Parallelizing python code - design/implementation questions

Philip Semanchuk philip at semanchuk.com
Sat Dec 13 21:27:50 CET 2008


On Dec 13, 2008, at 7:00 AM, stdazi wrote:

> Hello!
>
> I'm about to parallelize some algorithm that turned out to be too
> slow. Before I start doing it, I'd like to hear some suggestions/hints
> from you.

Hi stdazi,
If you're communicating between multiple processes with Python, you  
might find my IPC extensions useful. They're much less sophisticated  
than multiprocessing; they just give access to IPC semaphores and  
shared memory (no message queues yet) on Unix.

POSIX IPC:
http://semanchuk.com/philip/posix_ipc/

System V IPC:
http://semanchuk.com/philip/sysv_ipc/

More System V IPC:
http://nikitathespider.com/python/shm/

The System V IPC extensions are similar; the latter is older and  
better tested but won't be developed anymore. The former is newer, has  
a couple more features and is the future of System V IPC w/Python, at  
least as far as my work is concerned.

Good luck
Philip


>
>
> The algorithm essentially works like this:  There is a iterator
> function "foo" yielding a special kind permutation of [1,....n]. The
> main program then iterates through this permutations calculating some
> proprieties. Each time a calculation ends, a counter is incremented
> and each time the counter is divisible by 100, the current progress is
> printed.
>
> The classical idea is to spawn m threads and use some global lock when
> calling the instance of the iterator + one global lock for
> incrementing the progress counter. Is there any better way? I'm
> especially concerned with performance degradation due to locking - is
> there any way to somehow avoid it?
>
> I've also read about the `multiprocessing' module and as far as I've
> understood :
>
> ====
> permutation = foo()
> threadlst = []
> for i in xrange(m) :
>   p = Process(target=permutation.next)
>   threadlst.append(p)
>   p.start()
> for p in threadlst:
>   p.join()
> ====
>
> should do the trick. Am I right? Is there any better way other than
> this?
> --
> http://mail.python.org/mailman/listinfo/python-list




More information about the Python-list mailing list