Comparison between CP2K and Quantum Espresso

[tfran...@gmail.com]

Dear all,

I am currently testing my new machine, which is a cluster 2 Supermicro 2048U RT4 nodes with 4x2.6GHZ Intel Broadwell 10Core CPUs.

The point is that in order to check the scalability and the performances of the machine with 40 processors, I run several single point calculations of the same structure both with CP2K version 4.1 ssmp and Quantum Espresso 5.1.1 

It turns out that for the time spent by CP2K to converge a single point calculation with 40 processors, for a system of 256 atoms, was of the order of 13316.63 seconds (taken from the output in the TIMING BLOCK in the end of the output file [TOTAL TIME AVERAGE); whereas for the QE, the same calculation, with the same system ect. took just 1845.76 seconds. 

So, because we strongly want to work with CP2K, also for running some dynamics and exploit its scalability, may i ask you to kindly check if the input file of the CP2K I used is properly set? I am more expert in QE rather than CP2K, so I might have miss something essential.

For matter of comparison I attached the two input files used, one for the CP2K and the other of the QE.

Thanks in advance for the kind help!

Best Regards,

Tommaso 

[hut...@chem.uzh.ch]

Hi

without additional information (e.g. output files or complete input)
I will have to do guessing.

There are two parts to this calculation
- time per SCF step
- number of SCF steps needed for convergence

SCF convergence: you start from an atomic guess (CP2K) and a random
wfn (QE). Although the convergence asked for is the same (1.e-7)
this means different things in the two codes (check change in energy
for the last step). For CP2K you can optimize the settings for the
convergence, but as you do start from scratch only once the more
important part is what is the timing per step.
(You probably also don't need CHOLESKY OFF?,
you use the most expensive PRECONDITIONER,
DIIS might not be good to start with ...)

Your settings are ok, but rather tight in many respect, e.g. large and
diffuse basis set, high cutoff. Non orthorhombic cell is also not
optimal for CP2K.
So this is a very expensive calculation in CP2K standards.

regards

Juerg

--------------------------------------------------------------
Juerg Hutter                         Phone : ++41 44 635 4491
Institut für Chemie C                FAX   : ++41 44 635 6838
Universität Zürich                   E-mail: hut...@chem.uzh.ch
Winterthurerstrasse 190
CH-8057 Zürich, Switzerland
---------------------------------------------------------------

-----cp...@googlegroups.com wrote: -----To: cp2k <cp...@googlegroups.com>
From: tfran...@gmail.com
Sent by: cp...@googlegroups.com
Date: 02/05/2017 08:06PM
Subject: [CP2K:8645] Comparison between CP2K and Quantum Espresso

Dear all,I am currently testing my new machine, which is a cluster 2 Supermicro 2048U RT4 nodes with 4x2.6GHZ Intel Broadwell 10Core CPUs.The point is that in order to check the scalability and the performances of the machine with 40 processors, I run several single point calculations of the same structure both with CP2K version 4.1 ssmp and Quantum Espresso 5.1.1 It turns out that for the time spent by CP2K to converge a single point calculation with 40 processors, for a system of 256 atoms, was of the order of 13316.63 seconds (taken from the output in the TIMING BLOCK in the end of the output file [TOTAL TIME AVERAGE); whereas for the QE, the same calculation, with the same system ect. took just 1845.76 seconds. So, because we strongly want to work with CP2K, also for running some dynamics and exploit its scalability, may i ask you to kindly check if the input file of the CP2K I used is properly set? I am more expert in QE rather than CP2K, so I might have miss something essential.

For matter of comparison I attached the two input files used, one for the CP2K and the other of the QE.
Thanks in advance for the kind help!


















Best Regards,Tommaso 

--

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+uns...@googlegroups.com.

To post to this group, send email to cp...@googlegroups.com.

Visit this group at https://groups.google.com/group/cp2k.

For more options, visit https://groups.google.com/d/optout.



[attachment "cp2k.inp" removed by Jürg Hutter/at/UZH]
[attachment "qe.inp" removed by Jürg Hutter/at/UZH]

[tfran...@gmail.com]

Dear Prof. Hutter,

first of all, thanks a lot for your kind reply.

Il giorno lunedì 6 febbraio 2017 16:04:41 UTC+1, jgh ha scritto:

Hi

without additional information (e.g. output files or complete input)
I will have to do guessing.

There are two parts to this calculation
- time per SCF step
- number of SCF steps needed for convergence

SCF convergence: you start from an atomic guess (CP2K) and a random
wfn (QE). Although the convergence asked for is the same (1.e-7)
this means different things in the two codes (check change in energy
for the last step). For CP2K you can optimize the settings for the  

convergence, but as you do start from scratch only once the more
important part is what is the timing per step.
(You probably also don't need CHOLESKY OFF?,
 you use the most expensive PRECONDITIONER,
 DIIS might not be good to start with ...)

Ok, i will check this part in detail in order to better tune the input to my and to get rid of the "most" expensive part as much as I can.

 

Your settings are ok, but rather tight in many respect, e.g. large and
diffuse basis set, high cutoff. Non orthorhombic cell is also not
optimal for CP2K.

Ok, I will decrease the settings parameters in order to allow a better flexibility in terms of the cutoff eat. 

Unfortunately all the systems I am dealing with are or monoclinic or triclinic. Now that I am aware of this pitfall, I will reset all the parameters in order to enhance the jobs as much as they can.

 

So this is a very expensive calculation in CP2K standards.

regards

Juerg

Thanks a lot again for clarifying me these crucial points!

Best Regards,

Tommaso 

[David T]

Hi Tommaso

as you I am more expert on QE and only recently moved on CP2K.
On my experience CP2K is much more quicker and allows to work with bigger systems that planewave code could not afford. On the other hand it is true that this is really system depended so the fact that for nano-porous material CP2K is more efficient could not be true for other systems.
After having read the inputs a comment I can give to you is that you are probably not making a fair comparison.
I've see you use ultrasoft pseudo which allows a very low cut-off and can be "though" as minimal basis-set. So you should probably use some corresponding low basis-set in CP2K (for instance a dobule zera instead of triple zeta).

Probably more important another thing I can tell you is that I've seen that if CP2K is not properly compiled, its performance can be slow. For instance my own version with mpi, mkl, libxsmm and elpa is about 30% faster than the standard one I found on our cluster.
Moreover the speediness of the code can be further boosted if you use also GPU and hybrid openMP-MPI.

P.S. a curiosity, in QE why are you using a single point shifted from gamma? if there is not a major reason using the gamma algorithm on QE can accelerate you calc up to 30%

[tfran...@gmail.com]

Dear David,

first of all thanks a lot for your very useful comments.

I will benchmark my system in CP2K with a lower basis sets, in order to test again the parameters and to have better reference with respect to QE. 

Regarding the compilation of the program in our cluster, I think that it is the pitfall and that our technician missed something. He told me ( because I am not allowed to install anything, but only the technician ) that he compiled the precompiled version of CP2K. But I am pretty sure he missed some of the libraries you listed above. Indeed I am insisting in a more "serious" compilation of the program, that we wanna massively exploit for our dynamics. 

Regarding the single point shift from the gamma point in QE, that was suggested by one of my old Professors. At that time ( almost three years ago) he told us that the shift is recommended. 

I always used that setting, without honestly taking care too much about the change in performances. By the way, thanks a lot for the really useful hints. Now I will re-benchemark the new machine testing several levels of theory as you suggested.

Best Regards,

Tommaso Francese

[Ari Paavo Seitsonen]

Dear Tommaso and David,

  I would definitely _not_ call using Vanderbilt pseudo potentials as having "minimal basis set", usually the precision is better than with norm-conserving pseudo potentials (due to several reference energies/projectors). The reduction of cut-off is due to a good reason. So lowering the basis set in CP2K would be unfair, as the results are of lower precision. Rather using GAPW (instead of GPW) would have kind-of the same effect, as this would allow a lower cut-off energy being used in the plane wave part of QuickStep.

  The idea of shifting the k point might refer to the scheme of Baldereschi's special point (I do not know much, I was just told that this is the "best" point if using a single point; since most people do not know it and Monkhorst-Pack is so "black box"-convenient, younger generations have probably not become users of it). Just my guess.

    Greetings from Paris,

       apsi

--

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+unsubscribe@googlegroups.com.

To post to this group, send email to cp...@googlegroups.com.
Visit this group at https://groups.google.com/group/cp2k.
For more options, visit https://groups.google.com/d/optout.

--

-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-
  Ari Paavo Seitsonen / Ari.P.S...@iki.fi / http://www.iki.fi/~apsi/
    Ecole Normale Supérieure (ENS), Département de Chimie, Paris
    Mobile (F) : +33 789 37 24 25    (CH) : +41 79 71 90 935

[Tommaso Francese]

Dear Apsi,

thanks for the reply. I will take into considerations also your suggestions!

Best,

Tommaso

You received this message because you are subscribed to a topic in the Google Groups "cp2k" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/cp2k/YXGli4aI2kw/unsubscribe.
To unsubscribe from this group and all its topics, send an email to cp2k+uns...@googlegroups.com.