Most likely, changing NHGSS and BLKSIZ will not be of help.
Both of them only slightly affects the memory needs.
An increase of NHGSS could primary help one to capture
correct roots during initial guess procedure.
As to BLKSIZ, smaller BLKSIZ could reduce heap use
but not the use of the Firefly's primary memory pool.
On the other hand, migrating to larger number of CPUs (cores),
is a very good idea. You'll get speedup due to larger maxv and
additional speedup due to larger number of cores. I believe
CAS(15,15) should be scalable up to at least 256 cores with
current CAS implementation.
Hope this helps.
All the best,
On Fri Apr 18 '14 7:24am, Pavlo Solntsev wrote
>I am doing SA-CASSCF. Based on preliminary calculations (DFT,CIS-NO, MCSCF) and NBO analysis i realized that i have to use 15e15o active space at least for now to check the contribution of the orbitals into the MCSCF solution. My ground state is doublet. And as i expected there is a second excited doublet and quartet very close to my ground state. Therefore, i am doing SA-CASSCF (aldet) over two doublets and one quartet. I use 16 CPUs for this job. Unfortunately, this job is to large to allocate the amount of memory withing my CPU# window. I need ~550 MW. I found that MAXV variable in $det may reduce an amount of memory by cost of calculation time. In my case only MAXV=1 provides appropriate amount of memory, but job running so slowly. There are some variables in $det that can also affect amount of memory. I believe they are NHGSS, NSTGSS, MXXPAN and BLKSIZ.
>I am interesting in three states but use 8 states with appropriate wstate(1) variable.
>Based on the description NHGSS and BLKSIZ are suitable for a modification. Can you please advise me what is the best strategy here to overcome the memory problem? I also did a runtyp=check with distci=32 and it looks like it solves the memory problem for MAXV=4. I will try to get more CPUs but I am wondering, is it worth to spend time and play with 16 CPUs and settings or the best strategy is to switch to a larger number of processors.