memory requirement ....

[crajesh]

Hi all,

ISTART = 0
ICHARG= 0
ISMEAR = 0
SIGMA = 0.1
# NBANDS =200
EDIFF = 0.0001
EDIFFG = 0.001
IBRION =2
POTIM = 1.0
NSW =100
PREC =LOW
GGA=PE
NPAR=2
ISIF=2
NELM=120
#--------------- save memory--------------
LCHARG = F
LWAVE=F
ISYM=2
# ISPIN=2
#----------------------------------------

This is my simple INCAR file and 1X1X1 gamma Kpoint
My job is running on 8 nodes, each node contains 8 processors and 2GB memory for each proccessor.
So, My OUTCAR file says I require 1546 MB for writing wavefunctions and 3381 MB for predicting wavefunctions
and my NPAR=2
so what exactly is this memory requirement hinting at ..?
Is my configuration sufficient not to have any memory bottlenecks to slow down my calculations ...

awaiting a reply
rajesh

[alex]

Hello rajesh,

if I understand correctly you provide 256MB per process. This is on the very low end, even for small system sizes. You may try to use less processes but all physical nodes (in openmpi you would have to use the mpirun switch --bynode to distribute correctly)

Hth,

alex

[Danny]

If you are running VASP5.X the OUTCAR tels you roughly how much memory you need per process(used core). Since you are doing a gamma point calculation I assume you are doing atoms/molecules in a large empty box. The only way to reduce memory is
1) make the box smaller
2) use lower ENCUT
(both will decrease your accuracy)
3) play with NPAR(this can lower the memory usage, but generally you will lose efficiency)

The alternative is of course try to get your hands on a machine where there is RAM available (for VASP 2gb per core is near the minimum in my opinion)

Danny

[rajeshc]

@Alex
Thank you for the reply. If I have to make myself more clear. Our cluster has 8 processors per node and each node has a total 16GB RAM which is like 2GB per processor. I use openmpi i will checkk this switching on the moirun swith. I hope to get more input from you after my clarification.
@Danny
we use vasp 4.6. I am doing a 300 atom system of pervoskites trying to see som impurity site preference, where experiments were carried out at low dopant concentration ~1mol%. with ABO3 system, to get to a ~1.6mol% also i need to use 4*4*4 super cell.
So i am doing a gamma point and kpoints 1*1*1.
I atleat want some trends ...though the numbers r not accurate.
your comment "2GB per core is near minimum" do you mean 2GB per node" is miminum ..?
awaiing for your comments tooo

[Danny]

ok, that changes issues a bit :-)
With 2Gb per core I meant 2Gb per core. This way, you take also into account that there exist single core processors (standard few years back for personal machines) and multicores (nowadays you easily get quadcore processors in personal machines, supercomputers can go further having even hexacore processors. A node containing 8 processors could this way have between 8 and 128 cores. If your machine has single core cpu's then 16Gb is what I would suggest. (check the header of your OUTCAR to see how many cores are involved)


For your system(and VASP 4.6) you can make a good estimate of the required memory (scales linear with k-points and size, I did some scalingtests in my thesis). Make sure you have enough memory available (use a smaller cell--> 1x1x1 and 2x1x1 to check your memory requirements)
300 atoms is a lot, so you might actually need more nodes to provide the required memory.
The 3.3Gb indication by VASP might actually be 3.3 Gb per core in your current setup (ie > 200Gb total).


In addition, have you already done a calculation with a single unitcell? I see you want to relax that system...make sure that your calculation can succeed in finishing a single static calculation within your available walltime.

And one more addition set your ENCUT manually, otherwise you will have trouble calculating formation energies. (also check which cutoff is required to get sufficiently converged results)

[crajesh]

@danny
What happens if we are low on memory requirement ..? does the job get thrown out or it is going to take long time ..?
thanks

[Danny]

depends, If your system admin setup the system such that you need to ask for memory, the job will be blocked when you ask to much, and normally killed when you use more. If that is not the case the machine will start swapping (which will slow your job down a lot 10x or more slower easily), when it runs out of memory for swapping, the job will crash.