[gmx-users] More Athlon talk (or not?)

[Lynne Bilston]

>
>[...]
> > I haven't had much time to test single cpu jobs, so I can't answer this
> > one other than to say I haven't had a problem with them. I also have ECC
> > memory. However, I do still have occasional crashes in gromacs which
> > show up as settle errors in the logs. There are no odd things happening
> > as far as I can tell from the simulation data. If I restart from just
> > before the crash, they don't occur again. This is often several
> > nanoseconds into a run. If I look at the coordinate files from the step
> > before the crash, some of the bits of the protein are missing. I'm not
> > sure if any of this is helpful or relevant.
>
>Just a note to avoid confusion. When you restart a run, even using the
>.trr file and tpbconv, your continuation will not be *completely* identical
>to the original, the difference lies in the effects of temperature coupling.
>And even a very slight change will cause your simulation to diverge rapidly.
>So it is not surprising that you do not reproduce the settle errors in your
>restarted run, and you cannot conclude from this that you must have hardware
>problems.

I basically just shrug my shoulders and restart the simulation. I'm not 
sure if it has anything to do with the other Athlon-specific problems, as I 
said in my original post.

>Do the missing bits of the protein have 'nan' coordinates (check the
>.gro/.pdb file with a simple text viewer/editor)?
In one simulation (which is handy) the previous step coordinates have no 
nan's although one bit of the protein is in a weird place (separated), but 
the current (crashed) step coordinates have some waters with nan 
coordinates. The "missing bits" may be a viewer thing, I guess.

Any ideas about what is really causing the problem with settle and how to 
avoid it? The pressures and so on seem normal.

Cheers

Lynne