# numpy-sharedmem /

Filename Size Date modified Message
dist
doc
examples
sharedmem
tests
65 B
7.3 KB
57 B
46 B
2.3 KB
206 B
5.6 KB

# numpy-sharedmem

A collection of shared memory modules for numpy packaged together for organization and comparison purposes.

This package contains two different implementations of shared memory for numpy arrays with similar interfaces.

1. sharedmem: A shared memory module for numpy by Sturla Molden and G. Varoquaux, adapted from Robert Kern. Uses "sysv" style shared memory. Code found posted on scipy list at http://folk.uio.no/sturlamo/python/sharedmem-feb13-2009.zip See also Sturla's tutorial.

2. shmarray.py : an elegant piece of code from David Baddeley distributed under the BSD license. Uses multiprocessing to allocate shared memory. Underlying allocator is mmap. Does not allow pickling and passing shared array handles through pipes.

You might want to check Sturla's sourceforge project http://sharedmemoryarr.sourceforge.net/ thought it didn't have any content as of March 2010.

## Features/Documentation

I'm working on a table of the features of these different packages.

feature sharedmem shmarray
share array related processes yes yes
share array unrelated processes yes ?
pickle handle to shared array yes no
tested on 32-bit linux
tested on 64-bit linux yes yes
tested on 32-bit windows XP
tested on 64-bit windows 7
tested on 64-bit MacOS 10.x
tested on 32-bit MacOS 10.x

see test_sharedmem_common.py to see how to pickle the shared array and send it to another process See also http://folk.uio.no/sturlamo/python/multiprocessing-tutorial.pdf which shows examples of using ctypes and multiprocessing to create shared memory. [Note, I have placed a local copy of his tutorial in the doc/ directory of the repository]

I also have a copy of a C-types based code from Nadav Horesh adapted from other code in docs to look at.

## Requirements:

The original versions are suppose to work on unix and windows, at least in 32bit for not too big arrays. I've only done basic tests on winxp so far using mingw32 (Pythonxy 2.6 distribution) numpy 1.3 and cython 12.1

I'm now using 64-bit linux so I've made edits to Sturla's and Gael code to allow use on 64-bit. It needs to be tested back on 32-bit to make sure I haven't broken things.

The tests require multiprocessing, which is available for python 2.5 and is part of the standard library in python 2.6 and above. They are written for the nose test framework.

## Installation:

on windows, need to use mingw and assuming you are using mercurial.

Using the command line:

$hg clone https://cleemesser@bitbucket.org/cleemesser/numpy-sharedmem/$ cd numpy-sharedmem
$copy setup.cfg.template to setup.cfg # to set mingw as the compiler$ python setup.py install


I have also started putting binary distributions in the dist/ directory in case I don't have a compiler handy.

## To do

• test under linux [x] basic smoke tests done on 32bit linux (ubuntu 8.04)
• [] setup way to automate tests under linux 32 bit, 64 bit, windows xp 32bit, ?MacOS
• test larger files: Nadav Horesh mentions that it doesn't work on "large" memory regions (but still under 1GB)
• see if packaging is ok. trying out setuptools to use for development -> problem: cython build_ext doesn't seem to work with setuptools
• decide if this is the right namespace--it requires numpy so perhaps it should be added to that namespace: e.g. "import numpy.sharedmem"
• add doc on how to find allocated shared memory seg under linux and makesure they aren't leaking. e.g. ipcrm
• contact Sturla to see if he is working on this

## References and issues:

The original comment::

Ga�l Varoquaux and I did some work on that some months ago. It's not as trivial as it seems, but we have a working solution.

Basically it uses named shared memory (Sys V IPC on Unix) as buffer. The ndarray is pickled by its kernel name, the buffer is not copied. Thus you can quickly communicate shared memory ndarrays between processes (using multiprocessing.Queue).

Note that there is a pesky memory we could get rid of on Unix: It stems from multiprocessing using os._exit instead of sys.exit to terminate worker processes, preventing any clean-up code from executing. The bug is in multiprocessing, not our code. Windows refcounts shared memory segments in the kernel and is not affected. Change any occurrence of os._exit in multiprocessing to sys.exit and it will work just fine.

Well, it's not quite up to date for all 64 bits systems. I'll fix that some day.

Sturla Molden

Extended module that I used for some useful work. Comments:

1. Sturla's module is better designed, but did not work with very large (although sub GB) arrays
2. Tested on 64 bit linux (amd64) + python-2.6.4 + numpy-1.4.0
http://old.nabble.com/Multiprocessing-and-shared-memory-td25949044.html

## recent tests results

nosetests -v (on 64-bit linux, 2012-07-01):

test_sharedmem_bigarray test allocation of array 2**26 * np.uint64 size ... ok
test sharedmem.ones across common small single axis  types ... ok
test sharedmem.zeros for small single axis types ... ok
test sharedmem.zeros for arrays on the order of 2**16, single axis types ... ok
test sharedmem.zeros for arrays on the order 2**21 bytyes, single axis uint8 ... ok
test_sharedmem_common.test_two_subprocesses_no_pickle ... ok
test_sharedmem_common.test_two_subprocesses_with_pickle ... ok
test_shmarray.test_shared_ones ... ok
test_shmarray.test_shared_zeros ... ok
test_shmarray.test_KiB_shared_zeros ... ok
test_shmarray.test_MiB_shared_zeros ... ok
test_shmarray.test_two_subprocesses_no_pickle ... ok
test_shmarray.test_two_subprocesses_with_pickle ... ERROR