Why does the same code run 50 times faster in 2008b than in 2012a?

Hi,

I'm runing a simple Monte Carlo loop, using the exact same code, in Matlab R2008b and it takes 29 seconds to do 30,000 iterations, while the same thing in Matlab R2012 a takes 26 minutes.

There is nothing fancy in the loop, I just generate a multivariate randome number and then do the monte carlo simulations (see code below). My code calls a function (psimat) that just reorganizes the data. I just can't understand the huge difference in performance, and it's not the PC, I'm running both on a server.

Any ideas?

Thank you,

Sebastian

for num = 1:monte;  % 'monte' times of iterations

    smpi_rnd = mvnrnd(smpi,kron(omega,iQ)/nobs);    %Multivar. N random number

    pi_rnd = reshape(smpi_rnd,K*p+dis*(q+1),K);    
    psi_rnd = psimat(p,q,K,dis,st,pi_rnd);

    % Store mean responses 
    for i = 1:dis
        for j = 1:st+1
            mr_rnd(j+(i-1)*(st+1),num) = psi_rnd(K*j,i);
        end;
    end;    

    % Store cumulative effects
    for i = 1:dis
        ce(i,num) = sum(mr_rnd(1+(i-1)*(st+1):4+(i-1)*(st+1),num));
    end;

 end;  % End of Monte Carlo loop

[Oleg Komarov]

Profile the code and check if there is a bottleneck in the execution on the 2012a. For example, what is psimat()? And, are you sure you're using the same psimat()?

[Sean de Wolski]

32bit v. 64bit MATLAB?

[Matt Fig]

Is one a function and the other a script?

[Walter Roberson]

We've seen some reports of this having to do with 64 bit MATLAB on i5 and i7 processors. I seem to recall the reports were more about R2011b, but I would have to check that.

[Jan]

@Walter: Are there explanations of this effect, or even workarounds? If you do not know any, does TMW do?

[per isakson]

Sebastian,

Would it be possible for you to provide a piece of code, with which we could try to reproduce your findings?

Answer 1

[Ryan G]

In general you should expect some minor slowdown between versions for the same code on the same machine with a new version of MATLAB. This is because of added features, maybe some error checking, who knows. However, 26 minutes is a huge slowdown that would be unexpected.. I would also suggest you utilize the MATLAB Profiler to determine where the slowdown occurs.

To speed this up from there you can use PARFOR. See examples below:

Monte Carlo Simulation

[Sebastian]

Thanks for the answer.