Alex Granovsky
gran@classic.chem.msu.su
rsize1 mainly affects memory needs of the first half-transformation stage.
mxi mainly affects memory needs of the MP2 part of second half-transformation stage.
rsize2 mainly affects memory needs of the DM2 part of second half-transformation stage.
MXPQ cannot be lowered below some threshold defined by the particular basis set in use, and affects DM2 part of the second half-transformation stage as well as 2-e gradient calculation stage.
Turning off all async I/O options will save you some extra memory.
Best regards,
Alex Granovsky
On Wed Oct 3 '07, Roman Zubatyuk wrote
--------------------------------------
>Dear Dr. Granovsky,
> I have tryed changing MXPQ and RSIZE2, but it seems that it has no effect on memory needed for second half-transformation. At least as reported by exetyp=check..
>all options are default:
># OF WORDS USED FOR MP2 = 123097625
>$mp2grd MXPQ=2 rsize2=3000 $end :
># OF WORDS USED FOR MP2 = 123097625
>Is it a bug of check procedures?
>How can I guess the optimal parameters for large MP2 calculation?
>Best regards,
> Roman Zubatyuk
>
>On Mon Oct 1 '07, Alex Granovsky wrote
>--------------------------------------
>>Hi,
>>I would recommend to halve RSIZE1 and/or RSIZE2 as the first attempt.
>>
>>
>>>1. If I'll halve MXI (or MXPQ), integral half-transformation (first or second) will be performed in two passes and thus approximately twice less memory will be needed.
>>This is not exactly the case. The new algorithm is really complex.
>>MXI affects the amount of disk I/O. Setting MXI to the half of
>>the number of active occupied orbitals do decrease memory
>>requirements to some extent but will increase disk I/O.
>>Smaller MXPQ will increase memory requirements of some
>>parts of code and decrease requirements of other parts.
>>>2. Is there a formula to estimate memory requirement for various steps of MP2 calculation? Or the only solution is to perform check run? Should I run it in parallel, or serial check run will be sufficient?
>>The formulas are so complex so that it is much simpler to perform check run. You can use single-node run as an approximate upper limit,
>>however, parallel run typically uses much less memory. What is even worse, the memory requirements can slightly vary with the molecular geometry.
>>
>>
>>Best regards,
>>Alex Granovsky
>>
>>
>>On Mon Oct 1 '07, Roman Zubatyuk wrote
>>--------------------------------------
>>>Dear PC GAMESS developers/users.
>>>I got insufficient memory error while performing large MP2 gradient calculation (Nocc=129, Ncore=52, Nvirt=981, distributed over 24 nodes each having 1GB of RAM). There are four options for the code those 'seriously affect amount of memory required': MXI, MXPQ, RSIZE1 and RSIZE2. Regarding this I have two questions:
>>>1. If I'll halve MXI (or MXPQ), integral half-transformation (first or second) will be performed in two passes and thus approximately twice less memory will be needed. Am I right? Should I also change RSIZE1 (or RSIZE2)?
>>>2. Is there a formula to estimate memory requirement for various steps of MP2 calculation? Or the only solution is to perform check run? Should I run it in parallel, or serial check run will be sufficient?
>>>Thanks in advance.
>>> Roman Zubatyuk