[MMTK] Normal Modes problem

khinsen at cea.fr khinsen at cea.fr
Wed Jun 1 12:04:11 CEST 2005


On May 31, 2005, at 18:08, Vasisht R. Tadigotla wrote:

> I'm trying to calculate the normal modes for yeast RNA polymerase II 
> (PDB
> code: 1NIK) using the deformation force field. I'm using the
> deformation_modes.py file in the examples directory. The protein has 
> ~3300
> c-alpha atoms.
>
> After applying the cutoff, it tries to calculate 780 low frequency 
> normal
> modes but the calculation doesn't finish even after 15 hrs. I'm using
> MMTK-2.5.7 on Fedora Core 2. The machine is a P4 2Ghz with 512MB RAM.

Check what "top" says. I suspect your machine is swapping like crazy. 
On my 512 MB machine, I can't even run that job, it crashes with a 
"memory error" immediately. Yours may have more swap space, or a 
different configuration in other respects.

> Am I correct in assuming that the largest stored matrix in this case 
> would
> be 780x3x3300 ?

No. The largest array allocated will be (3*3000)^2.

> Here I think it's allocating 3300x3 X 3300x3 size array, though I'm 
> doing
> a reduced normal modes calculation. Is there an upper limit on the 
> size of
> the protein I can do approximate normal modes on?
>
The example script that you use is optimized for speed rather than for 
memory size. I calculates the full force-constant matrix and then 
projects onto the subspace basis. The subspace is introduced only to 
reduce the size of the eigenvalue problem, and hence CPU time.

For really big systems such as yours, there are two options:

1) Use numerical differentiation for the force constants. That's a bit 
less precise, but the derivatives are calculated only for the subspace 
basis vectors, eliminating the biggest array.

2) Use a sparse force constant matrix.  That's a bit slower than using 
the full one, and not supported by all force field terms, but it does 
work for the deformation force field (as well as its successor, the 
CalphaForceField) and it should be quite efficient for those 
short-ranged interactions.

My recommendation is to try 2) first, and if that still uses too much 
memory, use 1).

Konrad.
--
---------------------------------------------------------------------
Konrad Hinsen
Laboratoire Léon Brillouin, CEA Saclay,
91191 Gif-sur-Yvette Cedex, France
Tel.: +33-1 69 08 79 25
Fax: +33-1 69 08 82 61
E-Mail: khinsen at cea.fr
---------------------------------------------------------------------




More information about the mmtk mailing list