search

glotzerlab / hoomd-blue

Molecular dynamics and Monte Carlo soft matter simulation on GPUs.
http://glotzerlab.engin.umich.edu/hoomd-blue
BSD 3-Clause "New" or "Revised" License
335 stars 131 forks source link

Issues running table potential (v3.0.0beta008) on GPU #1105

Closed YouriRan closed 3 years ago

YouriRan commented 3 years ago

Description

While running a membrane simulation with tabulated potential for a cosine/squared potential. The script will run locally on CPU, but on a GPU node the process exits with an Error computing cell list (particle no longer in simulation box).

This system runs to completion with a LJ-potential in stead of a table potential on GPU (albeit not with 'chemical' stability). I run into the same GPU issues with the cosine/squared as a table potential on the molecular dynamics tutorial.

Another issue I have run into is when formulating a zeropotential for the table (V=F=np.array(n*[0]) ) a cudaErrorIllegalAddress occurs on GPU. This too will run on CPU. This zero potential is quite important when using multiple types of potentials between pairs.

Script

Included files at groups:

* init.gsd (initial coordinates)
* test.py (hoomd script)
* exectest.sh (gpu batch script)
* nvidia-smi.txt (output nvidia-smi)
* errorCellList.txt (error message 1)
* cudaErrorIllegalAddress.txt (error message 2)

Output

In included files.

Expected output

Stable simulation run. Can provide a trajectory ran on CPU if necessary.

Configuration

OS: Debian 10 GPU: GTX1080ti HOOMD version: 3.0.0dev8 Build: conda gpu_py39hc6c7ca5_0 Python version: 3.9.6

Developer

joaander commented 3 years ago

This is fixed by #1146 not #1145.