[molpro-user] Large System MCSCF/MRCIC Calculations in Parallel

Jason Goodpaster goodpast at caltech.edu
Fri Mar 1 00:48:17 GMT 2013

Dear Molpro Users,

I am trying to run MCSCF/MRCIC calculations on large transition metal
complexes where the active spaces are not particularly large, for instance
8 electron in 10 orbitals, but there are a large number of closed orbitals
using Molpro2012.1.  I am using a large basis set, a re-contraction of the
ANO-RCC basis, and the two-electron integral take up about 30-50 GB of disk
and has 600-800 basis functions.  I have two questions:

(1) How parallel is the MCSCF code and how is it parallelized?  When I run
the code in serial verse any kind of a parallelization, the MCSCF program
appears to take approximately (within 20%) the same time. Additionally,
when I run the code across nodes, I overrun the InfiniBand adapter on the
nodes and cause communication errors.  Effectively requiring the use a
single node.  When I print the CPU timings it appears as the majority of
the time is spent in the integral transformation.  Is the integral
transformation not parallel? Does the integral transformation require
significant node communication?  Am I able to over come this by
specifications/implementation of the scratch directory?

(2) Does anyone has any suggestions on parallelization/efficiency for MCSCF
calculations?  I haven't tested running in integral direct mode
significantly, but my first tests didn't show a significant difference in
timings.  It seems like I must be making a mistake somewhere since I can
run (in parallel) HF/CCSD(T) calculation using a similar size basis about
twice as fast as the MCSCF calculation alone (in serial or parallel), but
perhaps not since the CCSD(T) is highly parallelized.   Any input would be

Thanks in advance,
Jason Goodpaster

 A sample input would be:




-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.molpro.net/pipermail/molpro-user/attachments/20130228/cd3a63bd/attachment.html>

More information about the Molpro-user mailing list