working with Tinker8 (CPU, 25.5.3), but not Tinker9 (GPU, 1.4.0)

[qinghualiao]

Hello,

I have Tinker 25.5.3 and Tinker-GPU installed on the cluster.

I set up one system of a metal ion model in water (a cubic box 40 angstrom in one dimension).
I found that I can simulate it with CPUs without problem, however, I could not simulate it with the GPU version.
(The GPU version works well with a pure water system)

I also tried to used a well equilibrated system with CPUs as the starting point for the GPU run, but it still did not work.

I then used analyze9 to analyze the system:
analyze9 md.xyz -k md.key
(md.dyn is in the same folder, a dyn file from a run with CPU)

Total Potential Energy : -26071.5803 Kcal/mole
Energy Component Breakdown : Kcal/mole Interactions
Bond Stretching 304.3641 4176
Angle Bending 396.9066 2100
Urey-Bradley -6.0692 2085
Van der Waals 13517.2155 2210423
Atomic Multipoles -27659.7965 436884
Polarization -12624.2007 436884

Total Electric Charge : -0.00000 Electrons
Dipole Moment Magnitude : 522.564 Debye
Dipole X,Y,Z-Components : 436.227 197.585 -209.141
Quadrupole Moment Tensor : 1020.781 -2131.926 -153.053
(Buckinghams) -2131.926 -909.583 -4719.338
-153.053 -4719.338 -111.198
Principal Axes Quadrupole : -5676.135 936.053 4740.081
Radius of Gyration : 19.964 Angstroms
Center of Mass Coordinates : -0.008805 0.128612 0.114992
Euler Angles (Phi/Theta/Psi) : 71.144 -16.654 -120.139
Moments of Inertia and Principal Axes :
Moments (amu Ang^2) X-, Y- and Z-Components of Axes
9933448.950 0.309634 0.555260 -0.771889
9976678.509 0.906639 0.072266 0.415672
10075158.532 0.286588 -0.828531 -0.481045

Internal Virial Tensor : 2261.581 217.633 -138.696
217.633 2350.198 -148.794
-138.696 -148.794 2075.553
Pressure (Temp 298 K) : 1586.017 Atmospheres
Pressure From Virial -2388.228 Atmospheres

All the items above seem to be OK. However, when I start to run the MD with:
dynamic9 md.xyz -k md.key
(the same key file was used with the CPU version)

The simulation crashed immediately, and it shows that the potential energy goes extremely high, unlike the one reported by analyze9.
13 O 6581886.988884 349 14 15
14 H ****************************************** 350 13
15 H ****************************************** 350 13
16 O 9385356.813412 349 17 18
17 H ****************************************** 350 16
18 H ****************************************** 350 16
19 O ****************************************** 349 20 21
20 H ****************************************** 350 19
21 H 802531.078907 350 19

Did I miss something when going from the CPU version to the GPU version? Any suggestions to make the GPU version working?
I appreciate for your comments and help!

All the best,
Qinghua

[jayponder]

Clearly the GPU simulation is "exploding". This is very difficult to debug without actually seeing your files. Could you post the various input and output files from both CPU and GPU (coordinates, keyfile, parameters, "dyn" file, etc)? Either put them here or email them to me (ponder@dasher.wustl.edu).

Another thing you might try is to remove the CPU dyn file, then minimize the ion+water box. Use the "minimize" program on CPU and/or the "minimize9" program on GPU, and only run the minimization to an input convergence of 1.0. Then try running MD on the GPU from that coordinate file without any dyn file being present. That may also fail, but is an easy check to try. Actually, I suspect the issue may be with something in your parameter file that is not being handled the same on the CPU and GPU. But there are also other possibilities.

[jayponder]

The problem is that you have set the polarizability values of your dummy particles attached to the zinc to exactly zero. This works in the CPU code as the multipole and polarization potentials are computed completely separately. In the GPU code, for efficiency reasons, the multipole and polarization energy/force codes are combined into a single routine. This is then causing an issue (in the force, but not the energy) when a site has nonzero multipoles but no (or zero) polarizability.

We can likely patch this without losing efficiency in a future release of the GPU code. But in the meantime, you should just set the polarizability of your dummy particles to some very small nonzero value. For example, replace the following lines in your modified amoebabio18.prm parameter file:

polarize 367 0.0000 0.0000 366
polarize 368 0.0000 0.0000 366
polarize 369 0.0000 0.0000 366

with something like the following:

polarize 367 0.00000001 0.0000 366
polarize 368 0.00000001 0.0000 366
polarize 369 0.00000001 0.0000 366

Those polarizability values are so small as to have no effect on the simulation results, but this change will allow the current GPU code to run correctly.

[Qinghua-Liao]

Dear Prof. Ponder,

Thanks so much for your detailed explanation!
After following your suggestion, it is working well with GPU now!

All the best,
Qinghua