MATLAB: Convolution – Parfor loops not giving speedup on 12 core than 4 core machine

Question

Hello,

I am using "Matcovnet" to work on the Imagenet dataset for training and inference using fixed point operands. I am working on two machines simultaneously, both having different number of cores

1) 4 cores

2) 12 cores

I am running same dataset and same fixed point training code on both the machines simultaneously. I am getting 0.2 Hz frequency (0.2 images/sec) on both the machines, which ideally should not happen as the number of cores are different for both machines. I am having various parfor loops in the code.

The parfor loops are in the codegen files which are converted in the MEX functions by Matlabcoder toolbox.

My main question is:

1) Even if I use for loops instead of parfor my both machines uses all the cores by default with around 90% CPU-Utilization (just having parallel computing toolbox has this effect?) and also, my speed of computation/training is still the same that is 0.2 Hz, Can you please let me know about it ?

2) My 2nd question is, if I use parfor loop and run the code on this two machines why am I not getting any speedup on the machine with 12 cores compared to 4 cores ? Also, here on both machines I am getting the same training speed, that is 0.2 Hz.

Thank you in advance for your help !

Bhushan

Answer 1

Hello,

I am answering my own question.

After doing some research I found out that my compiler(Microsoft Windows SDK 7.1) does not support openMP according to the link.

https://www.mathworks.com/matlabcentral/answers/237411-can-i-make-use-of-openmp-in-my-matlab-mex-files

For using parfor loop while generating MEX file the compiler must support openMP. Below link gives this explanation.

http://www.mathworks.com/help/coder/ug/acceleration-of-matlab-algorithms-using-parallel-for-loops-parfor.html#btfitu2-9

So my question is, can you provide me the compiler which will support

1) Matcovnet 2) Parallel Computing toolbox 3) OpenMP 4) Matlab coder

I am not able to get the clear explanation about the compiler supporting openMP from the link.

 https://www.mathworks.com/support/compilers.html

Thank you for your kind help.