Re: Scratch space requirement for Gauss86
I've just subscribed to this list/newsgroup.
I'm a computer programmer who installed the NEC version of Gaussian86
on the NEC SX-1A Supercomputer. I'm now 'supporting' the software, but
I know next to nothing about computational chemistry.
The problem: G86 on the SX uses a tremendous amount of scratch filespace
for the INTEGRAL and RWF (readwrite) files. Both these
files are 'direct access' files whose sizes need to be
'allocated' before execution of the job. And there is no
'file expansion' facility. Most of the time, users just
allocate a large scratch space without really knowing how
much they actually need. If the space is insufficient,
the program aborts (after much wasted cpu hrs).
Question: Is there a way to calculate beforehand, the size of the
INTEGRAL and RWF files, given the input to the G86 program?
It would be most appreciated if anyone could provide some answers
or suggestions. A user is currently trying to run a G86 job on the
SX with 450MBytes of integral file space and 700MBytes of readwrite
file space. And this is NOT enough!!! The program aborted due to
insufficient filespace for the readwrite file.
Another question is: Is this sort of filesize for the integrals and
readwrite file typical?? I'm at a loss as to how to secure more disk
Thanks in advance for any response.
National University of Singapore
It is common for ab initio calculations to use large amounts of disk space.
The current version of Gaussian can trade CPU time for disk space for the
most common types of calculation (Hartree-Fock and MP2), and is much, much
faster, but these capabilities were not present in Gaussian 86.
The integral file size depends on the number of basis functions. It will be
roughly 1.5N^4 bytes or 2N^4 bytes long, depending on how the integral
labels are packed (whether they fill 32 or 64 bits). For open-shell
Hartree-Fock calculations, the size will be 2.5N^4 or 3N^4.
The read-write file size depends on N^2 for Hartree-Fock calculations, and
on N^4 in a complicated way for post-SCF calculations. The easiest way to
estimate the size is to run a smaller job an scale the size by the ratio
of OV^3 for the jobs (O=number of electrons, V=N-O=number of virtuals).
The limitation of not being able to extend these files is specific to the
NEC implementation of Gaussian and is not present on other versions. Contact
your NEC representative for an explanation.
All this is for Gaussian 86, of course. For SCF and MP2 calculations in
Gaussian 90 the integral file isn't needed. For MP2, whatever disk is made
available for the read-write file is used, with quantities being recomputed
instead of trying to use more disk than there is.