Speeding up TDDFT calculations?

[Jenni]

I'm running some singlet-triplet energy gap calculations with NWChem on a supercomputer but finding some of the calculations for bigger molecules are taking too long to complete and the job is timing out. I'm using the B3LYP functional and cc-pVDZ basis set. So far increasing the nodes or the cpus per task does not seem to speed up the job. Is there any way of speeding up TDDFT calculations with NWChem? 

[jeff.science]

Please attach the input file and post the environment variables used for your build configuration.

Jeff

[Jenni]

I have also attached here the .out file. It seems to get stuck at the Davidson iterations for the triplet state calculations, though it has worked with other functional and basis set combinations (PBE0/6-31G(d) and B3LYP/6-31G(d). 

[Jeff Hammond]

There are a few issues here.

First, we don't know how NWChem was configured, particularly w.r.t. ARMCI, which makes a huge difference in performance in some cases.  Since you are running with one MPI process, I know you're not using MPI-PR.  Most likely, you are getting the default of MPI-TS, which is usually terrible.

Second, you are running with one MPI process on a Cray machine, which is not a good use of the computer (I assume it has 64 or 128 cores per node).  I'm running on a 16-core desktop using 16 MPI processes and ARMCI-MPI, which reduces the run time of the ground-state DFT calculation by a factor of ~19 and the TDDFT iterations by a factor of ~11.

For both of the above reasons, please use NWChem directly, without the BigDFT stuff that doesn't know how to run it properly.  Also, please send me the results of "module show nwchem" in case I can see how it was built from that.

When you compile NWChem, please use ARMCI_NETWORK=MPI-PR or ARMCI_NETWORK=ARMCI, the latter after you run $NWCHEM_TOP/src/tools/install-armci-mpi and follow the directions provided by its output.  The former is easier, but doesn't let you run sequentially, which I find annoying.  There are other differences that are interesting to non-chemists.

Third, you are using the default of cartesian d functions with the Dunning basis sets, which were designed to be used with spherical d-functions.  This has a very small effect on the runtime and a large effect on the ground-state energy, but which mostly disappears with energy differences.  In any case, it's the correct way to use cc-pVDZ.  Note that Pople basis sets were designed to be used with cartesian d-functions, so you should not use "basis spherical" in those input files.

cc-pVDZ may converge more slowly because it is a larger basis set or perhaps because the use of cartesian d-functions increases the linear dependence of the basis set (the extra d-function manifests as a diffuse s-function, IIRC).

I will send you the log file when this run is done.

Jeff

--
You received this message because you are subscribed to a topic in the Google Groups "NWChem Forum" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/nwchem-forum/cQ_YQtzh_s0/unsubscribe.
To unsubscribe from this group and all its topics, send an email to nwchem-forum...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/nwchem-forum/ee7a4095-bacd-477e-a0d0-085657d1cb83n%40googlegroups.com.

--

Jeff Hammond
jeff.s...@gmail.com
http://jeffhammond.github.io/