[Neolith-users] Gaussian 03 revision E.01 on Neolith

[Torben Rasmussen]

Hi Neolith users,

We have installed revision E.01 of Gaussian 03 and this revision is  
now the default on Neolith.

The default Gaussian version is loaded with the command 'module load  
gaussian' or a line in your .modules file with the word gaussian.

You can review the release notes for revision E.01 here: http://www.gaussian.com/g_tech/g03_rel.htm
A new feature in revision E.01 is that each linda worker can use up to  
eight shared memory threads, however this is not the default behavior  
on Neolith. The default is to use up to four shared memory threads per  
linda worker and start two linda workers on each allocated node, if  
%NProcLinda is specified in the input. This is also the default for  
revision D.01. (Note that four shared memory threads is also the  
default for jobs that don't use linda, so remember to include  
%NProcShared=8 in the input for such jobs). We keep this combination  
of linda workers and shared memory threads as default because it gives  
the best performance for our test job. Note, however, that we have  
only performed extensive testing with this one job, a B3LYP geometry  
optimization, so you may find better performance with other  
combinations for other types of calculations. See below, how you can  
control the number of linda workers that is started on each node.

If you want to run e.g. four linda workers per node, you can add the  
following line to your sbatch script:
G_NEO_LW=4

NOTE that this line has to come BEFORE the source $g03root/g03/bsd/ 
g03.profile line otherwise the default value of 2 will be used. You,  
of course, also have to specify meaningful values for %NProcLinda and  
%NProcShared in your input. For example, if you want to run on two  
nodes using four linda workers on each node, you should use:
# gaussian input
%NProcLinda=8
%NProcShared=2

and

# sbatch script
#SBATCH -N 2
G_NEO_LW=4


For MP2 jobs using linda, each linda worker requires its own scratch  
space, so for large MP2 calculations you may want to run one linda  
worker per node so that this linda worker can use all the available  
scratch space ~190 GB. Remember to set MaxDisk in your input for such  
jobs, because the default is MaxDisk=90GB. Also, remember to include  
%NProcShared=8 in your input.

Note that you do not have to set G_NEO_LW if you want to run one or  
two linda workers per node. If you allocate two nodes (SBATCH -N 2)  
and specify %NProcLinda=2 in your gaussian input, gaussian will start  
one linda worker per node. It is only if you want to run more than two  
linda workers per node that you need to use G_NEO_LW.

Kind regards,
Torben
NSC support