CHPC Software: Gaussian 09
Gaussian 09 is the newest version of the Gaussian quantum chemistry package, replacing Gaussian 03.
- Current revision: D.01.
- Machines: ember, telluride, kingspeak
- Location of latest revision:
- telluride: /uufs/chpc.utah.edu/sys/pkg/gaussian09/EM64TL
- ember and kingspeak: /uufs/chpc.utah.edu/sys/pkg/gaussian09/EM64T
Please direct questions regarding Gaussian to the Gaussian developers. The home page for Gaussian is http://www.gaussian.com/. A user's guide and a programmer's reference manual are available from Gaussian and the user's guide is also available online at the Gaussian web site.
IMPORTANT NOTE: The licensing agreement with Gaussian allows for the use of this program ONLY for academic research purposes and only for research done in association with the University of Utah. NO commercial development or application in software being developed for commercial release is permitted. NO use of this program to compare the performance of Gaussian 09 with their competitors' products (i.e. Q-Chem, Schrodinger, etc) is allowed. The source code cannot be used or accessed by any individual involved in the development of computational algorithms that may compete with those of Gaussian Inc. If you have any questions concerning this, please contact Anita M. Orendt at firstname.lastname@example.org.
The general rule is that SCF/DFT/MP2 energies and optimizations as well as SCF/DFT frequencies scale well when run on multiple nodes. In the case of DFT jobs, the issue with larger cases (more than 60 unique, non-symmetry related atoms) not running parallel has been resolved. You no longer need to use Int=FMMNAtoms=n to turn off the use of the FMM, where n is the number of atoms in order to run these cases in a parallel fashion.
To run non-restartable serial jobs that will take longer than the queue time limits, you can request access to the long queue on a case by case basis in order to be able to complete these jobs.
As always, when running a new type of job, it is best that you first test the parallel nature of the runs by running a test job on both 1 and 2 nodes and looking for the timing differences before submitting a job that uses 4, 8 or even more nodes. The timing differences are easiest to see if you use the p printing option (replace the # at the beginning of the keyword line in the input file with a #p)
Please note that it is important that you use scratch space for file storage during the job. The files created by the program are very large.
When determining a value for the %mem variable, please allow at least 64 MB of the total available memory on the nodes you will be using for the operating system. Otherwise, your job will have problems, possibly die, and in some cases cause the node to go down. Sanddunearch nodes have 8GB, Updraft 16GB, Ember 24GB, and Kingspeak 64GB.
There are two levels of parallelization in Gaussian: shared memory and distributed. As all of our compute nodes have multiple cores per node, you will ALWAYS set %nprocs to the number of cores per node. This number should agree with the ppn value set in the PBS -l line in the batch script. The method to specify the use of multiple nodes has changed in G09. The %nprocl parameter is no longer used. Instead, in the PBS script a file containing the list of nodes in the proper format called Default.Route is created. This change has also caused a limitation in the number of nodes that can be used (at least for now). This limit is based on a line length limit, and is between 8-12 nodes, depending on the length of the cluster name.
In order for to successfully run G09 or GV jobs interactively you need to to be using the gaussian09 environmental variables as set in the .tcshrc script provided by CHPC. Therefore you must have this file in your home directory. You will need this even if your normal shell is bash, as Gaussian requires a csh to run.
The chpc.tcshrc is available at: chpc.tcshrc
An example PBS script may be obtained here: g09.script
Please read the comments in the script to learn the settings and make sure settings of these variables agree with settings in the % section of the com file and in the PBS directives in the PBS script.
Example of parallel scaling using G09
These timings can be compared with those given on the G03 page
The job is a DFT (B3PW91) opt job with 650 basis functions taking 8 optimization cycles to converge.
Number of Nodes
Number of Processors
|walltime to complete job|
The DFT frequency of the above case:
Number of Nodes
Complete wall time
|2||< 1.0 hrs|