Memory Issue with ALGO=ACFDTR

[s_ringe]

Dear developers,

I ran into an unexpected memory usage problem while calculating RPA using ALGO=ACFDTR.

In the calculation using 56 processes per node, the memory usage after assigning 28 processes to each node is more than twice as large as the memory usage after giving 56 processes to one node.

In the calculation using one node, each process consumes approximately 2 gigabytes of memory, so I expected similar or slightly larger memory usage (because of the communication buffer between nodes) for two or more nodes, but more than double (approximately 6 gigabytes) I thought the usage was abnormal.

After checking the memory through OUTCAR as well as directly through the terminal, I found that this was not just a reporting issue.

Below is the INCAR that I used.

Code: Select all

NPAR = 56; KPAR = 1

ALGO = ACFDTR; PREC = Accurate; ISTART = 1

ENCUT = 400; ISMEAR = 0; SIGMA = 0.05

NBANDS = 7952

NOMEGA = 24; ENCUTGW = 200

LWAVE = .FALSE.; LCHARG = .FALSE.

Thanks.

[s_ringe]

Dear developers,

For reference, the calculation environment is as follows.

Operating System: Rocky Linux 8.4 (Green Obsidian), Linux 4.18.0-305.el8.x86_64
Compiler: ifort (IFORT) 2021.5.0 20211109
openmpi, oneapi, and hdf5 used

Thanks.

[merzuk.kaltak]

Dear s_ringe,

judging from your INCAR, the memory requirement is estimated automatically.
This happens "under the hood" so to speak. On a Linux machine VASP looks into /proc/meminfo for the entry MemAvailable and sets MAXMEM automatically. This value of MAXMEM is used internally to set NTAUPAR which ultimately influences how much memory is used.
Note, by default VASP is quite "greedy" and grabs 90 percent of available RAM for the each run.

I suspect that NTAUPAR is higher for the job with 28 MPI ranks and thus uses more memory. However, I can say more if you upload a proper error report.

[s_ringe]

Dear merzuk.kaltak,


Thank you so much for your valuable comments!

I checked the output and saw that NTAUPAR had a larger value when using 28 processes.


Thank you and Best Regards,

s_ringe