Send for loop iterations to multiprocessors

Hello, My code consists of two parts, an outer portion that is independent and an inner portion that is dependent. I want to send the outer for loop to different processors to speed it up, but the inner loop needs to be run serially. At the end, I need to combine the results of the independent loop into one matrix.

Essentially the code can be visualized as for ii = 1:4 for i = 1:1000 a(i) = f(b) end A(ii) = a(i) end

If this makes sense, the ii loop is independent but the i loop is dependent. I want to send the ii loop to different processors so that i can be computed serially. I have not had very much luck with parfor in this instance.

Is what I am asking not possible? I estimate that if I cannot use multiple processors to execute the code, it will run for 1.5-2 days.

Thank you for any insight, Dan Stanaway

[Oleg Komarov]

You did preallocate right (if you don't have 2011a...)

[Sean de Wolski]

What do you mean by "I did not have very much luck"?

[Daniel Stanaway]

I meant that I was unable to index the variables that need to be compiled in a way that would allow the parfor loop to execute. The errors I am receiving are 1) Parfor loop cannot be run dur to way variable is used, and 2) valid indices for variables are restricted in parfor loops.

Answer 1

[Jason Ross]

Try putting your inner loop in a function and calling the function instead.

e.g.:

parfor ii = 1:4
somethingout = myloopfunction
end

then in myloopfunction you have the inner for loop.

If that doesn't work, you might also want to look into the spmd functionality.

[Daniel Stanaway]

Thanks Jason, even in using the function, I have an error saying value is non deterministic, but I think the spmd functionality is what I am looking for. I appreciate your answer.