ABINIT Discussion Forums

Hello,

I am trying to calculate phonons for 24 atoms system. I calculate it in three steps:

Dataset 1: Ground state calculation
Dataset 2: d/dk calculation
Dataset 3: Response function calculation at Gamma with electric field perturbation

some parameters I have used are:

ecut 46.
ngkpt 6 6 6
nband 48
ixc 11
fhi pseudopotential : Trouiller-Martins-type, GGA Perdew/Burke/Ernzerhof (1996)


abinit estimates the following memory usage:
DATASET 1: This job should need less than 111.604 Mbytes of memory.
DATASET 2: This job should need less than 146.624 Mbytes of memory.
DATASET 3: This job should need less than 176.287 Mbytes of memory.

However, the program always stops when memory usage exceeds 50 GB/node. I am currently using 36 nodes with 9 processors each. Still the program stopped because it exceeded 50 GB/node. Its currently using 1787.16 GB total memory. The 50 GB/node is a upper limit set by admn on the IBM system I am using.

Earlier I used 18 processor/node on 18 nodes and the abinit ran till the 1st direction of 11 atoms (892 GB memory was used). With 9 processors/node on 36 nodes it ran till 3rd direction of 22nd atoms. So, the memory used doubled from the 1st to the 2nd run (1787.16 GB).

Wonder if it is normal to expect such large memory usage.

Thanks,

Vivek

Hello,

0) As always, a post is halfway useless if you do not specify your version, platform and full details. Please follow the nettiquette.

1) are you really using 324 cores? You have relatively few k-points, so this may not work very efficiently. There is some band parallelization, which is automatic, but the i/o is in need of optimization (being done as we speak for the new versions of abinit). Further, the number of k is different for each q, as a function of the remaining symmetry operations. This may explain why some dtset are fine and others not.

2) the memory estimation is not great, but it should not miss by an order of magnitude. I did not understand your comments as to how much memory is being used by whom... 1.7TB for the whole job? This makes 47 Gb/node which should be ok based on what you say, but still a lot more than 9*176 MB.

Matthieu