Why is this simple parallel program much slower than the non-parallel version?

I have a very simple script that calls the built-in genetic algorithm function:

function test1(gen)
    options = gaoptimset('UseParallel', 'always', 'Vectorized', 'off');
    tic;
    x = ga(@dejong5fcn, 2, [], [], [], [], [], [], [], options);
    toc
end

First, I ran test1 without starting matlabpool. As expected, it runs fine but uses only one CPU core as observed with Windows Resource Monitor. It takes 4.2 seconds to run 20020 fitness evaluations. Then, I started the parallel engine with: "start matlabpool local 4" and then performed an otherwise identical run of test1. It runs and uses all four CPU cores, but takes about 90.7 seconds to perform 20020 fitness evaluations.

What am I not understanding about parallelism in Matlab R2012a (on Windows 7 64 bit)? Thanks for any help.

Answer 1

[Walter Roberson]

http://www.mathworks.co.uk/matlabcentral/answers/13404-serial-and-parallel#answer_18308

[Andy Yancy]

Thanks for your reply. I changed the objective function to a much more computationally-intensive one to reduce the effects of parallel processing overhead, and now the parallel version runs faster than the non-parallel version.

In fact, I have the opposite problem now: the parallel version is about 20x as fast as the non-parallel version. How is it possible for that to happen? Wouldn't 4x be the theoretically maximum possible speedup with four cores? Thanks for your help.

[Andy Yancy]

Anyone know about this?

[Walter Roberson]

Pr-allocation. When you use parfor(), MATLAB automatically pre-allocates for the outputs. If you have not pre-allocated in your code, the serial version could spend most of its time finding new memory and copying the results into it.

[Andy Yancy]

Great, thanks Walter. I appreciate you taking the time to answer my questions.