PC GAMESS/Firefly-related discussion club

Learn how to ask questions correctly

Alex Granovsky
gran@classic.chem.msu.su

Firefly-specific MP2 energy and energy gradient code was developed
with large jobs in mind as MP2 is relatively cheap. At the same time,
MP2 requires only the specific subset of transformed 2-e integrals,
and the theory is simple. As a results, the memory requirements for
MP2 are very modest with Firefly.

MP3 and especially MP4(full) use another (older) code to perform
2-e integral transformation. Next, the MP3/MP4 code itself
requires more memory than MP2. E.g., the (T) step of MP4(full)
requires at least V³ words of memory (where V is
the number of virtual orbitals)- and this amount is ca. two times
larger than the minimal amount required for MP4 integral
transformation. On the other hand, MP3 and MP4(SDQ) typically
uses much less memory than the (T) part.

The memory requirements of MP3 and esp. MP4 code were not
for a long period considered as a bottleneck as the calculations
themselves are very expensive - e.g., the expression for the number
of flops required by MP4(T) part can be found here
so one can estimate the required CPU time. MP3 and MP4(SDQ) are
cheaper but are also expensive...

However, we do plan to change the integral transformation code
of MP3/MP4, mainly because of efficiency considerations.
The memory demands for the (T) step of MP4(full) can also be
reduced by a factor of two or so...

regards,
Alex








On Sun Nov 15 '09 11:23pm, Sergey wrote
---------------------------------------
>Dear all,

>I have idea to recover a higher order correlation energy from MP4 single point energy run based on MP2 optimized geometry, however I failed to run my job because of memory limits. My job is 800 functions cc-pVTZ for stacked adenine system.

>Can you give me idea how MP4 job memory demands relate to MP2 (RHF, ENERGY, METHOD=1) demands, for single poin energy runs? So far I have for MP2 ~ 48MW and over 390 MW for MP4. Is that really normal or did I miss some input cards?
>
>
>Thank you!