Optimization/Parallelization in MKL

[Aaron_M_]

I have a 16 core system with hyperthreading (32 threads). The matrix calculations are on a 4096x4096 matrix. Using "dfeast_syev", it takes a few mins. And using "cblas_dgemm" it takes a few seconds. When monitoring my computer, it shows me that only 16 threads are being used in the calculation. I want to use all 32 to speed up the program. I am using open-mp when I can, but I thought before running my program, at the command line, I could just type :

"export OMP_NUM_THREADS=32"  and that would do it. But it doesn't work. It still only uses 16. Then I tried;

"export MKL_NUM_THREADS=32", and that does nothing either. When I enter "echo OMP(MKL)_NUM_THREADS" it returns 32 in both cases, but when I monitor my CPU it still only shows 16 threads being used.

Any suggestions?

Aaron.

[Bernard]

Probably there is contention between two threads running running on each core.As both of threads share the same SIMD vector stack there is possibility that one threads is stalled while waiting to access Vector stack.

[Gennady_F_Intel]

that's happens because of MKL_DYNAMMIC is true. in that case, MKL will choose the best number of threads. HypetThreading doesn't help for this sort computation. 

[Bernard]

Thanks for explanation.

[Aaron_M_]

Is there no way to speed up the computations? I have 8 dgemm functions in two nested for-loops. The for-loops run from 1-4000. So there are roughly 16million calculations. I was hoping that by parallelizing the BLAS functions further I could speed up the computations.

Thank you,

Aaron.