Quite different performace of HOOMD on simulations with L-J potential

Hao Dong

unread,

Mar 31, 2016, 6:16:58 AM3/31/16

to hoomd-users

Hello all,

I tested the simulations with two set of parameters by using the L-J potential. They are "epsilon=1.0, sigma=1.0" in job-1, and "epsilon=2.0, sigma=1.0" in job-2. All of the rest parameters are the same in the two jobs. They were run separately on the same machine with 2 GeForce GTX Titan X GPUs, and the CPU is Intel xeon E5-2630 v3 2.4GHz (2-cores). And the version I used is HOOMD-blue 1.3.3.

Here is the command I used to submit job: mpirun -np 2 hoomd job.py --mode=gpu >& job.log &

However, the performance of the two jobs are quite different: The calculation speed for job-1 is pretty stable, with a "TPS" around 2800. But for job-2, the performance drops significantly from "TPS 2800" to "TPS 180" after 20e5 steps run.

So I checked the trajectory of the two simulations, and found that most of the particles in job-2 aggregate into a large cluster (as expected), while in job-1 they are randomly distributed (I know that this description is inaccurate). Probably my setup for the neighbor list has some problems. So I checked the "nlist" reference, and tested different parameters including r_buff, check_period, etc. But none of them shows a TPS more than 200.

Could anybody tell me what is wrong with my input? Or how to setup simulations that particles may aggregate into cluster(s)?

For your reference, I paste part of the output from job-2, as well as my input file for job-2 at the end of this post.

Any suggestions is highly appreciated!!

Hao

## my input for job-1

from hoomd_script import *
context.initialize()

# create 5000 random particles of name A
init.create_random(N=5000, phi_p=0.01, name='A')

## specify Lennard-Jones interactions between particle pairs

# specify the pair interaction via a given potential energy and force
def ljg(r, rmin, rmax, epsilon, sigma):
    V = 4 * epsilon * ( (sigma / r)**12 - (sigma / r)**6);
    F = 4 * epsilon / r * ( 12 * (sigma / r)**12 - 6 * (sigma / r)**6);
    return (V, F)
table = pair.table(width=1000);
table.pair_coeff.set('A', 'A', func=ljg, rmin=0.001, rmax=12.0, coeff=dict(epsilon=2.0, sigma=1.0))

# integrate at constant temperature
all = group.all();
integrate.mode_standard(dt=0.005)
integrate.nvt(group=all, T=1.2, tau=0.5)

# dump an xmle file for the structure information
xml = dump.xml(filename='dump_dcd.xml', vis=True)

# dump a .dcd file for the trajectory
dump.dcd(filename='dump_dcd.dcd', period=10000)
#dcd = dump.dcd(filename"data/dump.dcd", period=50)

# run 10,000 time steps
run(10e7)

Joshua Anderson

unread,

Mar 31, 2016, 6:23:15 AM3/31/16

to hoomd...@googlegroups.com

1) Try pair.lj instead of pair.table. It will be faster.

2) Performance will decrease with increasing density of the aggregate, there are more pair forces to evaluate.

3) Run on a single GPU, not two. 5000 particles is barely enough to saturate a single GPU.

------
Joshua A. Anderson, Ph.D.
Research Area Specialist, Chemical Engineering, University of Michigan
Phone: 734-647-8244
http://www-personal.umich.edu/~joaander/

--
You received this message because you are subscribed to the Google Groups "hoomd-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hoomd-users...@googlegroups.com.
To post to this group, send email to hoomd...@googlegroups.com.
Visit this group at https://groups.google.com/group/hoomd-users.
For more options, visit https://groups.google.com/d/optout.

Michael Howard

unread,

Mar 31, 2016, 9:52:16 AM3/31/16

to hoomd-users

I agree 100% with Josh. To follow-up with a comment or two on the physics that you have -- in your system 1, you probably would not expect to get any aggregation (ever) because you have such a low density that it is outside the LJ binodal. You set the packing fraction to 0.01 when you initialize the system, so the density is about 0.02. The critical point of the LJ fluid is about Tc ~ 1.3 and rhoc ~ 0.3, so at T = 1.2 ~ 0.9Tc, you might guess that your density is so low that you are just in the gas phase. In system 2, where you double epsilon, you are effectively halving the temperature so that Teff ~ 0.6. The vapor density for the LJ fluid at this temperature is very low (almost zero), so you expect aggregation. The coexistence density at T ~ 0.6 is about 0.8, so the number of neighbors is 40x higher in the liquid phase of system 2 than the gas of system 1... and you must expect a corresponding slow down.

Regards... Mike

Reply all

Reply to author

Forward