[MMTK] nMoldyn timing?

[Konrad Hinsen]

On Nov 3, 2008, at 17:24, Jose Borreguero wrote:

> Does anyone have a rough estimate of nMoldyn computing time to  
> calculate the incoherent intermediate structure factor? I am using  
> the default settings:

I don't actually use nMoldyn much, so I will refrain from quoting any  
numbers. Just some general remarks:

- The incoherent scattering function and the associated structure  
factor are the quantities that take most time in nMoldyn.

- The time is pretty much proportional to the number of Q vectors  
that have to be treated (in your case that's 2500 (50 shells of 50 Q  
vectors)) and to the number of atoms.

- The time also grows with the length of the trajectory, though not  
linearly. The CPU time should be proportional to N log(N), where N is  
the number of time steps.

- I/O time can be important. nMoldyn reads the trajectory atom by  
atom, which is less efficient than time step by time step. You can  
improve the trajectory reading time by using a trajectory with a  
block_size parameter > 1 (try 100 or 1000, for example).

If you want to get an estimate for the time the computation will take  
on your machine, I suggest doing a test run for few atoms and few Q  
vectors first. From there it is safe to extrapolate linearly. Also  
check the percentage of CPU time that your job takes. If it is much  
less than 100% on an otherwise empty machine, then your bottleneck is  
I/O, not computation.

Konrad.
--
---------------------------------------------------------------------
Konrad Hinsen
Centre de Biophysique Moléculaire, CNRS Orléans
Synchrotron Soleil - Division Expériences
Saint Aubin - BP 48
91192 Gif sur Yvette Cedex, France
Tel. +33-1 69 35 97 15
E-Mail: hinsen at cnrs-orleans.fr
Web: http://dirac.cnrs-orleans.fr/~hinsen/
---------------------------------------------------------------------