parallel computing of optimization problem

Hi all,

I am trying to parallelize my optimization code. Below is my original code

   options=optimset('TolX',1e-8,'TolFun',1e-8);  
   [x,resnorm,residual,exitflag] = lsqnonlin(@objfun,x0,[],[],options);

To implement parallel computing, I rewrite it into

   matlabpool open 2
   options=optimset('TolX',1e-8,'TolFun',1e-8,'UseParallel', 'always');  
   [x,resnorm,residual,exitflag] = lsqnonlin(@objfun,x0,[],[],options);
   matlabpool close

After running matlab, I saw the following in the command window:

   Starting matlabpool using the 'local' profile ... connected to 2 labs.

I have tested np=2, 8, 12, but the time cost does not change with the number of processors. So is there anything I missed in my code? Thanks.

Answer 1

[Walter Roberson]

Explanation here: http://www.mathworks.co.uk/matlabcentral/answers/105573-about-parallel-computing-toolbox#comment_179397

The below may assist in understanding or in improving performance

http://www.mathworks.co.uk/matlabcentral/answers/83207-overhead-when-using-optimization-parfor-loops-or-useparallel-always-for-fmincon

http://www.mathworks.co.uk/matlabcentral/answers/41427-parfor-loop-how-to-calculate-its-transmission-limit

http://www.mathworks.co.uk/matlabcentral/answers/55381-can-t-get-speed-up#answer_67259

http://www.mathworks.co.uk/matlabcentral/answers/75629-why-code-using-parallel-processing-has-longer-running-time-than-the-other

[Likun]

Hi Walter,

Thanks for your answer. I notice that lsqnonlin does not support parallel computing, and that is probably the reason why changing number of processes does not influence computer performance.

For my problem, lsqnonlin gives best performance in serial computing. From profile report, the computation of my objective function (after vectorization) takes about 25% of the total time, and the rest of the time accounts for optimization. Since parallelization of lsqnonlin is not possible, could you please give any advice on speeding up my code further?

Thank you, Likun

[Walter Roberson]

Did you try the parallel profiler to see what it thinks the bottlenecks are in going parallel ? http://www.mathworks.com/help/distcomp/profiling-parallel-code.html

[Likun]

Thank you very much. Likun

[Walter Roberson]

I realized after I posted last night that the difficulty is that the optimizer code itself is not parallelized.

I am reading through a document to see if I can find some hints on performing this task more efficiently; it might take me a couple of days to get through.