well, there are myriads of testcases when fullnr converges
faster while soscf either converges slower or does not converge
at all. However, the exactly opposite statement is true as well.
The main problems with fullnr are:
1. It is slow by design.
2. It requires huge amount of memory for large systems.
3. It tends to converge even for badly chosen active spaces resulting in senseless answers.
The soscf is "clever" as it most likely will not converge at all.
With fullnr, you always need to double check the orbitals to be
sure the solution is what you really wanted to get.
On the other hand, soscf readily converges for reasonable starting
orbitals and active spaces; and is much faster and efficient than fullnr.
E.g., look at the fullnr-converged orbitals in the testcase you
have provided. As to me, at least one of orbitals looks suspiciously...
is that exactly the solution and active space that was actually intended to get?
As to your question on MCQDPT or XMCQDPT, you can run them using
multiple cores of multi-core system. In fact, this is one of our
standard benchmarks - you can find more information in the "Performance" section.
Look at http://classic.chem.msu.su/gran/gamess/smp.html for documentation on SMP
and multi-core support. In brief, you need to run a single process
and just properly set the mklnp variable.
On Tue Mar 9 '10 1:11pm, Alex Maiorov wrote
>Thank you very much, with mcscf it works good in parallel
>and quickly converges.
>But i am anxious about you consider FULLNR MCSCF code as
>obsolete and going to remove it in the future.
>I have an example of structure, whitch could be resolved
>with SOSCF and FOCAS without FULLNR method. I tried different
>initial vectors, but MCSCF is not converged by the means of SOSCF and FOCAS,
>and MCSCF converged quckly with FULLNR. The example is in
>the attachment. Am i right? And, by the way, can MRMP2 energy
>run in parallel?
[ This message was edited on Sun Mar 14 '10 at 0:25am by the author ]