Can I avoid this broadcast variable?

Hello,

I had two nested for loops that I would like to parallelise. It looked like this

for k = 1:size(S,2)
    for j = 1:length(b)
        fun(S(:,k),b(j));
    end
end

Here S is large (10^5 x 10^2), and b is not large (1 x 10).

Since all of the loops are independent I found the following solution

parfor i = 1:size(S,2)*length(b)
   k = ceil(i/length(b)); %S
   j = i-(k-1)*length(b); %b
       
   fun(S(:,k),b(j));
end

The problem is that S and b are broadcast variables.

Is there a way to eliminate S as a broadcast variable (b does not matter much)?

Thank you!

Answer 1

[Matt J]

Why not this?

parfor k = 1:size(S,2)
    tmp=S(:,k);
    for j = 1:length(b)
        fun(tmp,b(j));
    end
end

[Adam]

Hi Matt, because this way the computational time will be longer. My solution increases the number of CPUs length(b)-fold and yields overall shorter computational time. Am I correct?

[Matt J]

You would need a parpool with N>100=size(S,2) labs in order for the work per lab to be any less in your strategy than what I propose.

Method 1: parallelize over k only. The busiest lab will perform ceil(size(S,2)/N)*length(b) serial calls to fun.

Method 2: parallelize all calls to fun. The busiest lab will perform ceil(size(S,2)*length(b)/N) serial calls.

Both are therefore nearly the same unless N>size(S,2)