using variables in a parfor loop

Hey, I am very new to the parallel computing and i got some trouble.

So basicly i have so example code very similar to this one:

a = (normrnd(1,1,[100000,1]))+5;

i=0;

while i<=10000
    for j=1:length(a)
        
        a(j)=a(j)+1/a(j);
    end    
    
   i=i+1; 
end

Inside the nested for loop are many more alculations which takes a huge amount of time, so i basically want to cut a into as much pieces as i have physical cores, so in this case 4 *20000 and then do the same calculations of thos 20000 arrays parrallel.

But I direclty get some error cause i cannot use my input variable from outside the parfor loop because of classification (Code without the outside while loop)

parfor i=1:cores
   
    v=d(:,i);
    v=v(v>0);
    lim=size(v,1);
    
    for j=1:lim
        v(j)=v(j)+1/v(j);
    end
    
    d(1:length(v),i)=v;
    
end

Shouldnt my input d (20000x4 array) not work as an broadcast variable as defined in https://de.mathworks.com/help/parallel-computing/troubleshoot-variables-in-parfor-loops.html ?

Or am I missanderstanding something very basic in parallel computing?

Many thanks in advance

Best regards

Answer 1

[Daniel M]

I think it's because you're using d as the source and the sink in your loop, and you're assigning to d in a way that can make practically anything happen to the size of d, on different iterations. For example, what if on the first iteration d(:,i) contains no zeros, but on the second one it contains all zeros. Then 1:length(v) is changing every iteration and makes assignment to d impossible to determine a priori.

[Walter Roberson]

If so then,

parfor i=1:cores
   
    v=d(:,i);
    vc = v;
    v=v(v>0);
    
    v = v + 1./v;   %optimized out loop
    
    vc(1:length(v)) = v;
    d(:,i)=v;
    
end

Note that if your non-zeros are not consecutive that this can result in entries that were 0 being populated. [1 2 0 3] would be replaced with [1+1/1, 2+1/2, 3+1/3, 3] for example. It is not obvious that is what you would want.

[Marc Laub]

d can contain zeros and if so its only the last element, in cases when my total number in a cannot be eveny distributed on my workers without reminder. So in the case of 101 elements i want to calculate 26 on worker 1 and 25 on the other 3.

It is not important in which order the 25 or 26 elements in each arrays are calculatet, but it is very important taht i know a prior wihc elements are calculated by which worker.

So i want to be basically be like this:

num_workers =4;

a = (normrnd(1,1,[100000,1]))+5;

i=0;

while i<=10000
    
    b=randperm(length(a),length(a));
    a=a(b);
    d=reshape(a,[25000,num_workers]);
    
    for k=1:num_workers
        c=d(:,k);
        for j=1:length(c)
            
            c(j)=c(j)+1/c(j);
        end
        
        d(:,k)=c;
    end
    
    a=reshape(d,numel(d),1);
    
   i=i+1; 
end

but with the inner for loop k=1:4 run parallel instead of serial since the 4 arrays in d are independend.

How is this possible?

Edit: I think spmd is what i should be looking for but for beginners the examples or help is not very detailed i think.