Large comm time with bimodal system

LAMMPS LAMMPS General Discussion
1

Hello,

I am simulating a system with 2 particle types and the mie/cut potential as a “hard sphere”:

pair_coeff 1 1 400.0 20.0 50 49 20.40816
pair_coeff 2 2 1.0 1.0 50 49 1.020408

The system has 50000 atoms of those there are 6 large and 49994 small.

I’m using LAMMPS version “stable_29Aug2024” on 32 CPUs.

I am seeing huge comm contributions to the run time:

MPI task timing breakdown:
Section | min time | avg time | max time |%varavg| %total

Pair | 1567.9 | 1716 | 1892.2 | 199.2 | 3.99
Neigh | 1225.8 | 1317.7 | 1412 | 144.9 | 3.06
Comm | 37492 | 37966 | 38317 | 110.4 | 88.27
Output | 0.022999 | 0.024046 | 0.029119 | 0.6 | 0.00
Modify | 852.92 | 1097.6 | 1554.3 | 564.8 | 2.55
Other | | 914.2 | | | 2.13

My neighbor and comm settings:

comm_style tiled
neighbor 1.0 multi
neigh_modify delay 0 every 1 check yes
comm_modify mode multi

fix fix_balance all balance 20000 0.9 rcb

Are there any suggestions on how to lower the comm timing and perhaps get better run times?

Please let me know if there is any other information needed.

Thanks,
Frank

2

Maybe try comm_modify reduce/multi

3

comm_style tiled has generally more communication overhead. You may want to check if regular comm_style brick isn’t better despite having larger inbalance between MPI ranks

1 Like