[gmx-users] bug with bond type 6 and workaround

[Marc Baaden]

Hi,

further to my previous email about segfaults etc on Solaris,
David van der Spoel kindly helped me to pin down the problem.
It was already corrected in the current CVS code, but that is
not yet up for production. So here goes a brief description
and a workaround.


2. mdrun segmentation fault bug (Thu Dec  5 12:09:54 GMT 2002)
==============================================================

The bug is for vanilla Gromacs 3.1.4 and LAM (at least that's for
sure).

Mdrun gives a segmentation fault / core dump / signal 9 or 11 on a
system that uses harmonic bond potentials type 6. Strangely on some
systems/with some nb of processors this still runs, on others it
crashes. Like shows eg this table:

Bug reproducability

nb procs      Lab           Oswald        David         Oswell
............. ............. ............. ............. .............
1 processor   fine          crash         -             crash
2 processors  crash         fine          -             -
3 processors  crash         fine          -             -
4 processors  crash         crash         -             -
5 processors  crash         crash         -             -
6 processors  crash         crash         -             -
7 processors  crash         -             -             -
8 processors  crash         crash         fine/crash(*) -
12 processors -             -             -             crash
16 processors -             -             -             crash

(*) first tried the CVS code with another MPI library. That runs fine,
but I reproduced the problem with vanilla 3.1.4 and LAM

Explanation and Workaround
--------------------------

(David van der Spoel) It seems to be the definition of harmonic
potentials holding the iron atom. Can you not make it into normal
bonds type 1? Apparently this is fixed in the CVS code but not easily
backported (and CVS is not for production use right now).

(My comment) Yes indeed, changing the bond type 6 which held the iron
in place to a bond type 1 "solves" the problem.



  Marc Baaden

-- 
 Dr. Marc Baaden - Laboratory of Molecular Biophysics, Oxford University
 mailto:baaden at smplinux.de  - ICQ# 11466242 -  http://www.marc-baaden.de
 FAX/Voice +49 697912 39550  -  Tel: +44 1865 275380  or  +33 609 843217