Bug #1426
Group cutoff scheme yielding zero Coul-SR for SOL-SOL and tip4p, whereas verlet scheme does not
Description
I had problems with unexpected LINCS errors appearing in my simulation of a protein dimer in water+NaCl. The error only occurred with group cutoffs, however. I eventually discovered that the SOL-SOL short-range potential energy was zero with the group cutoff, but not with the verlet cutoff. amber99sb-ildn and tip4p-Ew was used to model the system.
Here are the energies after a few em steps with verlet and group kernels, respectively. mdp files are nearly identical:
[marklund@ajbchap3 em2]$ diff em_group.mdp em.mdp 72c72 < cutoff-scheme = group --- > cutoff-scheme = verlet 100c100 < vdw-modifier = Potential-shift --- > vdw-modifier = Potential-shift-verlet group Energy Average Err.Est. RMSD Tot-Drift ------------------------------------------------------------------------------- Coul-SR:SOL-SOL 0 0 -nan 0 (kJ/mol) verlet Energy Average Err.Est. RMSD Tot-Drift ------------------------------------------------------------------------------- Coul-SR:SOL-SOL -948221 15000 32896 -104103 (kJ/mol)
History
#2 Updated by Erik Marklund almost 6 years ago
- File em.tpr em.tpr added
- File em_group.tpr em_group.tpr added
Tpr files uploaded.
#3 Updated by Berk Hess over 5 years ago
Is this running on a machine with a GPU?
With the Verlet scheme on a GPU you don't get non-bonded energies split out per groups. Is that the issue here?
#4 Updated by Szilárd Páll over 5 years ago
Berk Hess wrote:
Is this running on a machine with a GPU?
With the Verlet scheme on a GPU you don't get non-bonded energies split out per groups. Is that the issue here?
The 0 energies were supposedly produced with the group scheme, hence I assume no GPU-s were used with the verlet scheme run.
#5 Updated by Erik Marklund over 5 years ago
No GPUs.
#6 Updated by Erik Lindahl over 5 years ago
- Status changed from New to Resolved
- Priority changed from High to Normal
I get realistic energies with the latest git version of the 4.6 branch. This was likely caused by combining shifted LJ with no modifier for PME, which might have led to the use of generic kernels, and then this was fixed by the previous patch that disabled solvent optimization for generic kernels ( https://gerrit.gromacs.org/#/c/3279/ ).
If you still think it's a problem, make sure that it's reproducible with the latest git version of 4.6.
#7 Updated by Erik Marklund over 5 years ago
That's good enough for me. I was merely comparing the performance of the group and verlet schemes for this system anyway.
#8 Updated by Rossen Apostolov over 5 years ago
- Status changed from Resolved to Closed