Segfault with non-interacting atoms with Verlet scheme
When Verlet cutoff scheme is used mrun crushes mysteriously if the non-interacting dummy molecule is present in the system. The minimal top file is below. Mdrun crashes despite the fact, that dummy atoms do not interact with anything and should not affect the system in any way:
Tolerance (Fmax) = 1.00000e+00
Number of steps = 5000
Step= 0, Dmax= 1.0e-03 nm, Epot= -nan Fmax= 8.47202e+02, atom= 607
Segmentation fault (core dumped)
When there are only two atoms in the dummy molecule everything is fine. If the third atom (named T) is added mdrun crashes immediately.
If group cutoff scheme is used everything works in both cases with no problem.
[ atomtypes ]
; Dummy atoms not interacting with anything!
;name bond_type mass charge ptype sigma epsilon
SW 12 0.00000 0.00000 A 0.0 0.0
DU 12 0.00000 0.00000 A 0.0 0.0
[ moleculetype ]
[ atoms ]
; nr type resnr residue atom cgnr charge mass
1 SW 1 SW SW 1 0.0 12.0
2 DU 1 SW DU 2 0.0 12.0
3 DU 1 SW T 3 0.0 12.0 ;Remove this atom and it will work!
[ bonds ]
1 2 6 0.0 10000
[ system ]
[ molecules ]
; Compound nmols
Add grompp check for unbound atoms
grompp now print a note for atoms that are not connected by
a potential or constraint to any other atom in the same moleculetype,
since this often means the user made a mistake.
Avoid numerical overflow with overlapping atoms
The verlet kernels did not allow overlapping atoms, even if they were
not interacting (in contrast to the group kernels). Fixed by clamping
the interaction distance so it can not become smaller than ~6e-4
in single and ~1e-18 in double, and when this number is later
multiplied by zero parameters it will not influence forces. The
clamping should never affect normal interactions; we would previously
crash for distances that were this small.
On Haswell, RF and PME kernels get 3% and 1% slower, respectively.
On CUDA, RF and PME kernels get 1% and 2% faster, respectively.
#1 Updated by Semen Yesylevskyy over 4 years ago
Such dummy particles are very useful in certain kind of simulations. The originate from complex production system where I fix DU atoms and let SW atoms to interact selectively with certain atom types in the system (a kind of wall invisible to certain atoms and visible to the others). T particles are used as anchors when such wall is moved using the com pulling.
It is possible to workaround this bug by moving T particles to separate moleculetype but such inconsistency between group and verlet schemes is not normal.
#3 Updated by Berk Hess over 4 years ago
Note that the dummies can overlap between themselves, since they are excluded from each other. The issue here is that all dummies overlap with the last hydrogen of the last water molecule. I think that for your real setup where the dummies do something useful there will be no issue.
#5 Updated by Berk Hess over 4 years ago
The actual issue in your setup is that atoms DU and T are not connected in any way. This will not result in a meaningful simulation. If you connect these by a chemical bonds (e.g. bond type 5) or exclude them manually, the system will run fine.
But the issue of non-excluded atoms not being allowed to overlap is still there in general.
It is not clear from your answer is these atoms should be able to overlap with atoms from other molecules. That would be an issue, since you can only exclude interactions within a molecule.
#6 Updated by Semen Yesylevskyy over 4 years ago
Ok, let me be more specific.
I have a set of atomtypes A1,A2... An which interact with atom type SW normally. And another set B1,B2...Bn which does not interact with SW at all. Types A and B could be in the same molecule or in different molecules. For example head of the lipid is interacting with SW while the tail is not OR lipid1 interacts but lipid2 does not. So atoms B1,B2..Bn can overlap with SW (SW is never in the same moleculetype with A or B). Something like this:
[ atomtypes ] ;name bond_type mass charge ptype sigma epsilon SW 12 0.00000 0.00000 A 0.0 0.0 DU 12 0.00000 0.00000 A 0.0 0.0 [ nonbond_params ] ; i j func SW A1 1 0.85 1e-5 SW A2 1 0.85 1e-5 SW A3 1 0.85 1e-5 ; All interactions with B1,B2,B3 are zero by default [ moleculetype ] SW 3 [ atoms ] ; nr type resnr residue atom cgnr charge mass 1 SW 1 SW SW 1 0.0 12.0 2 DU 1 SW DU 2 0.0 12.0 3 DU 1 SW T 3 0.0 12.0 ;Pull target particle [ bonds ] 1 2 6 0.0 10000 [ moleculetype ] LIP1 3 [ atoms ] 1 A1 ..... 2 A2 ..... 3 A3 ..... [ moleculetype ] LIP2 3 [ atoms ] 1 B1 ..... 2 B2 ..... 3 B3 ..... [ moleculetype ] LIP3 3 [ atoms ] 1 A1 ..... ; Interacts! 2 B2 ..... ; Do not interact! 3 B3 .....
#8 Updated by Berk Hess over 4 years ago
But the tail of the lipid would never come close to Sw, will it? So theoretically there is an issue, but in practice not.
Note that the actual nan/inf issue is still due to atoms DU and T not being excluded from each other in moleculetype SW.
PS I just pushed up a change to gerrit that checks if the energies are finite. This should catch most cases and give the user a hint of what is causing the issue. However, it doesn't allow non-interacting atoms to overlap.
#10 Updated by Berk Hess over 4 years ago
Ah, now I understand. (but you still need to add the missing exclusions/bond)
Maybe you could use the wall potential option for that, instead of using atoms?
With one non-interacting atom the chance that something goes wrong is around (REAL_MAX^(-1/12))^3 * number-density of the other atoms. In your case this would result in something around 10^-8. So you should be able to run for a reasonable amount of time.
#11 Updated by Semen Yesylevskyy over 4 years ago
No, unfortunately we can't use walls since our semi-transparent wall is of complex shape and we want to change its form during the simulation, which means that there is no alternative to atoms. This setup works nicely for group cutoffs and it would be very unfortunate if switch to Verlet will kill this possibility forever. We are using around 2000 dummies, so it will crash for sure.
I'm not a big expert in low-level Gromacs internals, so it is hard for me to understand why it works with group kernels but fails for Verlet.
Anyway, is it possible to introduce some kind of special particles for such semi-transparent walls which can safely overlap with normal atoms? Even if they will slow down simulations due to additional checks for overlaps, non-optimal kernels, etc. this is still much better then having nothing.
#12 Updated by Semen Yesylevskyy over 4 years ago
Btw, if the atoms have no charge how close they can be located without the crash in current implementation? May be it is possible to workaround the problem by defining very weak LJ repultion between them?
Another question. Does it help to define an energy group exclusions between SW and B1,B2... particles?
#15 Updated by Semen Yesylevskyy over 4 years ago
Well, this is not the best idea to run something, which is known to crash randomly. From the user point of view I'd prefer to run this in a stable way in single precision (probably with performance penalty).
I understand that my setup is a bit bizarre and 99% of the users will never use such black magic as semi-transparent particles, but at the same time it is really very useful.