Multithreading issues with schur complement computation

[Niels_G]

Hello,

Here are information that you often request:
- OS details and hardware :
Description: CentOS Linux release 7.8.2003 (Core)
Release: 7.8.2003
Intel(R) Xeon(R) Gold 6252 CPU @ 2.10GHz (max RAM 252G)
- Version Intel MKL : 2020.1.217

I am trying to solve a classical linear problem Ax=b, A being symmetric positiv definite. The matrix A is very sparse and has 1100 rows/columns. We have to compute the schur complement during the process (even if in the attached routine it seems useless) and decompose the process with phases 12, 331, 332 and 333. The following parameters are used :
iparm(1 ) = 1
iparm(2 ) = 1
iparm(5 ) = 0
iparm(6 ) = 0
iparm(24) = 1
iparm(35) = 0
iparm(36) = 2

Everything works fine if we set the number of threads to 1 with mkl_set_num_threads. But with 2 threads I get a segmentation fault. With 3 threads and more the results are wrong and change with the number of threads. I tried to switch iparm(24) to 1 (which does not seem logical to me) and it worked it this case. But for bigger cases it looks like we have a deadlock during phase 12 : this phase takes a few seconds in sequential and it runs for hours (I stopped before end of calculation) with only 2 threads.
You will find attached to this email a routine that reads the data from files (in data.zip) and proceed to all phases. it is compiled with command "ifort -mkl -qopenmp pardiso_schur_omp.f" and the path to data in the fortran code has to be adapted.

Niels

[AryanK_Intel]

Hi Niels_G,

Thank you for posting in Intel Communities.

We would like to request you to use the latest version of MKL as per the below provided system requirements link. And please let us know if the issue still persists.

https://www.intel.com/content/www/us/en/developer/articles/system-requirements/oneapi-math-kernel-library-system-requirements.html

Thanks & Regards,
Sri Raj Aryan.K

[Niels_G]

Hi AryanK,

Thank you for your answer. I have the exact same behavior with intel version 2023.1.0.

Niels

[AryanK_Intel]

Hi Neils_G,

>>Everything works fine if we set the number of threads to 1 with mkl_set_num_threads. But with 2 threads I get a segmentation fault. With 3 threads and more the results are wrong and change with the number of threads.

We are able to reproduce the above issue and we are working on it.

>>But for bigger cases it looks like we have a deadlock during phase 12 : this phase takes a few seconds in sequential and it runs for hours (I stopped before end of calculation) with only 2 threads.

Could you please elaborate on the issue and let us know for which bigger cases you are getting deadlock?

Thanks and Regards,

Sri Raj Aryan.K

[AryanK_Intel]

Hi Neils_G,

Gentle reminder:

We haven't heard back from you. Could you please provide an update on the above-required information?

Thanks and Regards,

Sri Raj Aryan.K

[AryanK_Intel]

Hi Neils_G,

Gentle reminder:

We haven't heard back from you. Could you please provide an update on the above-required information?

Thanks and Regards,

Sri Raj Aryan.K