topic Random parallelism with MKL Pardiso OpenMP in Intel® oneAPI Math Kernel Library

Random parallelism with MKL Pardiso OpenMP

cool__thomas — Mon, 29 Jan 2018 23:49:14 GMT

When linking the latest version of MKL with my c++ program with OpenMP and running the program multiple times, it sometimes parallelizes and sometimes it doesn't. I was wondering if someone has noticed something like this before?

Couple of notes:

the program uses both TBB (for parallelism in the rest of the code) and OpenMP for MKL. There are no nested MKL calls
if TBB is used for both, it works but pardiso is only partially parallelized in TBB and is thus significantly slower
different parallelism is observed without recompiling
1. sometimes one thread and sometimes the proper amount
the TBB library used is the one provided with MKL
MKL version is 2018 U1
the OS is windows 10
I have tried using mkl_set_num_threads, omp_set_num_threads, omp_set_dynamic(false) and mkl_set_dynamic(false). I have also tried setting the blocktime to no avail

Hopefully someone has an answer to this problem

Thanks,

Thomas Cool

Could you please clarify the

Gennady_F_Intel — Thu, 01 Feb 2018 04:36:21 GMT

Could you please clarify the following:

How do you link this application? i mean do you link with mkl_tbb_thread or mkl_intel_thread libs?

What is the typical problem size you solve?

what does this mean: >> if TBB is used for both, it works but pardiso is only partially parallelized in TBB and is thus significantly slower

Sorry for not being clear in

cool__thomas — Thu, 01 Feb 2018 21:18:45 GMT

Sorry for not being clear in the original message.

If the program is linked with mkl_tbb_thread, mkl runs in parallel but pardiso does not seem to be fully converted to support it yet.

If the program is linked with mkl_intel_thread, (in addition to libiomp5.lib) mkl routines seem to sometimes run in parallel and sometimes not.

This behavior change can be observed WITHOUT recompiling which is the confusing part

It should be noted that in both cases the program is also linked to tbb.lib since the rest of the program needs it

Problems are sparse but can vary from 10K*10K to 1.5M*1.5M

Thanks,