[gmx-users] Re: Re; Force Calculation

[Erik Lindahl]

Hi,

On Jun 29, 2006, at 4:41 PM, Justin M. Shorb wrote:

> Erik and Mark: Thanks for your help, I will begin diving into the  
> code today.
>
> Mark, if everything went ideally (which things don't always), I  
> would use the force tables to compute the total SR electrostatic  
> force on each atom in a specific group (around 2 to 7 atoms). The  
> same goal can be accomplished by running a trajectory and then  
> rerunning it to calculate the electrostatic forces exclusively.  
> This is what we have done before, but saving the necessary  
> trajectory file for our current system has gotten too costly.  
> Instead, I am hoping to collect the forces and save those  
> exclusively at each timestep (along with positions at larger  
> intervals). So far it has seemed that Gromacs creates a separate  
> coulombic and LJ table, but uses a combined table to actually  
> compute the forces. So it should be capable of computing the simple  
> ES force on the fly and saving them. Any suggestions on a better  
> tactic?

The separate/combined table is an optimization to improve cache  
efficiency. The memory latency can be pretty significant for table  
lookups (and you do billions of them in the kernels...), so by  
interleaving coulomb and LJ tables the LJ part will already be in  
level1 cache after doing the coulomb table interaction.

So, for all atoms that have both tabulated coulomb and tabulated LJ  
interactions we use the combined table, otherwise separate (since  
those are smaller).

Cheers,

Erik