topic Moving from GotoBLAS2 to Intel MKL in Intel® oneAPI Math Kernel Library

Moving from GotoBLAS2 to Intel MKL

marco_v — Mon, 03 May 2010 21:07:02 GMT

Dear Forum users,

Do you know if Intel MKL supports the new Nehalem-EXarchitecture?

I've been using the GotoBLAS2 optimized BLAS library to compile a C program where I call the function cblas_sgemm to do a simple matrix multiplication.
As I've been using a Nehalem-EP server (Dual Xeon 5660), and as the GotoBLAS2 library has been hand tunned for the Nehalem-EP architecture, I've preferred to use it.

But now I'm moving tothe new Nehalem-EX 7550 (Quad Octo processors server / 32 cpus). Since the GotoBLAS2 does not support the new Nehalem-EX architecture, I'll have to start using another library.

Would the current available version of Intel MKL provide good performance on the new Nehalem-EX ?

Thank you,
Marco

Moving from GotoBLAS2 to Intel MKL

Shane_S_Intel — Mon, 03 May 2010 21:33:22 GMT

Hi Marco, the first NHM EP optimizations were added to MKL over 2 years ago. The latest release -MKL 10.2.5 -has performance optimizations forthe recently releasedNHM EX and Westmere systems.

Moving from GotoBLAS2 to Intel MKL

marco_v — Mon, 03 May 2010 21:50:41 GMT

Thanks for replying Shane!

In GotoBLAS2 library there is a parameter wich lets you choose how many threads to use during the execution of the program compiled with such a library.
Is there a similar parameter in Intel MKL?

I mean, is it possible to compile the program so that it can execute a CBLAS function (like cblas_sgemm) in a parallel/multi-threaded way exploring all 32 cpus of a Quad Octo 7550 server?

In GotoBLAS2 the command is like:

$.\bash GOTO_NUM_THREADS=2 ./executable

Thanks again for your attention.
Best rgs,
Marco

Moving from GotoBLAS2 to Intel MKL

Shane_S_Intel — Mon, 03 May 2010 21:58:56 GMT

Yes, there are a few ways to do this, among them to call themkl_set_num_threads() service function or to set the MKL_NUM_THREADS environment variable. See this Knowledge Base (KB)article for more information - http://software.intel.com/en-us/articles/intel-math-kernel-library-intel-mkl-intel-mkl-100-threading/