Reax/c efficiency of simulations (speed)

Hi,

I am new in working with Reax/c force field.
My system contains about 9000 atoms. I am running on 96 CPUS of 2.80 GHz.

For Reax/c, the system could run ~3333 steps in one hour, while could run more than 1,000,000 steps (timestep 1) with COMPASS FF.

I was wondering if the reax FF is really this much slow compared to others (due to complicated calculations involved).
Your suggestions for efficiency improvements are appreciated.

If you need further info, I can present.

Thank you
-X

Hi,

I am new in working with Reax/c force field.
My system contains about 9000 atoms. I am running on 96 CPUS of 2.80 GHz.

​do you actually get sufficient strong scaling for this many CPUs?​

For Reax/c, the system could run ~3333 steps in one hour, while could run
more than 1,000,000 steps (timestep 1) with COMPASS FF.

I was wondering if the reax FF is really this much slow compared to others
(due to complicated calculations involved).
Your suggestions for efficiency improvements are appreciated.

ReaxFF is very slow compared to conventional force fields. please see the
force field speed comparison posted on the LAMMPS home page here:

​axel.

The current ADF version of ReaxFF, on my Dell T7910 36 core (dual 18 core Intel Xeon Processors E5-2699 v3 2.3 GHz cpus), for a 19757 atom system (NVT) runs at ~18,000 steps/hr with a step size of 0.25 fs at a density of 1 g/ml.

You did not provide any information on your environment (cluster vs workstation, interconnect if a cluster, etc.) If a cluster, and you are using Ethernet interconnect, you can expect slower processing speeds.

Jim

James Kress Ph.D., President

The KressWorks® Foundation

An IRS Approved 501 ©(3) Charitable, Nonprofit Organization

Engineering The Cure” ©

(248) 605-8770

Learn More and Donate At:

Website: http://www.kressworks.org

Facebook: https://www.facebook.com/KressWorks.Foundation/

Twitter: @KressWorksFnd

Confidentiality Notice | This e-mail message, including any attachments, is for the sole use of the intended recipient(s) and may contain confidential or proprietary information. Any unauthorized review, use, disclosure or distribution is prohibited. If you are not the intended recipient, immediately contact the sender by reply e-mail and destroy all copies of the original message.

You are getting about 1 step/sec, with 10^4 atoms and 10^2 cores. So that’s a compute rate of about 10^2 atom-steps/core-sec. A more typical rate would be 10^3 atom-steps/core-sec. Because you only have about 100 atoms/core, you are probably losing some efficiency on communication and redundant computation. If you use fewer processors, your rate will increase, but your steps/hour will drop. The Compass potential is running about 300 times faster, which is about the right ratio.

Aidan