Intel® oneAPI Math Kernel Library
Ask questions and share information with other developers who use Intel® Math Kernel Library.
7232 Discussions

Intel MKL forces processor affinity mask to use one processore, when 8 are available

intelmklpriv-id_com
545 Views
HI,

I run some Intel MKL code in an execuble. I start this executable 8 times on a duo processor, quad core AMD machine. It turns out that all 8 processes are assigned to the same core, due to the processor affinity mask being set to 0x01. This changing of the affinity mask occures upon (and WITHIN) my first call to an Intel MKL function.

Does anyone have a clue why this happens? I definitely would like Intel MKL to not touch these settings. What can I do to prevent this?

I am running Version: 9.1.027


Joop
0 Kudos
1 Reply
Ilya_B_Intel
Employee
545 Views
Hi Joop,

This is a known issue. First call of VML function can change affinity. This was fixed in 10.0 MKL. The workaround forit is to make some call of VML function on vector of at least 100 length, remember affinity befor it and set it back after it. Further calls of VML functions are not changing affinity.

Ilya


HI,

I run some Intel MKL code in an execuble. I start this executable 8 times on a duo processor, quad core AMD machine. It turns out that all 8 processes are assigned to the same core, due to the processor affinity mask being set to 0x01. This changing of the affinity mask occures upon (and WITHIN) my first call to an Intel MKL function.

Does anyone have a clue why this happens? I definitely would like Intel MKL to not touch these settings. What can I do to prevent this?

I am running Version: 9.1.027


Joop

0 Kudos
Reply