[Dock-fans] Dock 6.2 tests and parallelization problems

Scott Brozell sbrozell at scripps.edu
Tue Apr 15 10:06:16 PDT 2008


Hi,

On Sat, 12 Apr 2008, Dmitry Osolodkin wrote:

> Could help me to solve a few problems with the installation and 
> parallelization of DOCK 6.2 and 6.1? The system is a cluster of AMD 
> Opteron processors, RHEL3, gss 3.2.3.
> 
> 1. When I compile the serial version of DOCK6.2 with gcc 3.2.3 and 
> default configuration, it is compiled successfully. After running 'make 
> test', there is a great many of failures in almost all tests, rather 
> often concerning different number of conformations obtained or notable 
> differences in the atomic coordinates.

This is not unusual.  The underlying problems are numerical differences
compounded by grid creation and sampling itself in dock6.
We are working on improving the test suite.
However, our current standard test platform is a pentium4 cluster
Red Hat Enterprise Linux AS release 3 (Taroon Update 6)
gcc version 3.2.3 20030502 (Red Hat Linux 3.2.3-53)
I would not expect that pentium4 to opteron would break everything.
How did the continuous score and the amber score tests fare ?

> 2. When I compile the parallel version of DOCK6.2 with gcc 3.2.3 and 
> default configuration, it is compiled successfully. When I try to run 
> the parallel computation with 'mpirun -np _any_ -nolocal -machinefile 
> machine', it runs at only one processor. The same behavior is observed 
> without the 'nolocal' flag.
> Moreover, the computation can't be run from the desired directory, but 
> only from my home directory: it creates '~/dock.in' with default options 
> and tries to read all other files from '~/' even if all paths are 
> pointed explicitly.

This reads like a problem with your MPI installation.
Have you ran the installation tests of your MPI ?
Have you verified access to remote nodes ?
Are you doing everything exactly the same in comparison to your
dock 6.1 parallel ?  There are no parallel related DOCK
changes between 6.1 and 6.2.

Send verbatim and verbose details if these questions do not
lead to a fix.

> 3. Parallel version of DOCK 6.1 (all bugfixes applied) virtual screening 
> crashes with the SIGSEGV after about 300 of 2000 compounds screened. The 
> number of compounds depends on the input ligand database, but is 
> reprodicible for the different runs with the same database.

Presumably, the same calculation ran serially succeeds ?

We gratefully accept detailed bug reports that contain short input
files for reproducible bugs.  In fact those are our favorite ones :)
http://www.chiark.greenend.org.uk/~sgtatham/bugs.html

Scott




More information about the Dock-fans mailing list