Performance of Sevenn model in LAMMPS

[thangckt]

hi @YutackPark,

Can I have a question about the performance of running MD in LAMMPS.

I try a test with only 6 atoms in system, use 1 GPU to run LAMMPS with Sevenn model. with command

lmp -i input

Just run 30000 steps of NVT without any compute.

The total is almost 15 minutes.

Does this amount of time is reasonable?
I feel so slow.
Do you have any trick to deploy model for using in LAMMPS?

This is the output from LAMMPS:

Generated 0 of 1 mixed pair_coeff terms from geometric mixing rule
Per MPI rank memory allocation (min/avg/max) = 3.111 | 3.111 | 3.111 Mbytes
   Step          Temp          PotEng           Lx             Ly             Lz            Pxx            Pyy            Pzz      
         0   300           -45.354305      43.18          3.170347       2.7456011      2631.628      -47.045165     -191.52913    
      5000   353.41104     -45.372852      43.18          3.170347       2.7456011      6050.4796      1722.6888      1694.3101    
     10000   212.67402     -45.251991      43.18          3.170347       2.7456011      10696.388      2433.3746      5245.0921    
     15000   288.2721      -45.299927      43.18          3.170347       2.7456011      13409.972      5110.724       4350.4961    
     20000   261.78924     -45.266258      43.18          3.170347       2.7456011      12449.008      5292.5313      3094.6084    
     25000   206.22299     -45.227859      43.18          3.170347       2.7456011      14115.876      4913.9564      5119.1184    
     30000   227.59058     -45.255264      43.18          3.170347       2.7456011      15032.484      6413.9024      4453.1244    
Loop time of 779.865 on 1 procs for 30000 steps with 6 atoms

Performance: 3.324 ns/day, 7.221 hours/ns, 38.468 timesteps/s, 230.809 atom-step/s
99.3% CPU use with 1 MPI tasks x 1 OpenMP threads

MPI task timing breakdown:
Section |  min time  |  avg time  |  max time  |%varavg| %total
---------------------------------------------------------------
Pair    | 779.44     | 779.44     | 779.44     |   0.0 | 99.95
Neigh   | 0.0013798  | 0.0013798  | 0.0013798  |   0.0 |  0.00
Comm    | 0.14292    | 0.14292    | 0.14292    |   0.0 |  0.02
Output  | 0.00011363 | 0.00011363 | 0.00011363 |   0.0 |  0.00
Modify  | 0.10541    | 0.10541    | 0.10541    |   0.0 |  0.01
Other   |            | 0.1736     |            |       |  0.02

[YutackPark]

Total throughput of GPU decreases if is unsaturated (underutilized). This is different from CPU, which gives almost linear speed up or down with respect to the number of atoms in the simulation cell. For this reason, GPU could give almost the same or even slower speed compared to single core CPU. When I have to do MD simulations or relaxation with very small, different systems, I use multi core CPU to distribute the job, or write batched code in python to fully saturate GPU. You can see relevant figure in supporting information of SevenNet JCTC paper.

…

----------------------------------------------------------------- *Yutack Park* Ph.D. Candidate Materials Data and Informatics Lab. Department of Materials Science and Engineering Seoul National University *E-mail:* ***@***.*** ***@***.***> *Homepage:* http://mtcg.snu.ac.kr ----------------------------------------------------------------- 2025년 3월 20일 (목) 오후 10:21, C. Thang Nguyen ***@***.***>님이 작성:

hi @YutackPark <https://github.com/YutackPark>, Can I have a question about the performance of running MD in LAMMPS. I try a test with only 6 atoms in system, use 1 GPU to run LAMMPS with Sevenn model. with command lmp -i input Just run 30000 steps of NVT without any compute. The total is almost 15 minutes. Does this amount of time is reasonable? I feel so slow. Do you have any trick to deploy model for using in LAMMPS? — Reply to this email directly, view it on GitHub <#190>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A2RQY25P7A6PWJG3H36PTMD2VK6F3AVCNFSM6AAAAABZNN4CSCVHI2DSMVQWIX3LMV43ERDJONRXK43TNFXW4OZYGEYDSNBYGA> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[thangckt]

To run with CPU only, do we need to re-compile a separate lammps version? or can we use the version compiled with GPU?

Thanks.

[YutackPark]

You can use the same. There is no CPU-specialized compile option at this moment. and I think intel won't make it as GPU era comes...

…

----------------------------------------------------------------- *Yutack Park* Ph.D. Candidate Materials Data and Informatics Lab. Department of Materials Science and Engineering Seoul National University *E-mail:* ***@***.*** ***@***.***> *Homepage:* http://mtcg.snu.ac.kr ----------------------------------------------------------------- 2025년 3월 20일 (목) 오후 10:38, C. Thang Nguyen ***@***.***>님이 작성:

To run with CPU only, do we need to re-compile a separate lammps version? or can we use the version compiled with GPU? Thanks. — Reply to this email directly, view it on GitHub <#190 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A2RQY2ZPB2BHDYISKBTUD3T2VLAD7AVCNFSM6AAAAABZNN4CSCVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTENJWGQZTINA> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[thangckt]

hi @YutackPark

Thank you so much for your guide.
Can you have a little detail on how to disable Sevenn using GPU when it run?

I use Slurm for scheduling jobs, and even I do not ask GPU resource, the GPU existed on the node, and Sevenn automatically pickup it.

Do we have any way to "turn off" GPU (or tell Sevenn just use CPU)? For example export some runtime VARs?
Thanks.

[YutackPark]

In LAMMPS or Python, `export CUDA_VISIBLE_DEVICES= ` Alternatively, in python, you always have an appropriate 'device' argument to force 7net to use CPU. You can talk to your admin to automatically do the above behavior.

…

----------------------------------------------------------------- *Yutack Park* Ph.D. Candidate Materials Data and Informatics Lab. Department of Materials Science and Engineering Seoul National University *E-mail:* ***@***.*** ***@***.***> *Homepage:* http://mtcg.snu.ac.kr ----------------------------------------------------------------- 2025년 3월 21일 (금) 오후 1:49, C. Thang Nguyen ***@***.***>님이 작성:

hi @YutackPark <https://github.com/YutackPark> Thank you so much for your guide. Can you have a little detail on how to disable Sevenn using GPU when it run? I use Slurm for scheduling jobs, and even I do no accquire GPU resource, the GPU existed on the node, and Sevenn automatically pickup it. Do we have any way to "turn off" this? For example export some runtime VARs? Thanks. — Reply to this email directly, view it on GitHub <#190 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A2RQY23PHNOFMUFASF5KAIT2VOK5FAVCNFSM6AAAAABZNN4CSCVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTENJXGI2DQNA> . You are receiving this because you were mentioned.Message ID: ***@***.***>