NWChem
Overview
NWChem is a computational chemistry package that is designed to run on
high-performance parallel supercomputers as well as conventional workstation
clusters. It aims to be scalable both in its ability to treat large problems
efficiently, and in its usage of available parallel computing resources.
NWChem has been developed by the Molecular Sciences Software group of the
Environmental Molecular Sciences
Laboratory (EMSL) at the Pacific Northwest
National Laboratory (PNNL). Most of the implementation has been funded by
the EMSL Construction Project.
An extensive list of NWChem's capabilities and functionality
can be found by clicking on the Capabilities link on the navigation tabs on the NWChem website at URL
http://www.emsl.pnl.gov/docs/nwchem. It is highly recommended that you
review this information when evaluating the use of NWChem for your
research needs.
Release Notes
The latest NWChem release notes may be found by
clicking on the Release Notes link on the navigation tabs on the
NWChem website at URL
http://www.emsl.pnl.gov/docs/nwchem.
The release notes contain
very important information on the new features available in our
NWChem version, functional differences from earlier
releases you may have used, and other important information. It is
highly recommended you review these notes.
Citing NWChem
Our license agreement with EMSL requires that all scholarly works created using
the NWChem prominently cite the use of NWChem. The required and
proper citation information may be found by clicking on the Citation
link on the navigation tabs on the NWChem website at URL
http://www.emsl.pnl.gov/docs/nwchem/
Documentation
Documention on NWChem can be found by clicking on the
User's Manual link on the navigation tabs on the NWChem website at URL
http://www.emsl.pnl.gov/docs/nwchem.
Setup
NWChem requires a configuration file in your home directory so that
standard database files that contain the original force field information
can be located. These database files reside in a directory that is specified
in file $HOME/.nwchemrc.
You should link to the master system-wide configuration file by issuing the
following command:
ln -s /usr/global/NWChem-xc/data/default.nwchemrc $HOME/.nwchemrc
|
The will create the file .nwchemrc in your home directory that
NWChem will use to locate the database files.
Please note that you only need to issue this instruction once.
You will only need to execute it again should you delete or corrupt your
.nwchemrc file.
Usage
NWChem can be run interactively in serail and both serial and
parallel in batch mode.
Serial mode works on all systems while parallel
mode is available only on a few clusters at the moment.
lionxc and lionxb are the only parallel batch systems supported.
You should check with
the RCC group on parallel availability for availability on other systems.
Serial NWChem
To start an interactive serial NWChem job, run the following command:
/usr/global/NWChem-xc/serial/bin/nwchem input.nw
|
To run a serial batch NWChem job on
the batch clusters such as LION-XC
using the PBS queueing system, a
PBS script such as the following would be used.
#PBS -l nodes=1:ppn=1
#PBS -l walltime=0:10:00
#PBS -j oe
# change the current working directory to the directory where
# the input deck input.nw can be found
cd $PBS_O_WORKDIR
echo " "
echo "Starting job on `hostname` at `date`"
echo " "
# start serial NWChem with input deck input.nw
/usr/global/NWChem-xc/serial/bin/nwchem input.nw
echo " "
echo "Completing job on `hostname` at `date`"
echo " "
|
Additional information on PBS scripts and submitting jobs
to PBS can be found in the appropriate system's User
Guide in the User Guides section of this website.
Parallel NWChem
An MPI version of NWChem is available on clusters that support the
Infinipath MPI. Currently this includes clusters LION-XB and LION-XC.
Parallel version for non-Infinipath clusters will be made available.
Please check with beatnic for availability on non-Infinipath machines.
Since Infinipath MPI is only available on batch clusters no interactive
instructions will be provided.
Parallel NWChem does not require any special keywords be added
to the input file. Simply invoke the parallel version of NWChem to
run in parallel. No shared memory version is available but additional
processor cores can be used on a node by passing the proper arguements to
PBS. When using multiple cores per node please be sure to reference the
Performance Hints section below for important
information regarding memory usage. The following example uses 4 nodes
and 4 cores per node for a total of 16 processors.
#PBS -l nodes=4:ppn=4
#PBS -l walltime=0:10:00
#PBS -j oe
# change the current working directory to the directory where
# the input deck input.nw can be found
cd $PBS_O_WORKDIR
echo " "
echo "Starting job on `hostname` at `date`"
echo " "
# use mpirun to call parallel NWChem with input deck input.nw
/usr/global/bin /usr/global/NWChem-xc/mpi/nwchem input.nw
echo " "
echo "Completing job on `hostname` at `date`"
echo " "
|
Additional information on PBS scripts and submitting jobs
to PBS can be found in the appropriate system's User
Guide in the User Guides section of this website.
NWChem Performance Tips
Disk i/o and processor performance can affect the overall performance
of NWChem.
For most users a wise choice for the scratch keyword setting will have
the most impact on application performance. Use of the default local
temporary disk space will provide the highest i/o performance.
If your analysis requires large amounts of scratch disk space, first try to
run NWChem jobs on nodes with large local /tmp space before
switching to larger network-based scratch directories.
Make use of system memory to store integrals in-core rather then on disk
try adding this to the NWChem input file:
scf
semidirect memsize 2000000000 filesize 0
end
|
which will store 2 gwords (16 gbytes) of integrals in memory and none
on disk. You can try more by increasing that number. The job does not
crash your job if the integral cache cannot be allocated in the memory
available to NWChem, so there is no harm in setting this number to a
large value, as long as it does not roll over the maximum integer size
on your machine (shouldn't be an issue for a 64-bit machine).
This will work in a DFT calculation as well an SCF calculation.
If disk IO is very fast, the difference between
in-core and out-of-core algorithms is small. If you have fast disk,
an optimal approach would be to use both memory and disk for integral
caching, but limit the disk cache because huge files usually slow
things down. Experimentations find that past a ~2 gb disk access starts
to get slow. You opt for disk caching as well by
changing the filesize in the semidirect line to something other than
zero.
Semidirect is most useful when calculating the
integrals is expensive, ie when high angular momentum functions are
used, ie aug-cc-pVQZ. For sp basis sets like 6-31G*, the difference
with respect to direct should be small.
When running in parallel and using multiple core on the same node,
the memory setting is cumulative. For example:
memory stack 2048 mb heap 2048 mb global 1024 mb
|
will use a total of 5 GB (2 + 2 + 1) *per MPI process*. Thus 4 MPI procs i
on a node will use 4 * 5 or 20 GB. Node memory can be exhausted quite easily
if care is not taken.
For disk i/o and processor hints, feel free to contact
beatnic@cac.psu.edu
Further Information
Further information on NWChem may be found on the NWChem web site at URL
http://www.emsl.pnl.gov/docs/nwchem/.
Please send questions or suggestions about this web page to beatnic@aset.psu.edu
ASET | ITS | Penn State
|
