Problem with tersoff/cuda

Hello,

I have run across an issue where running LAMMPS with tersoff/cuda gives incorrect energies (or rather stripes in the energy) as a run progresses. (See attachment: “thermo_2400.png”) The green (red) data was computed with cuda turned on (off). In this plot, the x and y axes are the timestep and potential energy, respectively. This simulation was run at 2400K.

The simulations consisted of a spherical nanoparticle of ~2100 germanium atoms (~5nm in diameter).
Simulation cell size: 15nm x 15nm x 15nm with periodic boundary conditions.
Ensamble: NVT
LAMMPS version: 17Feb12
GPU: GTX560

Here is what I have discovered so far…

---------- The energy stripes persisted despite the following changes ----------

Tried reducing size of simulation cell: ±40 in x y and z dimensions.

Tried adding linear and angular momentum fixes:
fix LINMOM all momentum 1 linear 1 1 1
fix ANGMOM all momentum 100 angular

Tried initializing velocities prior to run:
velocity all create 2400 21642 mom yes rot yes dist gaussian

Tried with different nvt damping parameters: 0.1, 0.05 0.01

Tried adding an nvt drag factor: drag = 2.0

Tried modifying neighbor updates:
neigh_modify every 1 check no

Note: Kinetic energies showed no such nodal distributions.

---------- The energy stripes were not present under the following conditions -----------

Works with Cuda OFF. (See attachments)

Works with Cuda ON using, “pair_style tersoff” instead of “pair_style tersoff/cuda”

Works with Cuda ON using, “pair_style lj/cut/cuda 2.5”

----------- Other observations ------------

The number of energy stripes decrease at lower temperatures. (See attachment: thermo_1400.png) This simulation was run at 1400K.

I have also attached the geometry and input files. If you would like, I can send the output file (2.9MB) as well. Although I do not know how to program CUDA, please let me know if there is anything else I can do to help track down and resolve this bug.

Thank you,

– WM

thermo_2400.png

thermo_1400.png

Ge_5nm.geo (163 KB)

in.Ge (579 Bytes)

Hi

thanks for sending that bug report. This should already be plenty of information to hunt down the problem. I'll probably dont manage to do it today thoug.

Cheers
Christian

-------- Original-Nachricht --------