I got NaN using parfor while the same code works with for loop

Hi everyone

I found I got NaN using parfor while the same code works with for-loop.

My input is very large. So I tested it with a small part of my input (only three times loop) for both for-loop and parfor loop. I got the same results. Then I used "dbstop if naninf" to debug.The code stopped at the beginning of parfor loop for parfor. While the code can give me the same results for the for-loop as if the there is no "dbstop if naninf". Finally, I tested with my entire input for both loop without "dbstop if naninf". The parfor gave me NaN to the arrays which should have values.While the for loop works fine. The part code of parfor is listed below.

parfor i=1:num_atom
   tdelta_f(:,i)=xcorr(atom_for_v(:,1,i),atom_for_v(:,1,i),'biased')+xcorr(atom_for_v(:,2,i),atom_for_v(:,2,i),'biased')
                  +xcorr(atom_for_v(:,3,i),atom_for_v(:,3,i),'biased');
   tdelta_fv(:,i)=xcorr(atom_for_v(:,4,1),atom_for_v(:,1,i),'biased')+xcorr(atom_for_v(:,5,1),atom_for_v(:,2,i),'biased')
                  +xcorr(atom_for_v(:,6,1),atom_for_v(:,3,i),'biased');
end

(1)It is bug of my code or bug of parfor?

(2)If my code is fine. How can I use parfor for my code in this case?

Fan Li

[Edric Ellis]

Please could you post a minimal, self-contained example that shows the bad behaviour.

Answer 1

[Fan Li]

Hi Edric

This a part of my code. It reads the data and does the correlation calculations. You can reproduce the problem with code and attached files. Thanks for your time.

%tot_step=1000000;
  tot_step=40;%use the integer which is the multiple of ten, the maximum is 40. 
  dump_fre=10;
  num_atom=4323;
  file='outputfile.txt';
  save1='3.8atom_for_v_spce_rerun.mat';

    num_dump=(tot_step/dump_fre)+1;
    num_col=7;

    atom_for_v=zeros(num_dump,num_col-1,num_atom,'single');
    temcell=zeros(1,num_col+1,'single');
    tdelta_f=zeros(2*num_dump-1,num_atom,'single');
    tdelta_fv=zeros(2*num_dump-1,num_atom,'single');

    fileID = fopen(file);
    if (fileID==-1)
         disp('Can not open the file');

     else 
         disp('Open the file sucessfully');
    end

    dbstop if naninf;

    disp('Reading the data ...');
    for i=1:num_dump
        %skip the 9 lines header
        for j=1:9
            tline=fgetl(fileID);
        end
        %read the data and retrieve them to the multimensional array according
        %to the atom id
        for k=1:num_atom
            tline=fgetl(fileID);
            temcell=strsplit(tline);
            atom_for_v(i,:,k)=cellfun(@str2double,temcell(:,2:7));
        end

    end

    fclose(fileID);

    save(save1,'atom_for_v','-v7.3');

    disp ('Calculating the correlation function ...');

    %it can pass if use for instead of parfor
    parfor i=1:num_atom
       tdelta_f(:,i)=xcorr(atom_for_v(:,1,i),atom_for_v(:,1,i),'biased')+xcorr(atom_for_v(:,2,i),atom_for_v(:,2,i),'biased')+xcorr(atom_for_v(:,3,i),atom_for_v(:,3,i),'biased');
       tdelta_fv(:,i)=xcorr(atom_for_v(:,4,1),atom_for_v(:,1,i),'biased')+xcorr(atom_for_v(:,5,1),atom_for_v(:,2,i),'biased')+xcorr(atom_for_v(:,6,1),atom_for_v(:,3,i),'biased');
    end

    aucf_f=sum(tdelta_f,2);
    crof_fv=sum(tdelta_fv,2);

Answer 2

[Edric Ellis]

dbstop if naninf always causes execution to stop during setup of a parfor loop. This turns out to be expected behaviour (parallel_function.m has an explicit assignment to Inf), and I believe there is nothing to worry about - either in your code, or the implementation of parfor.