topic Pardiso crashes in 2017 MKL only when OMP_NUM_THREADS=2 in Intel® oneAPI Math Kernel Library

Pardiso crashes in 2017 MKL only when OMP_NUM_THREADS=2

Ferris_H_ — Mon, 26 Sep 2016 04:29:59 GMT

I have a small matrix that consistently causes pardiso to crash only when I use OMP_NUM_THREADS=2 or 3. My machine has 12 cpus and any other setting for OMP_NUM_THREADS works. The other odd thing is this same matrix solved without any issue in earlier version of Intel compiler ( 2016 ).

I modified the example cl_solver_sym_sp_0_based_c to read in my matrix and reproduce my error. I can share my test case if you like. Or is this a known issue with 2017 Intel compiler?

I have found a workaround is to set iparm[0]=1 . Then I can use any value for OMP_NUM_THREADS. With iparm[0]=0, the crash consistently appears. As mentioned this error is new to 2017 intel compiler. I did not ever experience it before with the same matrix.

yes, the problem with METIS

Gennady_F_Intel — Mon, 26 Sep 2016 05:59:35 GMT

yes, the problem with METIS has been addressed into MKL 2017. Could you please check it on your side and give us the results!

Thanks for the quick reply,

Ferris_H_ — Mon, 26 Sep 2016 15:27:41 GMT

Thanks for the quick reply, however, the issue I am experiencing is with 2017 MKL. The same matrix solves perfectly in 2016 and older MKL. Seems some bug with the 2017 release. I can provide my test case if you like.