Working around segmentation violations in parfor loops

3 次查看（过去 30 天）

Berk Ustun 2012-4-13

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/35342-working-around-segmentation-violations-in-parfor-loops

I am currently working on a project where I have to run multiple repetitions of a time-consuming MATLAB function in parallel. For the purposes of this question, let's refer to the function as myfunc.

myfunc uses a MEX file and ends up with a random segmentation violation once every 3 hours. I cannot diagnose the segmentation fault since it originates from a propriety API that I did not code myself. However, I do know that it occurs within the MEX file, and I also know that it is not deterministically related to any settings I can change.

I would like to work around the segmentation violation, and I would ideally also like to keep on using the parfor function in MATLAB. My idea right now is to use a try catch loop within the parfor loop as follows:

    %create an output cell to store nreps of output from 'myfunc'
    output = cell(1,nreps) 
    %create a vector to track # of runs that finish successfully
    successfulrun = zeros(1,nreps);
    % run myfunc in parallel
    parfor i = 1:nreps
       try
        output{i}
        successfulrun(i) = true
       end
    end
    %rerun experiments that did not end up successfully
    while sum(successulruns) < nreps
      %count # of experiments to rerun
      %initialize variables to store new results
      reps_to_rerun = find(successfulruns == 0);
      nreps_to_rerun = sum(reps_to_rerun);
      newoutput = cell(1,nreps_to_rerun);
      newsuccessfulrun = zeros(1,nreps_to_rerun)
      %rerun experiments
      parfor i = 1:nreps_to_rerun
         try
          newoutput{i};  
          newsuccessfulrun = true;  
         end  
      end
     %transfer contents to larger loop
     for i = 1:nreps_to_rerun
        rerun_index =  reps_to_rerun(i);
        successfulrun(rerun_index) = newsuccessfulrun(i)
        if newsuccessfulrun(i)
            output{i} = newoutput{i};
        end 
      end
    end

My questions are:

1. Will it be OK to keep continuing to run more repetitions like this even though there was a segmentation violation within the MEX file? Or should I clear the memory / restart the matlabpool? I'm assuming this shouldn't be problem since the segmentation violation was in C.

2. Is there any way to "break" out of a parfor loop?

5 个评论
显示 3更早的评论隐藏 3更早的评论

Sean de Wolski 2012-4-23

Does it occur with a regular for-loop?

Berk Ustun 2012-4-23

No, the errors occur only in parallel. See here for an update: http://www.mathworks.com/matlabcentral/answers/36312-segmentation-faults-and-thread-safety-in-parfor-loops-part-ii

请先登录，再进行评论。

请先登录，再回答此问题。