While using parfor the cpu usage drops by 1 core every something iterations

While using parfor loop for a long time everything seems to work fine loading all 8 cores of my CPU but after sometime Matlab releases cores one by one with some time in between exponentially increasing time working on the loop.

Triple for loop I use:

N = size(A,1); % around 5000

parfor i = 1:size(A,1)

      for j = i+1:N

          for k = j+1:N

              sumMatrix = [sumMatrix; [(A(i) + A(j) + A(k)) A(i) A(j) A(k)]];

          end
      end
  end

Answer 1

[Jan]

An iterative growing of an array is a very bad idea because it wastes resources. I guess that the cores are simply waiting for the OS to deliver RAM. Look for the term "pre-allocation".

[Stepan Ulyanin]

Well by my rough estimates I would need an array of 4 999 850 001 x 4 for just 100000 elements in vector A. I am not sure how to pre-allocate memory for such array?

[Walter Roberson]

If you cannot pre-allocate that much memory then you cannot store the result, which is just over 149 gigabytes, and your execution would eventually crash anyhow.

Your code as shown is not valid in a parfor loop as the results are dependent on the order of execution. Each output of a parfor must have a fixed output location.

[Edric Ellis]

@Walter - actually, that's not quite right - in this case, sumMatrix is treated as a parfor reduction variable. The following loop works just fine:

reduction = [];
parfor idx = 1:10
    for jdx = 1:10
        for kdx = 1:10
            reduction = [reduction; rand];
        end
    end
end