speed in scaling up

[Jiahao Zhang]

Dear people,

I was running molecular dynamics in aenet. The speeding of Aenet is relatively slow.

"Loop time of 627.95 on 48 procs for 5000 steps with 2592 atoms"

Any suggestion to improve it?

Best wishes

Jiahao

[Artrith, N. (Nong)]

Dear Jiahao,

Which MD codes did you use?

Best wishes,

Nong

=========================

Nongnuch Artrith, 

Assistant Professor

Debye Institute for Nanomaterials Science, Utrecht University

David de Wiedgebouw, Office 4.66, Universiteitsweg 99, 3584 CG Utrecht

Email: n.ar...@uu.nl

Phone: +31-6-28-15-97-99

Web: https://www.uu.nl/staff/NArtrith

---

From: ae...@googlegroups.com <ae...@googlegroups.com> on behalf of Jiahao Zhang <jia...@sas.upenn.edu>
Sent: Wednesday, February 2, 2022 7:32 PM
To: aenet <ae...@googlegroups.com>
Subject: [aenet] speed in scaling up

 

--
You received this message because you are subscribed to the Google Groups "aenet" group.
To unsubscribe from this group and stop receiving emails from it, send an email to aenet+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/aenet/13914a39-0ab4-4148-92c3-7799252e1501n%40googlegroups.com.

[Jiahao Zhang]

I used lammps. I also did a serial test.

30% efficiency for 48 processes compared to 1 proc

Best wishes

Jiahao

[Artrith, N. (Nong)]

Hi Jiahao,

The LAMMPS interface should be quite efficient (J. Chem. Phys. 155, 074801 (2021); https://doi.org/10.1063/5.0063880).  Though, I am not surprised by the 30% parallel efficiency because of the small number of atoms. For 5,000 atoms, 1-8 cores should be more than enough.

Upon reading your message again, I have a few more questions in addition to the MD code:

• What does “loop time” mean? Do you mean the computer time for a single MD time step?
• What is the unit of the time in your message? Is the time in seconds? Or seconds/core?
• What compiler did you use, and how exactly did you run the parallel simulation?
• What is your ANN architecture? And which descriptor/featurization method did you use?
• For the benefit of the aenet community, could you please sign your messages with your affiliation/status?

Best,

Nong

=========================

Nongnuch Artrith, 

Assistant Professor

Debye Institute for Nanomaterials Science, Utrecht University

David de Wiedgebouw, Office 4.66, Universiteitsweg 99, 3584 CG Utrecht

Email: n.ar...@uu.nl

Phone: +31-6-28-15-97-99

Web: https://www.uu.nl/staff/NArtrith

---

From: Jiahao Zhang <jia...@sas.upenn.edu>
Sent: Wednesday, February 2, 2022 7:42 PM
To: Artrith, N. (Nong) <n.ar...@uu.nl>
Cc: aenet <ae...@googlegroups.com>
Subject: Re: [aenet] speed in scaling up

 

[Jiahao Zhang]

Dear Prof. Nong,

We figure out that in the github https://github.com/HidekiMori-CIT/aenet-lammps

In the step 4:

make -f makefiles/Makefile.gfortran_serial lib

It’s actually using a serial version of annet.  Do you have a tutorial how to compile parallel?

Best wishes

Jiahao

To view this discussion on the web visit https://groups.google.com/d/msgid/aenet/DBAPR05MB7477C2AA9A2A0150649D5DE6F6279%40DBAPR05MB7477.eurprd05.prod.outlook.com.

[Artrith, N. (Nong)]

Hi Jiahao,

It is correct that LAMMPS is linked with the serial version of 'libaenet'.  LAMMPS takes care of the parallelization.

I see that you are using the GNU Fortran compiler.  LAMMPS is much more efficient when compiled with the Intel Fortran compiler and Intel MPI, if you have access to it.  The Intel compilers are available on most computer clusters?

Best,

Nong

Did you also take a look at these tutorials? The notebook can be obtained from the provided GitHub links [1] and can also be run online via Google Colaboratory [2] aenet-lammps (a tutorial of short MD simulation liquid bulk water):

[1] https://github.com/atomisticnet/aenet-lammps/tree/master/tutorial 

[2] https://colab.research.google.com/drive/1Km8JVFM2DCeElAeE2n_WgMgLa7IU_IFh?usp=sharing

---

From: Jiahao Zhang <jia...@sas.upenn.edu>
Sent: Wednesday, February 2, 2022 8:20 PM

To: Artrith, N. (Nong) <n.ar...@uu.nl>

Cc: aenet <ae...@googlegroups.com>; Drew Behrendt <drew...@sas.upenn.edu>