Skip to content
Distributed process pool for Python
Branch: master
Clone or download
Latest commit c1f5639 Mar 17, 2019
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
distex eventkit integration Mar 16, 2019
docs Fix RTD build Mar 17, 2019
tests eventkit integration Mar 16, 2019
.gitignore flake8 compliant code formatting Jan 15, 2019
.readthedocs.yml Fix RTD build Mar 17, 2019
.travis.yml Test with uvloop Jan 15, 2019
LICENSE Year update Jan 15, 2019
README.rst Doc URL set to RTD Mar 17, 2019
dist.sh flake8 compliant code formatting Jan 15, 2019
requirements.txt eventkit integration Mar 16, 2019
setup.py version 0.6.0 Mar 16, 2019

README.rst

Build PyPi Documentation

Introduction

Distex offers a distributed process pool to utilize multiple CPUs or machines. It uses asyncio to efficiently manage the worker processes.

Features:

  • Scales from 1 to 1000's of processors;
  • Can handle in the order of 50.000 small tasks per second;
  • Easy to use with SSH (secure shell) hosts;
  • Full async support;
  • Maps over unbounded iterables;
  • Compatible with concurrent.futures.ProcessPool (or PEP3148).

Installation

pip3 install -U distex

When using remote hosts then distex must be installed on those too. Make sure that the distex_proc script can be found in the path.

For SSH hosts: Authentication should be done with SSH keys since there is no support for passwords. The remote installation can be tested with:

ssh <host> distex_proc

Dependencies:

  • Python version 3.6 or higher;
  • On Unix the uvloop package is recommended: pip3 install uvloop
  • SSH client and server (optional).

Examples

A process pool can have local and remote workers. Here is a pool that uses 4 local workers:

from distex import Pool

def f(x):
    return x*x

pool = Pool(4)
for y in pool.map(f, range(100)):
    print(y)

To create a pool that also uses 8 workers on host maxi, using ssh:

pool = Pool(4, 'ssh://maxi/8')

To use a pool in combination with eventkit:

from distex import Pool
import eventkit as ev
import bz2

pool = Pool()
# await pool  # un-comment in Jupyter
data = [b'A' * 1000000] * 1000

pipe = ev.Sequence(data).poolmap(pool, bz2.compress).map(len).mean().last()

print(pipe.run())  # in Jupyter: print(await pipe)
pool.shutdown()

There is full support for every asynchronous construct imaginable:

import asyncio
from distex import Pool

def init():
    # pool initializer: set the start time for every worker
    import time
    import builtins
    builtins.t0 = time.time()

async def timer(i=0):
    # async code running in the pool
    import time
    import asyncio
    await asyncio.sleep(1)
    return time.time() - t0

async def ait():
    # async iterator running on the user side
    for i in range(20):
        await asyncio.sleep(0.1)
        yield i

async def main():
    async with Pool(4, initializer=init, qsize=1) as pool:
        async for t in pool.map_async(timer, ait()):
            print(t)
        print(await pool.run_on_all_async(timer))


loop = asyncio.get_event_loop()
loop.run_until_complete(main())

High level architecture

Distex does not use remote 'task servers'. Instead it is done the other way around: A local server is started first; Then the local and remote workers are started and each of them will connect on its own back to the server. When all workers have connected then the pool is ready for duty.

Each worker consists of a single-threaded process that is running an asyncio event loop. This loop is used both for communication and for running asynchronous tasks. Synchronous tasks are run in a blocking fashion.

When using ssh, a remote (or 'reverse') tunnel is created from a remote Unix socket to the local Unix socket that the local server is listening on. Multiple workers on a remote machine will use the same Unix socket and share the same ssh tunnel.

The plain ssh executable is used instead of much nicer solutions such as AsyncSSH. This is to keep the CPU usage of encrypting/decrypting outside of the event loop and offload it to the ssh process(es).

Documentation

Distex documentation

author:Ewald de Wit <ewald.de.wit@gmail.com>
You can’t perform that action at this time.