[gmx-developers] Segfault during EM in recent git version

[Jochen Hub]

Hi,

thanks for your suggestions.

I have played now a bit with steep and cg, with and without vsites, 
constraints = all-bonds, only hbonds, or none, and different proteins in 
water (and with lincs-order/iter and emstep), so pretty much everything 
I could think of.

The problem seems to be a combination of things:

1) EM *with* vsites is sometimes unstable *without* applying any bond 
constraints (as you suggested, Berk). Quite often however, EM with 
vistes runs fine also without constraints.
2) conjugate gradient either with constraints or with vsites (or both) 
is less stable than steepest descent. That came as a surprise to me, 
since I usually had better experience with cg compared to steep.
3) Using constraints = hbonds usually more stable than constraints = 
all-bonds

Therefore, I would probably suggest the following for EM with vsites:

1) First try EM without constraints wither with cg or steep. Probably 
steep is is more stable than cg.
2) Try steep (not cg!) with constraints=hbonds or all-bonds (and 
remember to use lincs-order >= 8 and lincs-iter 4-8, as pointed out in 
the manual/mdp comments).
3) Always use flexible water (-DFLEXIBLE)

I do generally not recommend to use cg with constraints, either with 
vsites or not.

So, thanks again for the help. I think I will also write a brief wiki 
site on that, so other people benefit from that struggling...

Best,
Jochen



Berk Hess wrote:
> Hi,
>
> Steepest descents does run without too many Lincs warnings.
> I don't see why you would want to run conjugate gradients on a protein
> in water.
>   
I usually had better experience with cg compared to steep, that's why I 
typically stick to cg.
> But if cg runs without virtual sites, it should also run with virtual sites.
> Have you checked if the same system does miminize properly with cg without
> virtual sites, but with constraints?
>   
I have tried that. cg does not run with constraints without vsites, so 
the constraints seem to be the main problem with cg. But cg with vsites 
and without constrains, may also be unstable (but less unstable than 
with constraints).

Apparently, cg seems to have trouble with constraints (very often) and 
vsites (sometimes). So it is probably best to just stick to steep when 
having vsites, and then to try with or without constraints.