What kind of cluster you use? Try the relation 4Gbram*Nºcpu or 5 should be enough. In this case I would use 80GB ram in total.

On the other hand, I think it is strange. If it needs more memory, computer use the disk, it goes slower, but It does the opt. Maybe it is a problem of structure more than a RAM problem.

I can use 80 GB as %mem. Then how much will be in maxdisk in the route section?

In my calculation I don't put disk. For MP2 the best way to perform de calculation is with RAM. Maybe could be an option don't give a maxdisk. 

why are you using scf=qc? Is there any convergence problem? usually scf=qc takes more time though it is reliable. I would love to know the actual number of atoms you are dealing with and the atom type. You can increase the %mem and maxdisk depending on your server availability. But I do agree with Joaquim's comment that if system need more memory then either it will take that or it will terminate by asking for that. Please check your input carefully. Hope these may help.  

I used SCF=qc, bacause I tried to optimize the system by hf and dft at the beginning. But using normal DIIS method, there were lots of problem in convergence. So I switched it SCF=qc.

I have seen for MP2 energies, required maxdisk is 2ON^2. I think N is number of basis function. Do you know what is represented by O?

In the "2ON^2" requirement, N is the number of basis functions, while O is the number of occupied orbitals (some methods also scale with V which is the number of virtual orbitals).

I would definitely add maxdisk with some large value to this calculations if %mem=80Gb doesn't work. MP2 has huge memory requirements: i've seen scratch files of hundreds of Gbs for moderate-size systems.

But I think that should be rare that depending on memory, program do not opt the water molecules, and when you remove some metallic centre, do it. Did you try this opt with a dft method? Could be interesting to see if opt works properly.

The exact requirement for your case is 138MW, considering 16 procs, 396 basis functions and hypothesizing that the BS are of d/f type only (which is almost certainly the case). An amount of 138 MW corresponds to 1100 Mb on a 64bit machine, according to Gaussian definition of a W (word). Please note that the memory requirement is not simply 2ON^2, since there is a threshold value.

This said, the amount of memory assigned does not affect the optimization result. If you assign a small amount of memory your machine will only reduce the number of shared processors used (which in turn could slow down your calculation). This is clearly stated in the log file (PrsmSu: requested number of processors reduced to ...). If you use a larger amount of memory (15GB) you will again slow down the calculation, especially if you exceed the amount of RAM available.

Hope this helps.

I coded a small utility software for evaluating the required memory:

http://people.unica.it/massimilianoarca/lang/en-uk/ricerca-scientifica/development-of-software-of-chemical-interest/gaussmem-2014/

%mem=15GB is not enough.

Use at least 128 GB but it would be a very long calculation.Use scf=xqc: it will take more time but you will get  almost surely the stationary state.

Regards, Juan.

Sorry, I fully do not understood your question but this is gonna help to solve your

problem. But i can suggest one solution.

If total energy is fluctuating, the scf=vshift is help to solve your problem.

You mentioned scf=qc in route section --> This keyword gonna make your

calculation much heavier. Please use qc keyword if you have no more chance to

solve convergence problem. 

Hope this help

following