MATLAB: How to distinguish sliced variables with reduction variables when I use “parfor” in parallel computing

parfor

Hello,everyone! Nowadays I am trying to do some parallel computing on Matlab.I get a question and the codes below may explain it.
a=1:1000;
b=zeros(10,1);
for i=1:1000
   j=mod(i,10);
  b(j+1)=b(j+1)+a(i);
end
In this example,I want to calculate a sum to a in every ten numbers,and save the results in an array b. And as it can be seen,it works well in a serial program. However,when I try to do it in parallel(below),an error appear:
a=1:1000;
b=zeros(10,1);
matlabpool local;
parfor i=1:1000
  j=mod(i,10);
  b(j+1)=b(j+1)+a(i);
end
matlabpool close;
"Error: The variable b in a parfor cannot be classified."
I think variable b is a reduction variable,because in line 5 of the code,it has the form "X=X+expr" (expr is any expression). But the examples of reduction variables in the "product help" of Matlab are all simple variable,while b is an array.So I turn to "sliced variable" for help.It is a pity that the index of b is not i(the loop variable).So maybe Matlab cannot make sure what kind of variable the "_b_" is. So,my question is:How to achieve my goal(sum a in a sliced way) in parallel? In other words, how to modify the code in parallel so as to get the same results as that in serial way? Your answer or idea is important for me.Thank you in advance!

Best Answer

Hi, I think you can try to use spmd programming to reach your goal. It can deal with your problem regardless of the form of j(i). See below for example:
matlabpool local; % startup of the parallel environment
tic;
spmd
a=1:100000;
D = codistributed(a); % distribute a in parallel workers  
b_in_worker=zeros(10,1); % temporal b array in each parallel worker
% do the operation on the partial a array in each parallel worker
% and restore the partial results in b_in_worker.
for i = drange(1:length(D))
    j = mod(i,10);
    b_in_worker(j+1) = b_in_worker(j+1)+D(i);
end
end
% sum up all partial results
b = zeros(10,1);
for i = 1:length(b_in_worker)
    b = b+ b_in_worker{i};
end
disp Parallel:
toc;
matlabpool close;
I have not tested the parallel efficiency. However, since matlab is based on java, the communication between prallel workers can be very consuming. I advise that you need to make some effort on reducing the data exchange between workers.
The following are some selected official references of spmd which I think could be helpful:
Working with Codistributed Arrays
Execute Simultaneously on Multiple Data Sets
for
Related Question