MPIRE (MultiProcessing Is Really Easy)
MPIRE, short for MultiProcessing Is Really Easy, is a Python package for multiprocessing, but faster and more user-friendly than the default multiprocessing package. It combines the convenient map like functions of multiprocessing.Pool with the benefits of using copy-on-write shared objects of multiprocessing.Process, together with easy-to-use worker state, worker insights, and progress bar functionality.
Full documentation is available at https://slimmer-ai.github.io/mpire/.
Features
- Faster execution than other multiprocessing libraries. See benchmarks.
- Intuitive, Pythonic syntax
- Multiprocessing with
map/map_unordered/imap/imap_unorderedfunctions - Easy use of copy-on-write shared objects with a pool of workers
- Each worker can have its own state and with convenient worker init and exit functionality this state can be easily manipulated (e.g., to load a memory-intensive model only once for each worker without the need of sending it through a queue)
- Progress bar support using tqdm
- Progress dashboard support
- Worker insights to provide insight into your multiprocessing efficiency
- Graceful and user-friendly exception handling
- Automatic task chunking for all available map functions to speed up processing of small task queues (including numpy arrays)
- Adjustable maximum number of active tasks to avoid memory problems
- Automatic restarting of workers after a specified number of tasks to reduce memory footprint
- Nested pool of workers are allowed when setting the
daemonoption - Child processes can be pinned to specific or a range of CPUs
- Optionally utilizes dill as serialization backend through multiprocess, enabling parallelizing more exotic objects, lambdas, and functions in iPython and Jupyter notebooks.
Installation
Note
MPIRE currently only supports Linux based operating systems that support 'fork' as start method. Support for Windows is coming soon.
Through pip (PyPi):
pip install mpire
From source:
python setup.py install
Getting started
Suppose you have a time consuming function that receives some input and returns its results. Simple functions like these are known as embarrassingly parellel problems, functions that require little to no effort to turn into a parellel task. Parallelizing a simple function as this can be as easy as importing multiprocessing and using the multiprocessing.Pool class:
import time
from multiprocessing import Pool
def time_consuming_function(x):
time.sleep(1) # Simulate that this function takes long to complete
return ...
with Pool(processes=5) as pool:
results = pool.map(time_consuming_function, range(10))
MPIRE can be used almost as a drop-in replacement to multiprocessing. We use the mpire.WorkerPool class and call one of the available map functions:
from mpire import WorkerPool
with WorkerPool(n_jobs=5) as pool:
results = pool.map(time_consuming_function, range(10))
The differences in code are small: there's no need to learn a completely new multiprocessing syntax, if you're used to vanilla multiprocessing. The additional available functionality, though, is what sets MPIRE apart.
Progress bar
Suppose we want to know the status of the current task: how many tasks are completed, how long before the work is ready? It's as simple as setting the progress_bar parameter to True:
with WorkerPool(n_jobs=5) as pool:
results = pool.map(time_consuming_function, range(10), progress_bar=True)
And it will output a nicely formatted tqdm progress bar. In case you're running your code inside a notebook it will automatically switch to a widget.
MPIRE also offers a dashboard, for which you need to install additional dependencies. See Dashboard for more information.
Shared objects
If you have one or more objects that you want to share between all workers you can make use of the copy-on-write shared_objects option of MPIRE. MPIRE will pass on these objects only once for each worker without copying/serialization. Only when you alter the object in the worker function it will start copying it for that worker.
def time_consuming_function(some_object, x):
time.sleep(1) # Simulate that this function takes long to complete
return ...
def main():
some_object = ...
with WorkerPool(n_jobs=5, shared_objects=some_object) as pool:
results = pool.map(time_consuming_function, range(10), progress_bar=True)
See shared_objects for more details.
Worker initialization
Workers can be initialized using the worker_init feature. Together with worker_state you can load a model, or set up a database connection, etc.:
def init(worker_state):
# Load a big dataset or model and store it in a worker specific worker_state
worker_state['dataset'] = ...
worker_state['model'] = ...
def task(worker_state, idx):
# Let the model predict a specific instance of the dataset
return worker_state['model'].predict(worker_state['dataset'][idx])
with WorkerPool(n_jobs=5, use_worker_state=True) as pool:
results = pool.map(task, range(10), worker_init=init)
Similarly, you can use the worker_exit feature to let MPIRE call a function whenever a worker terminates. You can even let this exit function return results, which can be obtained later on. See the worker_init and worker_exit section for more information.
Worker insights
When you're multiprocessing setup isn't performing as you want it to and you have no clue what's causing it, there's the worker insights functionality. This will give you insight in your setup, but it will not profile the function you're running (there are other libraries for that). Instead, it profiles the worker start up time, waiting time and working time. When worker init and exit functions are provided it will time those as well.
Perhaps you're sending a lot of data over the task queue, which makes the waiting time go up. Whatever the case, you can enable and grab the insights using the enable_insights flag and mpire.WorkerPool.get_insights function, respectively:
with WorkerPool(n_jobs=5) as pool:
results = pool.map(time_consuming_function, range(10), enable_insights=True)
insights = pool.get_insights()
See worker insights for a more detailed example and expected output.
Documentation
See the full documentation at https://slimmer-ai.github.io/mpire/ for information on all the other features of MPIRE.
If you want to build the documentation yourself, please install the documentation dependencies by executing:
pip install mpire[docs]
or
pip install .[docs]
Documentation can then be build by executing:
python setup.py build_docs
Documentation can also be build from the docs folder directly. In that case MPIRE should be installed and available in your current working environment. Then execute:
make html
in the docs folder.
573 Dec 27, 2022
3.8k Jan 8, 2023
5k Jan 7, 2023
102 Jan 6, 2023
8 Jul 27, 2022
1.7k Jan 5, 2023
9 Jul 4, 2022
1.4k Jan 1, 2023
7.1k Jan 1, 2023
4.1k Jan 3, 2023
1.3k Dec 28, 2022
0 Nov 9, 2022
229 Jan 2, 2023
67 Dec 31, 2022
1 Nov 10, 2021
1.7k Jan 3, 2023
0 Nov 6, 2021
2 Oct 31, 2021