topic Re:Weird time overhead during first call to DGEQRF and DSYRK in Intel® oneAPI Math Kernel Library

Weird time overhead during first call to DGEQRF and DSYRK

szachista — Tue, 18 Mar 2025 10:26:52 GMT

Hello everyone,

I encountered odd behaviour of those procedures when called for the first time - subsequent calls take shorter execution time. I assume it is caused by some allocations made by MKL library to speed-up. What can be done to avoid such situations?

I attach source code of my test. I've set environmental variables OMP_NUM_THREADS and MKL_NUM_THREADS to 1.

Thank you for any help,
Jacek

Re:Weird time overhead during first call to DGEQRF and DSYRK

Aleksandra_K — Mon, 31 Mar 2025 07:39:53 GMT

HI Jacek,

When you run this function for the first time, many things happen behind the scenes, so it's normal for it to take longer. For smaller sizes, you can try using direct call or sequential mode (Using MKL_DIRECT_CALL in C Applications).

Also, could you provide us with more context about what your application does and what kind of usage of oneMKL you expect? We can try to find a better solution for your case.

Regards,

Aleksandra

Re: Re:Weird time overhead during first call to DGEQRF and DSYRK

szachista — Mon, 31 Mar 2025 07:59:02 GMT

Hello Aleksandra,

Thank you for your answer. Actually I was measuring time (in microseconds) of the problem involving very small matrix - later I checked for a much larger matrix and the difference was negligible, so the topic may be closed. I am sorry for confusion.

With regards,

Jacek