[gmx-developers] Wrong results when using Verlet cut-off scheme and ewald-geometry= 3DC in version 4.6 and higher

[Sheng Bi]

Here we find gromacs goes wrong when using verlet and 3dc an run in parallel.
For slab geometry, we need to do slab correction based on total charge dipole ( mu_tot=sum(qi*zi) ). If one performs such simulation in gromacs with Verlet and 3dc in parallel (especially using domain decomposition), the mu_tot is wrong. For example, if you run simulation in 16 core, and the right global charge dipole is A, however in gmx 4.6 and higher each core would get 16*A!
I think the problem come out on this communication, for version 4.6, gmx sum charge dipole twice. In sim_util.c, line 1052:
            /* When we don't need the total dipole we sum it in global_stat */
            if (bStateChanged && NEED_MUTOT(*inputrec))
            {
                gmx_sumd(2*DIM, mu, cr);
            }
if using domain decomposition, gmx will sum the dipole for the first time. However in line 1090:
     if (bStateChanged && NEED_MUTOT(*inputrec))
    {
        if (PAR(cr))
        {
            gmx_sumd(2*DIM, mu, cr);
        }

        for (i = 0; i < 2; i++)
        {
            for (j = 0; j < DIM; j++)
            {
                fr->mu_tot[i][j] = mu[i*DIM + j];
            }
        }
    }
gmx sum the dipole for second time only if run in parallel. So, if you perform simulation in 3dc with verlet and in domain decomposition, the mu_tot will go wrong.
Since we performed a lot of slab geometry simulatin and gmx worked perfect in version 3.3 and 4.5, when we use 4.6 and higher, we find this problem. Is this a bug or my setup error in mdp？

--
PhD student
State Key Laboratory of Coal Combustion,
School of Energy and Power Engineering,
Huazhong University of Science and Technology (HUST),
Room 302 Power Building, 
1037 Luoyu Road, Wuhan, Hubei 430074 China
Email: chrishengbee at hust.edu.cn
Phone: (86)27-87548122 
http://itp.energy.hust.edu.cn 


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20160308/f0913662/attachment.html>