Broadcast Variable Warning in MATLAB Parallel Computing Toolbox

When I ran a MATLAB code using MATLAB Parallel Computing Toolbox, I got messages like:

The entire array or structure 'a' is a broadcast variable. This might result in unnecessary communication overhead.

The entire array or structure 'b' is a broadcast variable. This might result in unnecessary communication overhead.

The entire array or structure 'L' is a broadcast variable. This might result in unnecessary communication overhead.

I then got lost. The code is still running though, but I don't know what to do exactly to avoid the messages above. Below is a description of the main part of my code:

a = (5:5:100)'; b = (-0.4:0.1:2)'; L = (0.25:0.25:10)'; la = length(a); lb = length(b); lL = length(L);
parfor jj = 1:2
    err1 = zeros(la,lb,lL);
    for i = 1:la
        if jj == 1, bla bla bla; else, bla bla bla; end
        for j = 1:lb
            for k = 1:lL
                err1(i,j,k) = methods(a(i),b(j),L(k));
            end
        end
    end
end

Answer 1

[Matt J]

In this case, the warnings can probably be ignored because a, b, and L are very small.

However, I don't really understand the parallelization stratgy, since the outer parfor loop is very short, and you seem to be doing almost nothing with the loop variable jj. If the intensive work is being done by methods(), I would think you would want to paralleilize over (i,j,k).

[Kareem Elgindy]

Thanks @Matt J. The intensive work is on methods, but methods is running for two values of jj. One for jj = 1 and one for jj = 2. For each case, certain parameters are passed to methods but they are irrelevant to i, j, and k.

[Matt J]

Your posted code doesn't show that. Nothing jj-dependent is being passed to methods(). I hope you are not doing it with global variables.

Regardless, it still doesn't explain why you would be looping over jj instead of (i,j,k). Why not as below? That way, assuming you have more than 2 cores, you can parallelize across more than 2 parpool workers.

a = (5:5:100)'; b = (-0.4:0.1:2)'; L = (0.25:0.25:10)'; 
la = length(a); lb = length(b); lL = length(L);

err1 = zeros(la,lb,lL);

parfor m=1:la*lb*lL
    
    [i,j,k]=ind2sub([la,lb,lL],m);
    
    
    for jj = 1:2
        
        if jj == 1, bla bla bla; else, bla bla bla; end
        
        err1(m) = methods(a(i),b(j),L(k));
        
    end
 
end

[Kareem Elgindy]

@Matt J Thanks again for your time and efforts. Don't worry, I'm not using global variables. I removed the parameters passed to methods from the presented code above to keep it simple and short.

Your last code looks quite interesting. Now, that we calculate err1(m) for each m, how could we recover err1(i,j,k) back?

[Walter Roberson]

reshape(err1, la, lb, lL)

[Matt J]

Your last code looks quite interesting. Now, that we calculate err1(m) for each m, how could we recover err1(i,j,k) back?

I'm not sure what you mean by "recover" it? err1(m) is the same thing as err1(i,j,k). There shouldn't be any need to reshape() or do anything else.

[Kareem Elgindy]

Hi @Matt J. I still need err1 as a matrix of 3 dimensions to do further postprocessing using commands like squeeze along the i dimension, etc. So, reshape here remains useful as suggested by @Walter Roberson.

[Kareem Elgindy]

@Walter Roberson... there is a missing comma between la and lb in the command reshape; right?

[Matt J]

I still need err1 as a matrix of 3 dimensions to do further postprocessing using commands like squeeze along the i dimension, etc. So, reshape here remains useful as suggested by

It should have been 3-dimensional from the beginning and remained so throughout. The shape was established in the declaration line,

err1 = zeros(la,lb,lL);

[Kareem Elgindy]

@Matt J You are right. The size is already predefined. Thanks.