Is the GNU GSL Mersenne twister PRNG faster than the MKL VSL version?

[oscar_m_]

Hello everybody

I would like to perform some montecarlo simulations using the well-known Mersenne Twister PRNG. Before to start to write my code I decided to test the performance of this PRNG using the GSL and MKL libraries, so I wrote a small code for such comparison. The results I got suggested that the GSL version is faster.  May it be possible or am I doing something wrong?. I am compiling using

[]$icpc benchmark.c -O3 -c

[]$icpc benchmark.o -o benchmark -L/opt/intel/parallel_studio_xe_2017/compilers_and_libraries_2017/linux/mkl/lib/intel64 -mkl:parallel -lgsl -lgslcblas

Do you have a feeling about what is happening? The code is in attachment. It takes as argument the number of random variables to be generated.

Thanks

[Gennady_F_Intel]

This is an unexpected perf results. could you try the latest mkl 2018? what the performance numbers do you obtain?

[Gennady_F_Intel · Accepted solution]

Oscar,  i didn't noticed at the attached the example of you benchmark. The picture is clear now -  we don't recommend to use MKL RNG on the vectore size 1 like you did. 

[oscar_m_]

If I create a buffer with size equal to the number of random variables e.g. 10^8 will the mersenne twister from MKL be faster than the GNU version? Can you try and confirm it please.

[Gennady_F_Intel]

yes, it will. The performance difference will depend on system where you will execute this code.

[oscar_m_]

Thanks Gennady. Luckily, that was the trick, BTW my application will use MPI so I will have to use the MT2203 generator to obtain different sequences for each MPI task. Do you know about a toy code showing how to use the MT2203? May be a link.