[gmx-users] Protein simulation, improve GPU/CPU ratio

[Matilde Viegas]

Dear all,

I have a few questions regarding a simulation system containing a protein
in a box of water, with ions. Just like the lysozyme tutorial:

my system is a 5000 residue enzyme (75600 atoms), in a triclinic TIP3P
water box of a total 1567103 atoms (around 500000 water molecules). My
input is:

; minim.mdp - used as input into grompp to generate em.tpr
integrator      = steep         ; Algorithm (steep = steepest descent
minimization)
emtol           = 1000.0        ; Stop minimization when the maximum force
< 1000.0 kJ/mol/nm
emstep      = 0.01      ; Energy step size
nsteps          = 50000         ; Maximum number of (minimization) steps to
perform
define = -DFLEXIBLE

; Parameters describing how to find the neighbors of each atom and how to
calculate the interactions
nstlist             = 40            ; Frequency to update the neighbor list
and long range forces
cutoff-scheme   = Verlet
ns_type             = grid              ; Method to determine neighbor list
(simple, grid)
coulombtype         = Reaction-Field    ; Treatment of long range
electrostatic interactions
rcoulomb            = 1.0               ; Short-range electrostatic cut-off
rvdw                = 1.0               ; Short-range Van der Waals cut-off
pbc                     = xyz           ; Periodic Boundary Conditions
(yes/no)

Unfortunately, my GPU/CPU ratio is 16= Force evaluation time GPU/CPU:
345.675 ms/20.818 ms = 16.605

Any ideia on how I can improve the ratio?

Thank you all,
Matilde