Executing parfor takes more time than executing standard for loop

Hi,

I have some big data to compute and I've read Parallel Computing Toolbox would speed up some calculations. As we have access to Parallel Computing Toolbox, I though I would give it a go. I read just a few pages of a manual and tried to do some tests. Unfortunately, the time to get through standard for-loop is significantly less than using PCT. I did the same test on 2 different computers and the results are the same. I have validated that PCT can work on each computer before I did tests. Probably I am missing something but I don't know that would be. I tried different loops: less iteration with heavier computations, much more iteration with less computation etc. but always PCT take more time to get through than standard for-loop.

So what I did was I executed command

parloop

to start Parallel Computing Toolbox, waited until the toolbox got lunched and then executed this script

    clc;

    N = 10*1024*1024;
    M = 10;
    t = 0:1/N:1-1/N;
    A = zeros(N,M);

    for it = 1:10
        tic;
        for f = 1:M
            A(:,f) = sin(2*pi*f*t);
        end
        tt = toc;
        fprintf('%30s: %.2f secs\n', 'Non-parallel computing', tt);

        tic
        parfor f = 1:M
            A(:,f) = sin(2*pi*f*t);
        end
        tt = toc;
        fprintf('%30s: %.2f secs.\n', 'Parallel computing', tt);
    end

Did I misunderstand something?

Thanks

Answer 1

[Image Analyst]

You have very very small loops. So few iterations (10) that the overhead of setting up processes on different CPUs is probably more than you gain from having it be parallel. You should see improvements as your iterations get bigger, like in the millions.

[Zbigniew Koziel]

I did not realised overhead would cause that. I set the loops for 1 million iterations and parallel is 4 times faster now. Thanks.