simulink in parfor and for: different results

Hello,

I have a Simulink-Simulation. I need many evaluations with different parameters, therefore, i tried parallel computing.

The strange thing is: when i run the simulation on the matlab "workers" in a parfor loop, i get a different result compared to when i run it in a normal for-loop.

What could be the reason?

Thank you in advance, Arco Bast

[Edric Ellis]

Do you have a specific example you can post so that we can try to reproduce the problem?

[Arco]

Since my model is realy complex, I first want to give you more information - maybe this already allows to find the mistake.

My model describes a relatively large nonlinear ODE, that is very stiff, therefore, I'm using ode15s. It's main functionality is inside a Matlab-Function-Block.

So let's run the following code:

load('myParameters');
simOutput=cell(3);  

    parfor i=1:1
    simOutput{i}=sim('mySimulation', ...
        'ReturnWorkspaceOutputs', 'on', ...
        'SimulationMode','normal', ...
        'AbsTol','1e-6',...
        'RelTol','1e-6');
    end

    for i=2:2
    simOutput{i}=sim('mySimulation', ...
        'ReturnWorkspaceOutputs', 'on', ...
        'SimulationMode','normal', ...
        'AbsTol','1e-6',...
        'RelTol','1e-6');
    end

    matlabpool;
    parfor i=3:3
    simOutput{i}=sim('mySimulation', ...
        'ReturnWorkspaceOutputs', 'on', ...
        'SimulationMode','normal', ...
        'AbsTol','1e-6',...
        'RelTol','1e-6');
    end

simOutput{1} and simOutput{2} are identical, but simOutput{3} is different. The difference is very slight, but nevertheless irritating. Might it be possible, that matlabpool has a slightly different numeric precission compared to Matlab itself?

[Edric Ellis]

There are very occasional slight numerical differences caused by the fact that workers run in single-computational-thread mode. Try launching the MATLAB desktop client with the "-singleCompThread" command-line argument and comparing results.

Answer 1

[Abhishek M]

Hi Arco, This problem arises when you aren't doing anything to save the results of the parfor loop iterations. You need introduce a new output variable and then assign your parfor output to this new variable in order to save them.