Is 'parfor' running correctly when mex-ed ?

I have mex-ed a function containing a parfor loop and I get 0 Errors, Warnings, Notices or Messages.

When I run the original parallelized Matlab function on 10 cores, I can see in Task Manager (or ProcessExplorer) 10 Matlab threads each taking about 6% for a total of about 62% of CPU processing power.

When I run the mex function, I can only see one Matlab thread taking 50% of the CPU power.

Is the mex function actually running on 10 parallel threads (cores) or is it executing in single threaded fashion ? Can I verify that the compiler/OpenMP actually functioned correctly when mexing the code ?

I am using: 2x Intel Xeon X5560 @2.80GHz, Windows Server 2008 R2 Enterprise, Matlab R2015b, Microsoft Visual C++ compiler 2012.

Please see the attached file with the compilation report, if needed.

Thank you,

Victor

Answer 1

[Fei Deng]

How did you run the mex-ed function with par-for loop? You can specify on how many threads (at most) it runs in parallel by executing function_with_par_for(number of threads) in command window. You can take a look of the example "Generate MEX for parfor" here https://www.mathworks.com/help/coder/ref/parfor.html

[Victor Aldea]

I have hard coded the number of threads in the function. It reads like this:

parfor (pixelIndex=1:(L1*L2), 10)

then I mex-ed it and ran it from a batch file. Thanks for the example from Matlab's help, but I had already followed it.

   It seems very strange that the CPU utilization was lower when running the mex-ed function than when running the function directly in Matlab (there were no other users on that PC at the time, so all cores were free).

Also, only 1 Matlab thread was present taking 50% of the CPU power, instead of having 10 Matlab threads each taking about 6%. Is this normal behaviour ?