[gmx-users] Energy conservation of crystal surface (specification)

[Janne Hirvi]

Hello again!

In earlier message I said that double precision (actually same with single
precision) simulation with 4 processors terminates pretty soon and distinct
momentary increases in energy at intervals of 0.05ps can be observed until that
point. 

Now I have made more tests and noticed that actually the time interval between
these "energy steps" is connected to output control parameters - at least a
value of nstxout has effect on the results. Multiple simulations with different
values of nstxout (other parameters won't change and nstenergy=1) gave
different kind of energy curves. Energy values were equal at the beginning but
deviated just after the time step, which coordinates were written to trajectory
file. Energy always increases sharply at that moment and equilibriates to the
corresponding value until next sharp increase is observed when coordinates are
written to trajectory file next time.

This sounds pretty strange and I would like to know if there is some limitations
for parallelization when using option pbc=full or if someone has good ideas
what I might be doing wrong? That hopefully might also solve my overall energy
conservation problem with this system.

Thanks for your time,

Janne


>Hello gmx-users!
>
>Sorry that I am bothering you all the time with energy conservation issues but
>this time I have problems with a crystal surface consisting of infinite
polymer
>chains.   
>
>I think that simulation and interaction parameters employed should be
>conserving the total energy: time step of 0.5fs, PME with a slab correction
>for Coulombic interactions (f-spacing=0.1, pme-order=6, ewald-rtol=1E-8), a
>plain cut-off or switch function for LJ interactions (neighborlist updated on
>every time step for test purposes), and double precision. At least there
>hasn't been such problems with these parameters (excluding pure LJ cut-off)
>earlier with other systems. New features in this system for me are positions
>restraints which are used for some atoms (160 of 46080 atoms) to maintain
>"original crystal shape" and option pbc=full due to the infinite chains.  
>
>Could it be that position restraints or full periodic boundary conditions are
>responsible for the energy drift or should I just still try to equilibriate
>system more in NVT ensemble? I guess that this could be the problem, because
>there seems to be also some huge problems depending on number of processors
>used for the simulation. Simulation with 2 processors progresses normally
>although energy drift is observed, but with 4 processors simulation terminates
>pretty soon and huge momentary increases in energy in intervals of 0.05ps can
>be observed earlier.  
>
>Thanks for any help,
>
>Janne