[gmx-users] single vs double precision issues

[Adrien Leygue]

Dear Gromacs users,

I am experiencing several problems which all seem to relate to single
vs. double precision.

1-) I am doing simulations of a liquid crystal system using a custom
force field but whenever I try to run them in parallel they all crash
after some time (238 ps) unless I run them in single precision. The
error message tend to vary but it is usually something like:

in md0.log:
...
Inner product between old and new vector <= 0.0!
constraint #20 atoms 1310 and 1311
    i     mi      j     mj      before       after   should be
 1297  15.04   1298  14.03     0.15300     0.15302     0.15300
...
...
Step 119954  Warning: Pressure scaling more than 1%.
Inner product between old and new vector <= 0.0!
constraint #5 atoms 492 and 493
...
...
Constraint error in algorithm Shake at step 119954
Wrote pdb files with previous and current coordinates
Grid: 69919375 x 69919375 x 69919375 cells
-------------------------------------------------------
Program mdrun_mpi, VERSION 3.3.1
Source code file: nsgrid.c, line: 220

Fatal error:
Number of grid cells is zero. Probably the system and box collapsed.

-------------------------------------------------------

If I run the very same simulation in single precision it does not
crash. I have tried compiling Gromacs with different options and used
different MPI implementations (mpich and openmpi)

2-) I am also doing simulations of coarse grained molecules using
virtual sites and home-made tabulated potentials (using several
table_X_Y.xvg files). I haven't tried parallel runs. These simulations
run prefectly in double precision but whenerver I try to go to single
precision they crash almost immediately (prior energy minimization is
performed in double precision). I have tried to re-run some energy
minimization on the system in single precision but this stops
immediately as well. Looking at the output I see huge forces and
potential energy which are not present in double precision. However if
I remove all non-bonded interactions (I do not have charges) in the
topology file the energy minimization works fine in single precision.

I can however simulate a LJ fluid using any of the above tabulated
potential. I assume there must be some coupling between constraints,
virtual sites, and tabulated potentials but then why does it work in
single precision.

I welcome any help or suggestion on how to investigate these issues. I
haven't attached any file (mdp, topology,...) but I can send them.

Thanks a lot for your help.


Adrien.

-- 
Adrien Leygue

Department of Materials Science and Engineering
School of Chemical Engineering
National Technical University of Athens
9 Heroon Polytechniou Street
Zografou Campus, Athens 157 80,
GREECE
Tel/Fax: +30210 772-3112