topic Have you tried checking the # in Intel® oneAPI Math Kernel Library

Switching between threaded and non-threaded libraries

mullervki — Tue, 28 Jul 2015 19:04:59 GMT

Hello,

I have an application that uses Intel's MKL libraries. In some sections of this application calls to MKL are made in parallel areas. In these areas, I want to use the sequential version of the MKL libraries; at times I am not in a parallel area in my application, so I would like to use the threadedversion of the MKL libraries here.

I understand I can only link my application with either the threaded or the sequential libraries. Given that I need to use the threaded version in certain places, I assume I must link with the threaded version. I set my environment variables OMP_NUM_THREADS and MKL_NUM_THREADS to 1. My understanding is that I should be able to call mkl_set_num_threads(n), where "n" is either 1 - in areas where I want to use MKL in sequential mode - or ">1", in threaded MKL calls.

But if I bring the Windows task manager I notice that calling mkl_set_num_threads(1) when I'm using the threaded version of the MKL library still starts many threads. If I use the sequential version, on the other hand, this is not a problem. It appears that mkl_set_num_threads does not have the effect I expected.

If I run my application entirely in sequential mode, then even if I call mkl_set_num_threads(1) I still see multiple threads running when MKL is being called. Since I know the threading is not coming from my application, it must be from MKL.

Is there a way to force MKL to run in sequential mode - even if I'm linking with the threaded MKL library?

Thanks.

Have you tried checking the #

VipinKumar_E_Intel — Wed, 29 Jul 2015 09:14:34 GMT

Have you tried checking the # of actual threads in your application using mkl_get_max_threads()?

Can you disable dynamic threading using MKL_DYNAMIC to FALSE and setting the MKL_NUM_THREADS=1 even when you use threaded MKL functions?

--vipin

Vipin,

mullervki — Wed, 29 Jul 2015 14:57:07 GMT

Vipin,

Thanks for the reply. I'll give your suggestion a try. I'm not familiar with MKL_DYNAMIC (I'm rather new to this issue). I assume this is an environment variable to be set to FALSE also.

I did some reading on MKL_DYNAMIC (https://software.intel.com/en-us/node/528547) and was surprised by the comment "Note also that if Intel MKL is called in a parallel region, it will use only one thread by default."

Does this mean I have no need to call mkl_set_num_threads at all? If you are familiar with the underlying reasons for the comment above, could you elaborate on it? Thanks.

Hi,

VipinKumar_E_Intel — Thu, 30 Jul 2015 05:42:39 GMT

Hi,

Yes, that's the environment variable.

That note is about when you call an MKL routine from a parallel region, say an OpenMP parallel region to have nested parallelism like, here

#pragma omp parallel for

for (int i; i < n; i++) {
mkl_function();

}

by default, mkl will use single thread. If you still want to have nested parallelism, then you can try with MKL_DYNAMC to FALSE, and manually set the # of threads than MKL to decide it. Also, please note, you are using the same omp compiler as MKL is using as mentioned there.

Vipin

Vipin,

mullervki — Thu, 30 Jul 2015 16:06:44 GMT

Vipin,

You've been most helpful. Thanks.

I'm multi-threading the application using Windows threads, not OpenMP. I think I may have found the problem: running the application from the command line seems to work as expected if the environment variables are set at the command prompt. So I started Visual Studio (devenv) from the command line to make sure the environment variables were properly set (I chose not to change the environment variable globally at this point).