parfor problem (broadcast variable)

Hi, I'm trying to use "parfor" in my MATLAB script.

It goes well without error, but It shows no significant speed boost compared to for loop.

And I got this warning message:

"the entire array ('A_Data' , 'A_ref') is a broadcast variable. This might result in unnecessary communication overhead."

How can I deal with this problem?

Note that:

A_Data: 1710203x5 double

A_ref: 5760x5 double

Please suggest the solutions.

% load 'A_Data' and 'A_ref' matrix
load dataq.mat 

% preallocation ('theta' ,'b_ref')
k=length(A_ref);
theta=zeros(k,5);
b_ref=zeros(k,1);

% regression
poolobj = parpool(8);

parfor i=1:k
 
     W = sqrt( ...
        exp( ...
        -( ...
        ( ( A_Data(:,2)-A_ref(i,2) ) / 0.6 ).^2+ ...
        ( ( A_Data(:,3)-A_ref(i,3) ) / 0.6 ).^2+ ...
        ( ( A_Data(:,4)-A_ref(i,4) ) / 0.6 ).^2+ ...
        ( ( A_Data(:,5)-A_ref(i,5) ) / 0.6 ).^2 ...
        ) /2 ...
        ) ...
        );
    
    theta(i,:)=regress(W.*b_Data,repmat(W,1,5).*A_Data);
    b_ref(i,1)=A_ref(i,:)*theta(i,:)';
    
end

delete(poolobj);  

Answer 1

[Matt J]

There's never any gaurantee parfor will be faster, but I would modify the code as follows,

B_Data=A_Data(:,2:5)/(0.6*2);
B_ref=A_ref(:,2:5)/(0.6*2);

parfor i=1:k   

    W = exp(sum(  -( B_Data-B_ref(i,:) ) ).^2 ,2) ;
    
    theta(i,:)=regress(W.*b_Data,W.*A_Data);
end

b_ref=A_ref*theta.';