Intel oneMKL cblas_dgemm performance degradation issue

[Rizwan1]

I have installed intel oneAPI BaseKit and HPCKit version 22.1.1

I am facing issue in performance degradation of MKL cblas_dgemm

When I execute mo_linpack bechmark I got  the performance as expected but when i execute cblas_dgemm I  got very less performance as expected. I don't know what is missing and why it gives very less performance

Could someone please help me to identify the environment variables and their values used by /opt/intel/oneapi/mkl/latest/benchmarks/mp_linpack routine this information may help me to identify the missing environment variables form MKL cblas_dgemm performance which may  degrade the performance.

Need urgent help

Thanks

[Rizwan1]

Could someone tell me what are the best possible setup of environmnet variables on intel xeon phi 7250 KNL to get the maximum performance of cblas_dgemm

Thanks

[VidyalathaB_Intel]

Hi,

Thanks for reaching out to us.

We are working on this issue, we will get back to you soon.

Regards,

Vidya.

[VidyalathaB_Intel]

Hi,

Could you please try exporting the mkl_enable_instructions environment variable as shown below to set the code-specific path to AVX-512 (we searched for your particular intel xeon phi 7250 KNL and noticed that it supports AVX-512 https://www.intel.com/content/www/us/en/products/sku/94035/intel-xeon-phi-processor-7250-16gb-1-40-g...) and see if it helps in improving the performance?

> icc -qmkl=parallel dgemm.c

> export MKL_ENABLE_INSTRUCTIONS=AVX512

> MKL_VERBOSE=1 ./a.out

Regards,

Vidya.

[Rizwan1]

I had already tried this but no difference. 

No I have uninstalled the current version and installed the previous one

[VidyalathaB_Intel]

Hi,

>> I have uninstalled the current version and installed the previous one

Could you please confirm whether you are facing the same performance issues with the previous version (please let us know the version you are using now)?

Regards,

Vidya.

[Rizwan1]

At previous version I am not facing this issue, at previous version performance is fine around 2TFlops but at current version 22.1.1 performance of cblas_dgemm is worst around 750GFlops

Thanks

[VidyalathaB_Intel]

Hi,

Thanks for the details.

>>...intel Xeon phi 7250 KNL to get the maximum performance of cblas_dgemm.....At previous version I am not facing this issue...at previous version performance is fine around 2TFlops but at current version 22.1.1 performance of cblas_dgemm is worst around 750GFlops,

We understand your concern and here are a few things that we would like to bring to your notice (as we could see that the same issue has been discussed in this previous thread https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/MKL-cblas-dgemm-huge-performance-gap...

The deprecation of KNL is already mentioned in oneMKL release notes,

" Dropped support for Intel® Xeon Phi™ Processor x200 “Knights Landing (KNL)” and Intel® Xeon Phi™ Processors “Knights Mill (KNM)” "

https://www.intel.com/content/www/us/en/developer/articles/release-notes/onemkl-release-notes.html

and it is also mentioned in the oneAPI HPCToolkit release notes what versions of parallel studio should be used by Xeon Phi Customers

" Support for Intel® Xeon Phi™ Processor x200 “Knights Landing (KNL)” and Intel® Xeon Phi™ Processors “Knights Mill (KNM)” is deprecated and will be removed in a future release. Intel® Xeon Phi™ customers should continue to use compilers, libraries, and tools from the Intel® Parallel Studio XE 2020 and older PSXE releases. "

For more details refer to https://www.intel.com/content/www/us/en/developer/articles/release-notes/intel-oneapi-hpc-toolkit-re...

Hope the provided information helps and let us know if we can close this thread from our end.

Regards,

Vidya.

[VidyalathaB_Intel]

Hi,

Reminder:

We haven't heard back from you. Could you please provide an update on your issue?

Regards,

Vidya.

[VidyalathaB_Intel]

Hi,

As we haven't heard back from you, we are closing this thread. Please submit a new question if you need any additional information from Intel as this thread will no longer be monitored.

Regards,

Vidya.