[CP2K:10266] Re: ELPA speed-up with Intel-compiled code

[Nico Holmberg]

Hi Jerry,

Just to expand on Juerg's answer, it appears that you are using the OT 
method together with the FULL_SINGLE_INVERSE preconditioner 
<https://manual.cp2k.org/trunk/CP2K_INPUT/FORCE_EVAL/DFT/SCF/OT.html#list_PRECONDITIONER>. 
ELPA won't accelerate such simulations because no direct diagonalization is 
performed during the SCF run.

If you swap the preconditioner to FULL_ALL, you should notice a ~50 % 
improvement in the *first step* of each inner SCF loop iteration when you 
use ELPA instead of ScaLAPACK. Futhermore, if you switch OT to a standard 
diagonalization based SCF solver, each SCF iteration should be accelerated 
by the same factor.  I would however recommend that you always use OT if 
the system is suitable for the method.


BR,

Nico       

torstai 3. toukokuuta 2018 19.24.34 UTC+3 jgh kirjoitti:
>
> Hi 
>
> how much time is used for Diagonalization in these runs? 
> I would guess it is a rather small fraction. We developed the OT 
> method especially to avoid diagonalization as much as possible. 
> If this is in fact the case, no speedup can be expected by using 
> ELPA. 
>
> best regards 
>
> Juerg 
> -------------------------------------------------------------- 
> Juerg Hutter                         Phone : ++41 44 635 4491 
> Institut für Chemie C                FAX   : ++41 44 635 6838 
> Universität Zürich                   E-mail: hut... at chem.uzh.ch 
> <javascript:> 
> Winterthurerstrasse 190 
> CH-8057 Zürich, Switzerland 
> --------------------------------------------------------------- 
>
> -----cp... at googlegroups.com <javascript:> wrote: ----- 
> To: cp2k <cp... at googlegroups.com <javascript:>> 
> From: Jerry Tanoury 
> Sent by: cp... at googlegroups.com <javascript:> 
> Date: 05/03/2018 06:15PM 
> Subject: [CP2K:10266] Re: ELPA speed-up with Intel-compiled code 
>
> Dear Alfio, 
> My apologies for not attaching these at the beginning.  The input and 
> output files are now attached.  Please note that the file names (coordinate 
> files, etc.) have been changed in the input and output files for 
> proprietary reasons.  Also, the ELPA run was killed rather soon after it 
> began because no speed-up was observed. 
>
> Best regards, 
> Jerry 
>
> On Wednesday, May 2, 2018 at 2:57:12 PM UTC-4, Jerry Tanoury wrote: 
> Dear forum, 
> I am running CP2K version 5.1 compiled with Intel 2017 update 5 compilers 
> and corresponding MKL libs.  The arch file is attached.  Everything runs as 
> expected.  I then build ELPA-2016.05.004 as shown below and built an 
> ELPA-enable cp2k version according to the attached arch file: 
>
>  ../configure 
> --prefix=/cluster/home/tanoury/CP2K/intelbuilt_packages/2017u5/elpa-2016.05.004 
> FC=mpiifort FCFLAGS=-O2 -xHost CC=mpiicc CFLAGS=-O2 -xHost 
> --enable-option-checking=fatal --enable-static=yes --enable-avx2=no 
> --enable-avx=no 
> SCALAPACK_LDFLAGS=-L/cluster/home/tanoury/intel/2017u5/compilers_and_libraries_2017.5.239/linux/mkl/lib/intel64 
> -lmkl_scalapack_lp64 -lmkl_intel_lp64 -lmkl_sequential -lmkl_core 
> -lmkl_blacs_intelmpi_lp64 -lpthread 
> -Wl,-rpath,/cluster/home/tanoury/intel/2017u5/compilers_and_libraries_2017.5.239/linux/mkl/lib/intel64 
> SCALAPACK_FCFLAGS=-L/cluster/home/tanoury/intel/2017u5/compilers_and_libraries_2017.5.239/linux/mkl/lib/intel64 
> -lmkl_scalapack_lp64 -lmkl_intel_lp64 -lmkl_sequential -lmkl_core 
> -lmkl_blacs_intelmpi_lp64 -lpthread 
> -I/cluster/home/tanoury/intel/2017u5/compilers_and_libraries_2017.5.239/linux/mkl/include/intel64/lp64 
>
>
> When doing a speed test on 80 cores, I saw no speed-up from ELPA.  Is this 
> unexpected?  Did I build ELPA correctly?  Perhaps I need to run on 100's of 
> cores. 
>
> Thank you for the help, 
> Jerry 
>
>   
>   -- 
>  You received this message because you are subscribed to the Google Groups 
> "cp2k" group. 
>  To unsubscribe from this group and stop receiving emails from it, send an 
> email to cp2k+... at googlegroups.com <javascript:>. 
>  To post to this group, send email to cp... at googlegroups.com <javascript:>. 
>
>  Visit this group at https://groups.google.com/group/cp2k. 
>  For more options, visit https://groups.google.com/d/optout. 
>   
>
> [attachment "CP2K_Test.inp" removed by Jürg Hutter/at/UZH] 
> [attachment "CP2K_Test1-ELPA.inp" removed by Jürg Hutter/at/UZH] 
> [attachment "CP2K_Test1-ELPA.output" removed by Jürg Hutter/at/UZH] 
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20180504/c6e6522f/attachment.htm>