understanding memory consumption

[kanna]

Hi,

In the log

| Size of dense matrices = 10516 x 10516 [1687 MB per matrix]                  |

|                                                                              |
| Storage of orbitals in real space is disabled.                               |
| Enabling storage requires an additional   607 MB.                            |
|                                                                              |
+------------------------------------------------------------------------------+
+------------------------------------------------------------------------------+
|                                                                              |
| Real space grid sampling is (61, 111, 773) in a, b, and c directions.        |
|                                                                              |
+------------------------------------------------------------------------------+
+------------------------------------------------------------------------------+
|                                                                              |
| Storage of atomic core grids requires    39 MB per grid                      |
|                                                                              |
+------------------------------------------------------------------------------+

• I think the dense matrices are formed from the sparse matrices in the calculation , but how many such dense matrices are used ? In my simulation each mpi thread takes 10.9 GB of memory.
• How is the real space sampling controlled ? I am guessing it is related to the resolution of k points we give in the calculation , am I correct to assume this ?
• How many atomic core grids are present ? Is this the number of atoms in the structure * 39MB in this case ?Does this depend on the basis set (SZP/DZP) and exchange correlation alone ?[\li]
  

Thanks
kanna

[Jess Wellendorff]

The of grid points for real space sampling is determined by the size of the simulation cell and by the "Density mesh cut-off" (in Hartree). The larger the cell and the high the cutoff, the more grid points are needed. k-point sampling has nothing to do with real space sampling.

[kanna]

Thanks Jess,
With reference to my above questions,
-Reducing cell size means that I need to have more k-points for the simulation so that I get good accuracy of results, even reducing the mesh cut-off reduces accuracy.
Is there a thumb rule for these parameters ?

Summarizing, to reduce memory consumption , the options are:
1. reduce k point sampling along c direction
2. reducing density mesh cutoff
3. basis set (DZP, SZP)
4. Does exchange correlation significantly affect memory / CPU time ?

[Anders Blom]

Where do you get the 10 GB number from? The grid sampling is how many points the grid contains. Each point is NOT 39 MB, it's the WHOLE grid which is 39 MB. You have 3-4 such grids in the calculation (density, effective potential, ...).

The main parameters to reduce memory consumption is mixing and basis set. See http://docs.quantumwise.com/tutorials/advanced_performance.html