

with ParallelMap(f, [2, 4, 6, 8], 4, result_shape=(None,)) as pmap:
  pmap.compute()

Example loading code:

import h5py

with h5py.File(pmap.results_container, "r") as h5f:
  out = h5f["result_0"][()] # returns a NumPy array of shape (4,)

Datasets support "unlimited" dimensions that do not have to be set a priori (use np.inf in result_shape to denote a dimension of arbitrary size)

# Assume only the channel count but not the number of samples is known
nChannels = 10
nSamples = 1234
mock_data = np.random.rand(nChannels, nSamples)
np.save("mock_data.npy", mock_data)

def mock_processing(val):
  data = np.load("mock_data.npy")
  return val * data

with ParallelMap(mock_processing, [2, 4, 6, 8], result_shape=(None, nChannels, np.inf)) as pmap:
  pmap.compute()

with h5py.File(pmap.results_container, "r") as h5f:
  out = h5f["result_0"][()] # returns a NumPy array of shape (4, nChannels, nSamples)

More examples can be found in Collect Results in Single Dataset

Collect Results in Local Memory

This is possible but not recommended.

def f(x, y, z=3):
  return (x + y) * z

with ParallelMap(f, [2, 4, 6, 8], 4, write_worker_results=False) as pmap:
  result = pmap.compute() # returns a 4-element list

Alternatively, create an in-memory NumPy array

with ParallelMap(f, [2, 4, 6, 8], 4, write_worker_results=False, result_shape=(None,)) as pmap:
  result = pmap.compute() # returns a NumPy array of shape (4,)

Debugging

Use the debug keyword to perform all function calls in the local thread of the active Python interpreter

def f(x, y, z=3):
  return (x + y) * z

with ParallelMap(f, [2, 4, 6, 8], 4, z=None) as pmap:
    results = pmap.compute(debug=True)

This way tools like pdb or %debug IPython magics can be used. More information can be found in the FAQ.

Documentation and Contact

To report bugs or ask questions please use our GitHub issue tracker. More usage details and background information is available in our online documentation.

Resources

• ACME Presentation at deRSE23 - Conference for Research Software Engineering in Germany
• ACME Demo presented at the 4th annual Data Scientist Community Meeting
• ACME Tutorials
• ACME FAQ