Why is parfor so slow?

Hi,

I want to run Matlab on a cluster node with 4 processors and 48 cores (128GB ram). Typically what I do is 3D ffts. Toy example: -----------------------------------

function test(a)

N = 100;

FT = @(x)(fftshift(fftn(ifftshift(x))));

iFT = @(x)(fftshift(ifftn(ifftshift(x))));

X = randn(64,64,64);

for ii = 1:N

    X = iFT(FT(X));

end

toc(a)

end

-----------------------------------

when calling this function from a parfor loop:

-----------------------------------

N = 8

poolSize = 11;

matlabpool(poolSize)

disp(['Pool created at ',num2str(toc(a))])

for ii = 1:N

    test(a);

end

disp('Done processing')

toc(a)

matlabpool close

disp('Pool closed')

toc(a)

disp('DONE!!');

toc(a)

-----------------------------------

I almost not getting any speedup compared to running it serially on a single core. WHY? Is it reading and writing from RAM that takes all time?

Thankful for all answers, Anders

[cr]

Just a comment to make sure that not so obvious things aren't taken for granted. Is the PCT installed?

[Edric Ellis]

We can assume PCT is installed otherwise the calls to matlabpool could not succeed.

Answer 1

[Edric Ellis]

Firstly, I assume you meant to put a PARFOR loop in your example code for the 1:N loop. The reason you are not seeing any speedup is that the FFT operation is itself intrinsically multi-threaded by MATLAB. Intrinsic multi-threading is always more efficient than sending the data to the workers and executing (in a single-threaded manner) there.