topic Re: Performance drop in dgetrf() in Intel® oneAPI Math Kernel Library

Performance drop in dgetrf()

techwinder_ — Tue, 07 May 2024 00:17:19 GMT

Hello,
I'm distributing an app which links to MKL - amazing library btw, thanks.
A user has recently reported a significant performance drop in dgetrf_() when switching to a new computer with an Intel® Core™ i9-13950HX Processor.
Matrix size is ~27000. Typically the factorization should only take 2 minutes or so in multithreaded mode, however it can take up to 40 minutes on the new platform. When disabling in-app multitheading the LU factorization takes roughly 6 minutes which seems normal.

Adjusting MKL setttings with

MKL_Set_Num_Threads_Local(); MKL_Set_Dynamic();

doesn't make any significant difference.

After disabling efficient cores in the BIOS so that only the performance cores are enabled, the factorization time is back to a minute and a half.
Do you have any suggestions to address this issue without having to tweak the BIOS settings?
Thanks in advance,
André

Re: Performance drop in dgetrf()

noffermans — Wed, 08 May 2024 07:32:16 GMT

Hi André,

Thanks for reaching out. Are you aware of the below doc page?

https://www.intel.com/content/www/us/en/docs/onemkl/developer-guide-linux/2024-1/managing-performance-with-heterogeneous-cores.html

It gives some useful tips on how to manage performance with heterogenous CPUs, without having to play with the BIOS. I hope this helps.

Best,
Nicolas

Re:Performance drop in dgetrf()

Fengrui — Fri, 10 May 2024 23:13:15 GMT

Hi André,

Did you get a chance to try the environment variables shared by Nicolas?

Thanks,

Fengrui

Re: Performance drop in dgetrf()

techwinder_ — Sat, 11 May 2024 02:20:38 GMT

Yes, it fixed the problem for the user.

Thank you all for the quick support.

André