[MMTK] Segmentation fault, normal modes calculation

Konrad Hinsen hinsen at cnrs-orleans.fr
Thu Jan 11 12:00:30 CET 2007


On Jan 5, 2007, at 11:23, Lars Skjærven wrote:

> I am trying to calculate the normal modes of a large chaperon (GroEL/
> GroES complex, pdb code: 2cgt/1aon/1xck/2eu1). We know that these pdb-
> files have some missing residues, which we think may cause the
> problem. I have tried different methods, but they all return the same
> error. Namely "Segmentation fault". No other details are written to
> the screen. Running throught the gdb we found that it crashes at line
> 668 in MMTK_forcefield.c. This happens both om x86 and AMD 64 system.

The cause of the crash is a bug in MMTK, which is fixed in the public  
Mercurial archives. It concerns the handling of force constant  
matrices whose size exceeds 4 GB. Your system has 7854 atoms, the  
force constant matrix thus has 9*7854**3 elements or 8*9*7854**3 bytes.

Note however that the fix in MMTK by itself is not sufficient. Arrays  
 > 4 GB require Python 2.5 and NumPy. I have verified that with the  
latest MMTK + Python 2.5 + NumPy, the force constant matrix for your  
system is calculated. However, I cannot test if the computation of  
the normal modes succeeds, because I don't have any machine with  
sufficient memory to run the job without excessive thrashing.

My suggestion for handling your big protein is to switch to a  
different algorithm, which requires only two changes:

1) Change
	from MMTK.NormalModes import NormalModes, SubspaceNormalModes
    to
	from MMTK.NormalModes import NormalModes,  
FiniteDifferenceSubspaceNormalModes

2) Change

      modes = SubspaceNormalModes(universe, subspace)

    to

      modes = FiniteDifferenceSubspaceNormalModes(universe, subspace)


With that change, the huge force constant matrix is never calculated.  
Instead, numerical differentiation is used to calculated only the  
1734**2 force constant matrix for the selected Fourier subspace. This  
is somewhat less accurate, but uses only 22 MB for the force constant  
matrix.

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
---------------------------------------------------------------------


-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 2745 bytes
Desc: not available
Url : http://starship.python.net/pipermail/mmtk/attachments/20070111/e2c4cc5f/attachment.bin 


More information about the mmtk mailing list