Virtual memory

[bourgeois]

Dear Abinit users,

I'm currently running calculations on quite big unit cells (let's say 50 to a bit more than 100 atoms, so the number of bands is quite high). I run abinit in parallel mode, and the problem appears whether I specify paral_kgb (and related input variables) or not.
The log file estimates reasonable memory needs (less than 1Gb) and allocate it sucessfully:

Code: Select all

 Biggest array : cg(disk), with    479.1432 MBytes.
-P-0000  leave_test : synchronization done...
 memana : allocated an array of    479.143 Mbytes, for testing purposes.
 memana : allocated     671.189 Mbytes, for testing purposes.
 The job will continue

When Abinit reads the Psp data and computes the splines, it crashes and returns an error saying "insufficient virtual memory":

Code: Select all

- pspatm: opening atomic psp file    O_Fabio.cpi
 Oxygen, fhi98PP : Troullier-type, GGA PBE, l=2 local (rs=rd=def.,rp=1.60)
   8.00000   6.00000  20050401                znucl, zion, pspdat
    6   11    2    2       473   0.00000      pspcod,pspxc,lmax,lloc,mmax,r2well
  No XC core correction.
  1.024700          amesh (Hamman grid)
forrtl: severe (41): insufficient virtual memory
rank 17 in job 1  compute-0-14.local_34170   caused collective abort of all ranks
  exit status of rank 17: killed by signal 9

Can anyone tell me about this virtual memory? What exactly is needed, for what purpose, how is it related to the input variables? I'm surprised cause each node I use for the parallel computation has 2Gb of RAM and the disk space available (which I though would be use as virtual memory) is really huge...

Thanks for your help!
Regards,

A. Bourgeois

[pmanglade]

Hi,
There is actually nothing special about the virtual memory. It's just referring to the RAM + swap space. Have you tried to monitor your memory usage on various node (using top or a system monitoring tool). It could be interesting to see if Abinit really launches a single job on each node (for instance an MPI configuration error could lead to running every instance of abinip on a single node); or if it really tries to allocate much more memory than what was foreseen in memana.
PMA

[mverstra]

As a follow up to PM's post: the memory estimation for kgb is unfortunately not very precise yet. I have also had these problems, and in practice it's just that abinit really does need much more memory (than 1Gb) and is trying to allocate it.

matthieu

[bourgeois]

Hi,

I did some tests on my own and read the post "Memory usage OpenMPI configuration", this looks closely related to the problem I got... So I'll post a comment on that thread instead of here
Thanks for your answers,

Antonin