topic MKL on multiple CPU XEON in Intel® oneAPI Math Kernel Library

MKL on multiple CPU XEON

FrancoFDF — Thu, 15 Feb 2024 01:43:23 GMT

I have issue (total loss of precision) when using MKL calling DGESV() to invert a matrix in Fortran 90 on dual Dual core XEON Gold 6248R with 1TB RAM system and compiling for multiple thread execution. For instance if I link my program using the following libraries

..\intel64_win\mkl_intel_lp64.lib
..\intel\intel64_win\mkl_core.lib
..\intel\intel64_win\mkl_tbb_thread_dll.lib

the code would run using all the computer cores and threads (96) and will only produce valid results if the size on the matrix is <7999 elements. At n=8000 the solution would be just noise.

One interesting feature if I use the exact same libraries on a different system with single CPU (ex Intel 13th gen with 64GB Ram) the code would run just fine without any loss of precision for any matrix size.

If instead I link the code on the Dual core XEON Gold 6248R against following library

..\intel\intel64_win\mkl_rt.lib

the code would run using all the physical cores (48) at about 80% of their speed and will produce valid results regardless the size of the matrix but at lower speed since half of the threads aren't used.

Any possible solutions on how to run my code using all the available 96 threads ?

Source code and bat files use for compilation are attached

Re: MKL on multiple CPU XEON

Gennady_F_Intel — Fri, 16 Feb 2024 07:31:15 GMT

MKL uses the physical cores by default. You see the developer guide for more details. If you are interested in using hyperthreads also, then you could set MKL_DYNAMIC environment variables.

Example, the local system, – 40 physycal cores, 80 hyperthreading :

Default behavior: - 40 threads have been used, see the verbose outpus:

MKL_VERBOSE=2 ./a.out 2000

MKL_VERBOSE oneMKL 2024.0 Product build 20231011 for Intel(R) 64 architecture Intel(R) Advanced Vector Extensions 512 (Intel(R) AVX-512) enabled processors, Lnx 2.40GHz lp64 intel_thread

MKL_VERBOSE DGEMM(N,N,2000,2000,2000,0x7ffd8d611e18,0x14d86a5d4080,2000,0x14d86874f080,2000,0x7ffd8d611e28,0x14d8668ca080,2000) 65.43ms CNR:OFF Dyn:1 FastMM:1 TID:0 NThr:40

And here when we explicitly say MKL to use HT threads as well:

MKL_DYNAMIC=FALSE MKL_VERBOSE=2 ./a.out 2000

MKL_VERBOSE oneMKL 2024.0 Product build 20231011 for Intel(R) 64 architecture Intel(R) Advanced Vector Extensions 512 (Intel(R) AVX-512) enabled processors, Lnx 2.40GHz lp64 intel_thread

MKL_VERBOSE DGEMM(N,N,2000,2000,2000,0x7ffe77399688,0x1526551d4080,2000,0x15265334f080,2000,0x7ffe77399698,0x1526514ca080,2000) 50.29ms CNR:OFF Dyn:0 FastMM:1 TID:0 NThr:80

--Gennady

Re: MKL on multiple CPU XEON

Gennady_F_Intel — Fri, 16 Feb 2024 07:33:53 GMT

and one more - check the current (latest ) version of MKL ( 2024.0 ) if the DGESV() problem still exists when, as you reported, the problem size >8K

Re: MKL on multiple CPU XEON

FrancoFDF — Sun, 18 Feb 2024 08:32:09 GMT

Ok thank you for both replies, to follow your latest comment instead of dealing with the old installation I update my compiler by installing the following :

w_Base_Kit_2024.0.1.45

w_fortran-compiler_p_2024.0.2.27

integrating with visual studio 2022. Initially the integration in visual studio fail, so following other people comments on the forum I downgraded from visual studio 2022.17.8 back to visual studio 2022.0.x.

After that all the installation went well no errors and no complains, but when I start visual studio I get the following message

The 'oneApiPackage' package did not load correctly (see attached picture) so I am wondering what went wrong ?

After that I will try to see if the 8k bug disappear in the new MKL

Thanks

Franco

Re:MKL on multiple CPU XEON

Gennady_F_Intel — Mon, 19 Feb 2024 05:32:32 GMT

please create the new thread to follow with new topic and set the new TAG there . The current topic is resolved.

--Gennady