Parfor loop problem with SVM classification

Hi. I have a problem for paralleling a SVM Classification. In my code I repeat 3 times a 5-fold cross validation of SVM and average the accuracy of it. I want to paralell the main loop of 3 times repeat of cross validation but when I use parfor in this loop the program doesn't use all performance of 2 Cores of my CPU and only use 60% of the performance!.

Thank you so much

Code :

    parfor L=1:NumOFLoops;

    indices = crossvalind('Kfold',QQ,NumOfKfold);
    cp = classperf(QQ);

    for m=1:NumOfKfold;

    test = (indices == m); Train = ~test;
    testI=zeros(numel(test>0),1);

    TrainInputs=INPUT(Train,:);
    TrainTargets=QQ(Train,:);
    TestInputs=INPUT(test,:);

    %%SVM Structure

      svmstruct=svmtrain(TrainInputs,TrainTargets,...
     'boxconstraint',Penalty,...
     'kernel_function','rbf','method','QP',...
     'rbf_sigma',Sigma)

      TestOutputs=svmclassify(svmstruct,TestInputs,'showplot',false);
      classperf(cp,TestOutputs,test);

    end
    Error(:,L)=cp.ErrorRate;         
    end
    Results.ErrorRate=(sum(Error))/NumOFLoops;

Answer 1

[Edric Ellis]

If you're using R2013a or earlier, you need to explicitly open a matlabpool before running the PARFOR loop using "matlabpool('local', 2)". R2013b does this for you automatically.

Also, if you only have 3 iterations that can run in parallel, the best speedup you can hope for is 1.5 since one worker will have to perform two of the iterations.

[Eghbal]

Thank you so much for your reply. Yes. I'm suing R2013b. I increase numbers of parfor loop to 8 but the max performance in two cores in Task Manager is 73-75!. Two matlab in task manager are running.Two 32-35% processing. When I compared it with normal processing(1 core), the total performance is similar in two cases. for and parfor.

Thanks,

[Edric Ellis]

It's possible then that the performance of your loop is limited by memory bandwidth rather than by CPU utilisation.

Answer 2

[Eghbal]

I checked this problem with other classification & regression methods like MLP neural network (Newff) and other, but i have same problem.

Thanks,