Global variables in parfor loop

Question

Abhinav 2018-6-29

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/408110-global-variables-in-parfor-loop

评论： Abhinav 2018-6-29

I am using global variables in parfor loop and I get following m-lint

 The code uses a global variable in a parfor loop. Because parfor loops run on several different machines simultaneously, the global workspace might not be the same on each machine. Therefore, using a global variable in a parfor loop could have unpredictable or unexpected results and Code Analyzer flags it as an error.

This is my code:

 r=cell(size(sample,1),1);
parfor samp=1:size(sample,1)
    theta_opt=theta_opt_array(samp,:);       
    Cp_opt=diag(Cp_opt_array(samp,:));    
    beta_opt=beta_opt_array(samp);          
    sigmaobs2_opt=sigmaobs2_opt_array(samp);  
    mu=strm(:,samp);                 
    n=10000;                          
    Ct_opt=CovarianceMatrixEstimation(Cp_opt,theta_opt,sigmaobs2_opt,...
        GLOBAL_DATA,GEOMORPH);  % these two are global variables
    r{samp}=mulgennormrnd(n,mu,Ct_opt,beta_opt);
end

Any suggestions on how to fix it? I don't understand the suggested fix by MATLAB. The suggestion by MATLAB is:

 Make local copies of the global variables that you want to use within the parfor loop before beginning the parfor loop

Also,I am not changing my global variables in parfor loop.

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Adam Danz 2018-6-29

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/408110-global-variables-in-parfor-loop#answer_326896

编辑：Adam Danz 2018-6-29

What was the suggestion by Matlab? Have you read through the plethora of answers to this topic?

There are lots of reasons not to use global variables at all and this is one of them. As the error message indicates, parfor() is executed in parallel potentially between >1 machine. Global variable are not accessible between machines. So you'll have to initialize them independently within your code.

6 个评论
显示 4更早的评论隐藏 4更早的评论

Steven Lord 2018-6-29

So your mulgennormrnd function needs information from the CovarianceMatrixEstimation call that took place earlier in the same loop iteration (not cross-iteration)?

The simplest approach, and the one that makes it clear that you're not trying to share data across iterations, would be to have your CovarianceMatrixEstimation function return that information as additional outputs and have your mulgennormrnd function accept that information as additional inputs.

Abhinav 2018-6-29