Why is my parfor loop running slower than a regular for loop

Hello all, I have a 3D array that takes an SVD of the first two dimensions for the size of the 3rd dimension. This seems like it would be extreemly parallelizable to me so I tried to run it in a parfor loop but for some reason the computation is like 10x slower running the parfor vs the standard for loop. Any help will be much appriciated!!

 im_vec = rand(4,4,1000);
 U = zeros(size(im_vec,1),size(im_vec,1),size(im_vec,3));
sig = zeros(size(im_vec));
V = zeros(size(im_vec,2),size(im_vec,2),size(im_vec,3));

%%% For loop time %%%
tic
for n = 1:size(im_vec,3)
        [U(:,:,n), sig(:,:,n),V(:,:,n)] = svd(im_vec(:,:,n));
end
toc

%%% Parfor loop time %%%
p = gcp('nocreate');
if isempty(p)
   parpool;
end
tic
parfor n = 1:size(im_vec,3)
        [U(:,:,n), sig(:,:,n),V(:,:,n)] = svd(im_vec(:,:,n));
end
toc

[Walter Roberson]

On my system, if I expand the work by a factor of 100, then the regular for loop takes about 0.58 seconds and the parfor loop takes about 0.22 seconds the first time (and about 0.15 seconds after that.)

Answer 1

[Christine Tobler]

The problem is likely that the cost of each iteration is still too small to warrant the start-up time of the parfor loop.

You could take a look at the pagesvd function, which applies the SVD to each page A(:, :, i) of the input array, just like what you're doing here. It uses threading on a lower level when the size of the array means we expect the threading to be worthwhile.