I'm not aware that we've encountered this, but I also don't think we've ever tried.
Just to clarify, the issue is that the answer changes between processes=1 or processes=2+? I don't think that any of FiPy's parallel code (e.g., using mpirun to invoke PyTrilinos in parallel) should be triggered by this use of multiprocessing, but we've never tried. I'll try to do some experiments. If you'd like to send me a complete (non)working example, I'll be happy to take a look at it. On Mar 13, 2015, at 2:46 PM, Tony Bartolotta <[email protected]> wrote: > FiPy Developers, > > I am currently using FiPy 3.1 with Python 2.7 on a laptop running Ubuntu > 14.04. I am attempting to find the impulse responses of a Fokker-Planck > equation. In the below snippets of code, funcCall(args) calculate an impulse > response with FiPy using a non-parallelized solver. > > > Serial Calls > > for k in range(N): > responses[k,:] = funcCall(args) > > > Parallel Calls > > pool = multiprocessing.Pool(processes=1) > responses = pool.map(funcCall, [args for k in range(N)]) > pool.close() > pool.join() > > > I am finding that FiPy returns different numerical results when I call this > method in serial or in parallel. Strangely, even if the parallel calls are > carried out by only one worker using the same solver, the results are still > different. > > Using a debugging tool, all arguments appear to be correctly passed to FiPy. > After considering a few test cases, the serial calls appear to arrive at the > correct numerical solutions while the parallel version does not. I am unsure > what is causing this behavior. Are you aware of any potential cause of this > behavior? > > If necessary, I can post a more complete version of the code. > > Thank you, > Tony > > _______________________________________________ > fipy mailing list > [email protected] > http://www.ctcms.nist.gov/fipy > [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ] _______________________________________________ fipy mailing list [email protected] http://www.ctcms.nist.gov/fipy [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ]
