[gmx-users] NMA and g_nmeig segmentation fault

[Mark Abraham]

On 7/08/2012 7:07 AM, Michael Howard wrote:
> Hello All -- I'm trying to do a normal modes analysis of a fairly
> large crystal system (~20,000 atoms) in double-precision GROMACS. The
> system is xy periodic with 2 9-3 walls, and has periodic molecules. I
> first minimized the structure with L-BFGS to 1e-5 tolerance with
> switched vdW interactions and PME electrostatics (3dc geometry).
>
> When I tried to run g_nmeig on the matrix, I received the following
> error message from the shell
>
> /var/spool/pbs/mom_priv/jobs/2480011.lionxj.rcc.psu.edu.SC: line 7:
> 3205 Segmentation fault      g_nmeig_d -f wall3_nm -s wall3_nm -T 300
> -first 1 -last 20000 -of wall3_efreq -ol wall3_eigval -v wall3_eigvec
>> &nm.g_nmeig
> The last thing written to the g_nmeig output is
> Reading file wall3_nm.tpr, VERSION 4.5.5 (double precision)
> Reading file wall3_nm.tpr, VERSION 4.5.5 (double precision)
> Reading double precision matrix generated by Gromacs VERSION 4.5.5
>
> I was able to successfully run NMA using the same code on a similar
> system that was only 1/3 of the current size (reaction-field-zero
> electrostatics) and also on a system 1/3 of the size with xyz
> periodicity and PME electrostatics, and received no segmentation fault
> error. I'm a little bit baffled about the source of the segmentation
> error, because I would have expected it to crash during the Hessian
> calculation if there were a problem with allocating system memory or
> essentially singular forces. Has anyone else encountered a similar
> problem?

Diagonalizing the Hessian requires memory proportional to N^2, so your 
symptoms seem consistent with not having enough memory for the job. You 
can probably probe for this by using only part of the system, or single 
precision, more virtual memory, or a bigger machine.

Mark