Need help with making AIMD simulations faster!

[Akanksha]

Hi,

I am trying to run an AIMD simulation of an explicitly solvated system of ~280 atoms at 300 K and 1 bar pressure. However, I have noticed my simulation is very slow (Only reached about 200 fs in a day). I do have a few motion constraints on my solute molecules, which I believe could also make the simulations slower. 
Each MD step takes about 200 seconds to run with 6-10 SCF iterations. I am new to CP2K and AIMD in general, so I am not sure if this speed is normal. Could anyone please help me figure out if I am missing something or if there are any ways to make these simulations faster? Your help will be very much appreciated.

Here are the parameters I am using:
1) CP2K/2024.2 version (mpirun -np 32 -c 4 cp2k.psmp)
2) Anthracene and 2-bromophenol in a DMF box (17*14*14)
2) Calculation started from the last snapshot of a 5 ns classical MD run.
2) PBE-D3(BJ)/DZVP-MOLOPT/GTH 
3) ASPC(3) extrapolation
4) DIIS as the OT minimizer with FULL_SINGLE_INVERSE preconditioner

5) Timestep: 0.5 fs
5) CSVR thermostat with 50 fs time constant.

I have also attached the full input file and a plot of the constant-of-motion wrt time for your reference.

Please let me know if you need more information from me. I would really appreciate any feedback you can provide.

Thank you,

Akanksha

[Krack, Matthias]

Hi

 

At a first glance your setup looks fine to me. I don’t understand why it runs so slow. Maybe, you are overcommitting the available resources which often causes a significant slowdown. I suggest to use only one OMP thread per MPI rank. Running your input using an increasing number of CPU cores shows a reasonable scaling up to 64 CPU cores (cp2k v2025.2):

 

CPU cores (MPI ranks)	Average time per MD step [sec]
8	78
16	40
32	22
64	14
128	12

 

Best

 

Matthias

 

--
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 view this discussion visit https://groups.google.com/d/msgid/cp2k/ffb6cc93-0861-43cb-ba6e-047ac09c83e5n%40googlegroups.com.

[Akanksha]

Hello Dr. Krack,

Thank you for responding so quickly. 
Based on your suggestion, I ran a few benchmarking tests and was able to identify the issue. I realized my SLURM script was getting overwritten during submission, which resulted in all of my calculations running on a single core.

I really appreciate your time and guidance. Thank you.

Regards,
Akanksha

To view this discussion visit https://groups.google.com/d/msgid/cp2k/ZR2P278MB11463EBBD559E46F49842AD7F403A%40ZR2P278MB1146.CHEP278.PROD.OUTLOOK.COM.