[gmx-developers] Suggestion: move bonded forces in mdrun

[Berk Hess]

Carsten Kutzner wrote:

>Hi there,
>
>I wonder if it is possible to move the calculation of the bonded forces
>(calc_bonds) before the do_pme routine in mdrun. This way, one could
>achieve slightly better scaling when using PME in parallel runs, in
>particular for systems that consist of a single large protein in water.
>
>In the current code, where the bondeds are evaluated after PME, a single
>CPU calculates the bonded forces for the big protein while the
>others are already finished and waiting. If the bondeds were before PME,
>one could give this CPU somewhat less atoms via the -load "..."
>option in grompp, so that all CPUs enter the PME routine at the same
>time. This would also enhance the potential speed gain for future dynamic
>load balancing.
>  
>
Do you mean that one could compensate for the larger bonded interaction load
by doing less non-bonded interactions on such a cpu?

>As I see it, this would only require a slight change in the force.c
>routine.
>  
>
I thought that there might be a problem with shifts.
But after looking in the code I was that also PME uses whole molecules,
so it is very easy to change indeed.

Berk.