[Beowulf] Here we go again

[Bill Broadley]

On 12/12/19 6:35 AM, Douglas Eadline wrote:
> 
> Anyone see anything like this with Epyc, i.e. poor AMD performance
> when using Intel compilers or MKL?
> 
> https://www.pugetsystems.com/labs/hpc/AMD-Ryzen-3900X-vs-Intel-Xeon-2175W-Python-numpy---MKL-vs-OpenBLAS-1560/

I as getting anomalously slow performance with Matlab on AMD Rome.  Turns out 
Matlab uses the MKL and Intel ignores the AMD vector unit, despite the CPU 
capability flag showing that AMD support AVX2-256.

Turns out "export MKL_DEBUG_CPU_TYPE=5" fixed it.  The improvement is highly 
variable, but in one case was 6x faster.

Awhile back I tinkered with AOCC, AMD's tweak of clang/flang for Zen.  Had 
issues compiling OpenMPI (and dependencies), fftw3, OpenBLAS and friends.  I 
tried again with AOCC-2.1 and had no problems.  Generally things are looking 
good and we are considering supporting an AOCC based platform for our users to 
make the most of the AMD Rome chips.  Kudos to AMD for working on AOCC and 
working on upstreaming the changes to the numerous projects involved.

Related projects of possible interest:

  AMD FFTW3 = https://github.com/amd/amd-fftw

  BLIS/BLAS = https://github.com/amd/blis

  AMD Math library/Libm = 
http://developer.amd.com/amd-cpu-libraries/amd-math-library-libm/

  AMD LAPACK/Libflame = https://github.com/amd/libflame

  AMD Clang/Flang/LLVM = https://developer.amd.com/amd-aocc/