Dear Kuilin.

Thank you for your suggestions. It seems that the the memory usage is critical. Finally I managed to run VDW with trunk-367 version. However, it needs a huge amount of memory. In my case (128 C atoms, DZP basis) I need to declare around 64GB. Unfortunately, I haven't succeeded with trunk-408 version yet.

Best regards.
Mariusz.



On Fri, 4 May 2012, kuilin lu wrote:

Dear Mariusz,

I guess 3 possibilities:

1.  Did the 128 atoms system really eat up the memory? Is there any
memory limitation on the compute/cluster system ? Some job scheduler
such as torque could kill jobs when process's memory is exceeding
limitation.

2.   I found siesta-trunk-408 made a lot of change in the parallel
scheme. I get a little different LDA result of cellxc when writing my
own program(a serial program, my program maybe wrong), So I rolled
back to trunk-387 and trunk-367, both could produce expected LDA
result. since trunk-387 and trunk-367 could also produce  VDW DRSLL
result, maybe you could try those two versions.

3.  Another possibility is in dhscf.F, I found ifort (version
10.1.018)tends to miscompile the statement similar as:
  DRho(:,ispin) = DRho(:,ispin) + rhoatm(:) / nsd
  in dhscf.F, even with "-g" option without any optimization. Though
not always the case, it has small chance to crash just after printing
"InitMesh". I don't know why.
  If one replace statement
  DRho(:,ispin) = DRho(:,ispin) + rhoatm(:) / nsd
  to
    do ispin = 1,nsd
       do ip= 1, ntpl
         DRho(ip,ispin) = DRho(ip,ispin) + rhoatm(ip)/nsd
       enddo
    endo
  everything goes fine.


Best Wishes,
Kuilin

On 5/4/12, Mariusz Krawiec <[email protected]> wrote:

Dear SIESTA users.

I face a problem with VDW DRSLL functional (siesta-trunk-408). For a small
number of C atoms (let's say 72) everything goes fine. However, if I
increase the number of atoms to 128, siesta crashes just after
initliaziation of the density matrix with the message:
"rank 2 in job 1 ... caused collective abort of all ranks
  exit status of rank: killed by signal 11"
This never happens for a fewer atom system. Note that LDA calculations
with large number of atoms also go fine.
It seems this is not related to the memory, since 72-atom system works
with 0.5 GB, while 128 - doesn't work even with 64 GB.

System: Xeon quad core cluster, ifort 11.1, mkl 10.2, impi 3.2.

Thanks in advance for any suggestion.
All the best.
Mariusz.




Responder a