Speed comparation: the new versions of cp2k are slower than the old

[li lincai]

Dear cp2k developers:

I found that the new versions of CP2K are slower than the old ones. I installed CP2K-2023.1, 2023.2, and 2025.1 on my computer. CP2K-2025.1 is much slower than CP2K-2023.1 when running the same task. 

My computer's CPU is EPYC-7V12, and my system is Rocky linux 9.2. I tested at least two systems; one is Pt-water system ( about 300 atoms), and another is the ionic liquid system (about 500 atoms, including C H O F S N), the results are same that CP2K-2025.1 is much slower than CP2K-2023.1. For example,  CP2K-2025.1 need 2.7 s while CP2K-2023.1. just need 1.9 s for one SCF step.
I compiled CP2K with GCC 11 and  MKL math library. 
Do these default parameters  has improved their accuracy, or I wrongly compiled my CP2K?

[Jürg Hutter]

Hi
Can you have a look at the detailed timings at the end of your test jobs?
Do you see a specific routine causing the slow down or are all parts running slower?
If you post that information we can have a deeper look?
regards
JH

________________________________________
From: cp...@googlegroups.com <cp...@googlegroups.com> on behalf of li lincai <lilin...@gmail.com>
Sent: Tuesday, February 25, 2025 2:54 AM
To: cp2k
Subject: [CP2K:21204] Speed comparation: the new versions of cp2k are slower than the old

--
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<mailto:cp2k+uns...@googlegroups.com>.
To view this discussion visit https://groups.google.com/d/msgid/cp2k/f7011633-4f1d-4ea3-841e-9744f37b6d3dn%40googlegroups.com<https://groups.google.com/d/msgid/cp2k/f7011633-4f1d-4ea3-841e-9744f37b6d3dn%40googlegroups.com?utm_medium=email&utm_source=footer>.

[li lincai]

I have tested a Pt-IL (ionic liquid) system on my workstation, and the input and log file are attached as below: Is is obvious that cp2k2025.1 is slower than cp2k2023.1.

[Jürg Hutter]

Hi

Yes, it is obvious, but why? Can you run the same job with both versions
such that the jobs finish correctly? Then there will be detailed timing
outputs that will give the necessary information to see what is going on.

regards
JH

________________________________________
From: cp...@googlegroups.com <cp...@googlegroups.com> on behalf of li lincai <lilin...@gmail.com>

Sent: Monday, April 7, 2025 9:31 AM
To: cp2k
Subject: Re: [CP2K:21333] Speed comparation: the new versions of cp2k are slower than the old

To view this discussion visit https://groups.google.com/d/msgid/cp2k/71c11b2b-4455-40fc-8138-a659187a241bn%40googlegroups.com<https://groups.google.com/d/msgid/cp2k/71c11b2b-4455-40fc-8138-a659187a241bn%40googlegroups.com?utm_medium=email&utm_source=footer>.

[li lincai]

I have tested two systems; one has atoms, and the other has 744 atoms. The two tasks ended normally, and their log information and input files are attached. 
Thanks for cp2k developers' answers.

[Jürg Hutter]

Hi

thank you for the outputs.
It looks to me that the main part of the difference comes from MPI library calls.
Check the timings for "mp_alltoall_z22v".
Can you check what MPI is used for the different versions of CP2K?

regards
JH

________________________________________
From: cp...@googlegroups.com <cp...@googlegroups.com> on behalf of li lincai <lilin...@gmail.com>

Sent: Friday, April 11, 2025 8:27 AM
To: cp2k
Subject: Re: [CP2K:21361] Speed comparation: the new versions of cp2k are slower than the old

To view this discussion visit https://groups.google.com/d/msgid/cp2k/63d07639-49cc-46d4-8cdf-8c75e4453fe2n%40googlegroups.com<https://groups.google.com/d/msgid/cp2k/63d07639-49cc-46d4-8cdf-8c75e4453fe2n%40googlegroups.com?utm_medium=email&utm_source=footer>.

[li lincai]

Hi 

  Thanks for your message!

Both CP2K versions are installed by toolchain. OpenMPI 4.14 for CP2K23.1 and OpenMPI 5.06 for CP2K25.1.

To manage multiple versions of CP2K, each version is separated in the .bashrc file to avoid messes. So, I can promise that the OpenMPI versions and CP2K versions are strictly corresponded. 

regards 

LCL

[li lincai]

I tried to compile CP2K2025.1 with OpenMPI 4.14, and it solved this problem. Maybe my computer is not very compatible with the high versions of OpenMPI.