Matlab parallel computing time effiency

Currently I've started to work with matlab's parallel computing toolbox. I wanted to test it with a code for face recognition. My first try was with the Eigenfaces function using 10.000 images:

Script:

    function [m, A, Eigenfaces] = EigenfaceCore(T)
    tic
    Train_Number = size(T,2);
    A = [];  

    for i = 1 : Train_Number
    temp = double(T(:,i)) - m; 
    A = [A temp,]; % Merging all centered images
    end

    L = A'*A;
    [V D] = eig(L); 

    x = zeros(1,size(D,1));
    for i = 1 : size(D,1)
    x (i) = D(i,i);
    end

    L_eig_vec = [];
    for i = 1 : size(V,2) 
        if( x(i) > 1)
            L_eig_vec = [L_eig_vec V(:,i)];
        end
    end

    Eigenfaces = A * L_eig_vec; % A: centered image vectors
    toc

For this code I need something that takes about 112 seconds.

If I open a POOL with 4 kernels, and use `parfor` instead of `for`, it takes 421 seconds. Can somebody help me? I thought by increasing the number of images (first it was something about 20 Images and parfor needing more time) the parallel computing version will pass the normal version.

Answer 1

[Walter Roberson]

eig() is multi-threaded and will use all of the available CPUs. Each worker only gets one CPU in current versions of MATLAB. Therefore each eig() call will individually take longer. However, since you would be doing more of them at the same time, the total time should be roughly the same.

If your pool of 4 workers is based upon 2 cores with hyperthreading then you would expect to roughly double your execution time when you do numeric calculations. hyperthreading is only efficient when one thread is waiting for interrupts or I/O, then the other thread can steal the CPU.

This could account for a doubling of execution time, but beyond doubling the reasons typically depend upon a lot of data transfer to the workers, which is extra work not needed in a plain for loop.

I notice you have

    A = [];  

    for i = 1 : Train_Number
    temp = double(T(:,i)) - m; 
    A = [A temp,]; % Merging all centered images
    end

You should really re-write that to pre-allocate A and assign to the appropriate row or column of it.

[Kerem Ciftci]

Hey Walter,

okay so this Code is not well suited for parallel computing? Hmm sadly, I just wanted to try to make it faster with parallel Toolbox.

Answer 2

[Matt J]

You have a lot of unnecessary and expensive loops. Here's a loop-free version that should run much faster,

   function [m, A, Eigenfaces] = EigenfaceCore(T)

       A=double(T)-m;

        L = A'*A;

        [V D] = eig(L); 

         x=diag(D);

        L_eig_vec = V(:,x>1);

       Eigenfaces = A * L_eig_vec; % A: centered image vectors