[gmx-developers] Re: Re: Re: Energy conservation

[Erik Lindahl]

Hi Michael,

On Apr 10, 2006, at 5:41 PM, Michael Shirts wrote:
>
> I'd love to get the .mdp and .gro for the system to test on my
> velocity verlet + Settle + Rattle integrator to see if I can get
> comparable performance.   Why do you think it was necessary to comment
> out SETTLE?

I'll see if I can find it, but don't hope too much since it was just  
a simple "playing" system with water - I might not have saved it.

Anyway, the important parts are:

1. Use buffers in the neighborlist to account both for diffusion and  
finite charge group size.
2. Use switch, or PME with a significantly lower value of pme_rtol.  
I'm pretty sure the reciprocal part will conserve energy anyhow, so  
you don't need to worry about accuracy due to lower pme_rtol.  
Alternatively you can use your own tabulated function that is a  
switched-off ewald interaction.
3. Equilibrate well

As I mentioned earlier today SETTLE conserves energy just fine - this  
was just a hack to be able to test LINCS using a tiny, well  
equilibrated, system instead of equilibrating a huge protein or  
something.

>
> Having the data + standards for these examples in the CVS (or
> otherwise posted) would allow other people.  I know there some
> regression testing that is done -- is that information (not only the
> input files, but the results) in the CVS?  Are the example files in
> the CVS the same as the ones used in the regression testing?  I'd also
> be hearing from Charlie Peck about how the examples could be better
> used from a CS / Quality control perspective to be able to better
> ensure a good development process.

Check out the "test" module - there is some scattered stuff there.  
The primary regression tests are the benchmarks, our research runs of  
the day, and the kernel regression tests.

Cheers,

Erik