[gmx-users] Query regarding performance varies hugely in two similar systems

[Mousumi Bhattacharyya]

Dear Gromacs users,

I am trying to study a self-assembly process of a polymer with three different composition ratio using Gromacs in CPU cluster (Intel(R) Xeon(R)). As the ratio of the polymer constituents are different, the number of atoms differs in each case (# of atoms ~ 170000,  190000, and 217000). Each system contains 40 chains of each polymer and N number of water molecules.

Due to constant failure and blowing up of the systems in NPT equilibration and further production run using Parrinello-Rahman pressure coupling I thought of using Berendsen for the last 2 systems. Using Berendsen pressure coupling there is no failure until now (4ns). However the performance has become very poor. Using 16 MPI processes and 8 OpenMP threads per MPI process the last 2 systems are achieving only 0.4-0.6ns/day.

On the other hand the first one can achieve nearly 4ns/day.  The differences in .mdp file of system 1 from the other two are that during nvt and npt I have used cutoff-scheme   = Verlet and in NPT and for  further production run the pressure coupling as Parrinello-Rahman. I have attached all the .mdp files here. The scheme I have used are:  1) Position restr NVT 2) position restr NPT 3) free NPT 4) production run. I agree that systems 2nd and 3rd would definitely take more time than the 1st one due to the difference of no. of atoms but should not this much. I am in doubt now whether I have messed up with my mdp files to get rid of the blowup from NPT simulations.

I would be highly grateful if you could look into the files and help me find out the problems so that I can get the performance back. Please let me know if I need to provide more information.


Thanks in advance.
Mousumi