Strange Fluctuations with NH+PR coupling in GROMACS 2018.3
I recently bought a workstation with Intel Xeon CPUs and 3 Nvidia Tesla P100 GPUs. I inteded to perform simulations on pure water. Since I also run water simulations on a CPU-HPC cluster using GROMACS 5.1.4 I wanted to use also this version on the workstation. I could, however, not compile this version on the workstation. Therefore, I decided to use the most recent GROMACS version (2018.3). It compiled fine, but test simulations produced drastically different fluctuations than the simulations using 5.1.4 (same mdp files used).
I also tried with 2018.3 (without GPU support), where I find similarly strange results, and 2016.5, where everything worked fine. The simulations I ran were
- NPT of 1000 TIP4P/2005 molecules at 1bar, 300K
- NPT of 1000 TIP4P/2005 molecules at 1bar, 240K
- Simulated annealing (- 30K/ns) of 1000 TIP4P/2005 molecules at 1 bar, starting from 240 K
(Nose-Hoover and Parinello-Rahman thermo- and barostats used.)
Attached you find simulation data of all runs showing potential energy (U), T, P, and density. At 300 K the results are similar for all versions used (T300.png). At 240 K (T240K.png) and when cooling (COOLING.png) the data obtained from 2018.3 (with and without GPU support) differ markedly from the results obtained with 5.1.4 and 2016.5.
I also performed the 2018.3 tests (including physical validation), which produced 1 error for the GPU version in the complex regression test (see 2018-GPU.txt attached) and 1 error for the CPU only version in the physical validation test ens_water_md_verlet_settle_pme_vr_pr (see 2018-CPU.txt attached).
I however, have the feeling that these fails are not responsible for the immensely large fluctuations I find in the simulations.
In the meantime I use 2016.5 for my simulations, but I still tought I'll let you know of this strange behaviour in the 2018.3 version.
In case you need additional information from my side, do not hesitate to contact me.
Department of Physical Chemistry
University of Innsbruck
Innrain 52c, A-6020 Innsbruck (Austria)
#3 Updated by Berk Hess almost 2 years ago
- Status changed from Feedback wanted to In Progress
The large fluctuations seem to be caused by using the same resonance time for Nose-Hoover and Parrinello-Rahman. We generally advise against this. Setting tau_p to 2 ps significantly reduces the fluctuations in gmx2018. You might need a larger distance to get acceptable fluctuations.
But I have not yet found out why there is a difference in the fluctuations between 2016 and 2018.
I suppose we should add a warning in grompp for setting tau_p equal for Nose-Hoover and Parrinello-Rahman (or four tau_p <= 1.5*tau_t) and advise to have tau_p at least twice as large as tau_t.