|
From: | Tuomo Keskitalo |
Subject: | Re: [Help-gsl] Asynchronous multidimensional minimization |
Date: | Fri, 21 Sep 2012 21:57:16 +0300 |
User-agent: | Mozilla/5.0 (X11; Linux x86_64; rv:15.0) Gecko/20120827 Thunderbird/15.0 |
Hello,you might consider several points for parallelization, depending on your problem. If you can parallelize somewhere inside function evaluation (independent parts, if there are any), that is probably easiest. Another point might be the evaluation of the Jacobian, if you must do it numerically and if you must optimize many parameters simultaneously. And of course you can start several optimizations from different initial values at the same, if that is useful.
BR, Tuomo On 09/21/2012 05:54 PM, Maxime Boissonneault wrote:
Hi, I have a multidimensional minimization problem for which the function is pretty long to compute (think hours or days). I coded a master-slave MPI communication structure to do the work. I would have the master who dispatches sets of parameters to compute to different slaves, on different computers. What I am wondering is if there is a way to use GSL multidimensional minimization algorithms in a "queue many and wait for result" fashion rather than in a "evaluate the function sequentially" fashion. Thanks, Maxime Boissonneault
-- address@hidden http://iki.fi/tuomo.keskitalo
[Prev in Thread] | Current Thread | [Next in Thread] |