Inquiry About Memory Consumption in VASP 6.6.1

[wilver_muriel]

Dear Developers,

I have observed some rather unusual behavior when using `vasp_gam` (VASP 6.6.1), and I would appreciate your help in identifying the possible cause.

I am trying to calculate the energy of an isolated Ar atom in a 25 × 15 × 25 Å cell, using only the Γ-point and 8 bands. However, the calculation exhibits extremely high memory consumption, which I find rather unreasonable for a system of this size, although I may be overlooking some technical aspect.

The calculations were performed on an HPC cluster using 8 MPI ranks. I tested several memory allocations, ranging from 700 MB to 2800 MB per core. The maximum allocation was therefore 22.4 GB of total memory (8 MPI ranks × 2800 MB per rank), which was the maximum available under the cluster's resource allocation policy. Nevertheless, the calculation consistently terminated due to `OOM Killed` events.

I would have expected the available memory to be sufficient for a calculation of this nature. Therefore, I would like to know whether this behavior is normal or whether it could be related to an issue with the input configuration, the compilation, or possibly a bug in the code.

I also tested VASP 6.3.1 and encountered exactly the same issue: the calculation terminates due to an out-of-memory (`OOM Killed`) event.

I have attached the input and output files for the calculations performed with both VASP versions, for which 2800 MB per core was allocated.

I would greatly appreciate any suggestions or guidance you could provide to help me determine whether the observed memory consumption is expected or if there is an issue that I should investigate.

Thank you very much in advance for your help.

Best regards,

Wilver Muriel

[alexey.tal]

Dear Wilver Muriel,

Thank you for your question.

This is a relatively large calculation: the cell contains a lot of vacuum, and in a plane-wave code vacuum is not free. The number of plane waves scales with the cell volume (ENCUT3/2), independent of how many atoms are in it.

I ran the calculation on 8 MPI ranks and the peak memory was about 30 GB in total (roughly 4 GB per rank). Since the wavefunctions are distributed over ranks, using more ranks lowers the memory per rank. You can reduce the total by shrinking the vacuum region or lowering ENCUT. VASP also prints a memory estimate near the top of OUTCAR, which is useful to check before launching a large run.

Best regards,
Alexey