[gmx-users] Load balancing between PME and PP on more than 12 processors

[Warren Gallin]

I have a question about the procedure for running a parallel  
simulation on more than 12 processors using GROMACS 4.0.7.

I understand that partitioning the PME calculations and PP  
calculations improves performance, and that there is an initial  
automated "guess" at deciding how many processors to devote to PME vs.  
PP calculations.

When I ran a simulation of a system on 16 processors, GROMACS  
automatically devoted 6 processors to PMA calculations and 10 to PP  
calculations.  A 1 nsec simulation using 2 fs steps took only  
marginally less time (5 h 12 m) as runningthe same simulation on 8  
processors (6 h 45 m), without the separation of PME and PP  
calculations and the following note was in the log file:


NOTE: 35.7 % performance was lost because the PME nodes
       had more work to do than the PP nodes.
       You might want to increase the number of PME nodes
       or increase the cut-off and the grid spacing.


I reran the same simulation, this time including -npme 8 in the mdrun  
call, and the simulation completed in 4 h 19 min, with the following  
note in the log file:

NOTE: 16.8 % performance was lost because the PME nodes
       had more work to do than the PP nodes.
       You might want to increase the number of PME nodes
       or increase the cut-off and the grid spacing.


So I conclude that I need to increase cut-off and grid spacing, since  
this is recommended in the manual and in the paper describing the  
GROMACS 4 algoithm changes.

Unfortunately
a) I am unclear on which parameters in the .mdp file represent cut-off  
and grid spacing

and

b) When the manual says "For changing the electrostatics settings it  
is useful to know the accuracy of the electrostatics remains nearly  
constant when the Coulomb cut-off and the PME grid spacing are scaled  
by the same factor."
Does this mean that cut-off and grid-spacing parameters need to be  
changed by the same proportion?

I hope this is sufficiently specific a question - if not, let me knwo  
what I need to be clearer on.

Thanks,

Warren Gallin