|
HXYZ | | Parallelization of frequency analysis over coordinates. This is
the default. |
|
HSCF | | Parallelization of frequency analysis over the SCF. |
|
GRAM | | Store inverse Coulomb matrix in RAM. |
Description:
In a parallel single-point calculation, the work is distributed inside the
SCF. That is, the integral calculation and linear algebra are parallelized
and therefore the various CPUs all work on the same SCF calculation. We call
this parallelization over the SCF. For the calculation of second derivatives
(see 4.8.1 and 4.6.5), a different parallel assignment is
made. Here each CPU works alone on a distinct SCF cycle. We call this
parallelization over the coordinates. This is the default for the numerical
second derivative calculation. But such parallelization also means that
each process uses a complete set of matrices and files, just as a calculation
on a single CPU would. As a result, performance problems can occur on some
computational architectures, mainly because of swapping and/or intensive I/O.
In those cases, use of the option HSCF will change the parallelization scheme to
the one which is used for a single-point SCF calculation, i.e. all
CPUs work on the same SCF. That approach is less efficient and should be
used only if necessary.
The GRAM option of the PARALLEL keyword triggers the in-core storage of the
inverse Coulomb matrix [25] in frequency calculations. This increases
the RAM demand of the calculation but reduces I/O operations. We recommend this
option for parallel runs of smaller systems with less than 500 atoms. This
option has no effect when the option HSCF is also active.
