[gpaw-users] Calculation freezes after memory estimate

[Marcin Dulak]

On 07/03/12 12:17, Juho Arjoranta wrote:
> Lainaus "Marcin Dulak"<Marcin.Dulak at fysik.dtu.dk>:
>
>> On 06/26/12 08:41, Juho Arjoranta wrote:
>>> Hello,
>>>
>>> I have been converging the vacuum around a copper surface with two
>>> bottom layers fixed and with 4 and 6 layers of copper everything went
>>> fine. Now I have an 8 layer surface and the calculation freezes after
>>> the evaluation of memory usage is made. The calculation for the
>>> surface was restarted from a bulk calculation where the initial state
>>> for the surface was set. The bulk was relaxed to make sure that the
>>> structure is really correct.
>> could freezing be ralated to scalapack?
>> I remember similar problems with scalapack on, for example
>> this hangs:
>> https://trac.fysik.dtu.dk/projects/gpaw/browser/trunk/gpaw/test/parallel/scalapack_pdlasrt_hang.py
>>
>> Marcin
> It would seem that the freezing was related to scalapack. I tried to
> run the same calculations with the default eigensolver and without
> scalapack and it seems to be working now. To get the density and wave
> functions to converge I also made some modifications to the calculator:
did you get different energies without/with scalapack?
Do you know which version of scalapack do you use (on our systems
scalapack version 2.0.1 seems to behave better than 1.8.0)?
Can you run, on 4 cores:
https://trac.fysik.dtu.dk/projects/gpaw/browser/trunk/gpaw/test/parallel/scalapack_mpirecv_crash.py
https://trac.fysik.dtu.dk/projects/gpaw/browser/trunk/gpaw/test/parallel/scalapack_pdlasrt_hang.py

Marcin

>
> surface, calc = restart('bulk-8-initial.gpw',
>                           txt = name + '.txt',
>                           xc = 'PBE',
>                           basis = 'szp(dzp)',
>                           mixer=Mixer(beta=0.1, nmaxold=2, weight=100.0),
>                           poissonsolver=PoissonSolver(nn=3, relax='GS',
> eps=1e-12),
>                           parallel={'domain': (1,2,7)},
>                           kpts = (8, 8, 1),
>                           h = 0.12)
>
> And for a different (but similiar system) I also increased the maximum
> number of iterations for the Poisson solver as explained in this
> thread:
> https://listserv.fysik.dtu.dk/pipermail/gpaw-users/2010-October/000408.html
>
> Juho
>
>>> Initially when I used the same calculator options as for the 6 layer
>>> surface the calculation for 8 layers froze. After that I have tried
>>> less agressive mixing, different Poisson solvers, extracting the
>>> single-zeta polarization basis set from the double-zeta polarization
>>> basis sets, and eventually even a different eigensolver.
>>>
>>> These calculation were run at Louhi and the last one I tried was this one:
>>>
>>> from ase.parallel import paropen
>>> from gpaw import restart, Mixer
>>> from gpaw.poisson import PoissonSolver
>>> from ase.constraints import FixAtoms
>>> from ase.optimize import QuasiNewton
>>> from numpy import zeros
>>>
>>> resultfile = paropen('8-layer-cg-results.txt', 'w')
>>>
>>> for vac in range(5, 15):
>>>
>>>           name = '8-layer-cg-%.i-vacuum' % vac
>>>
>>>           k = 8
>>>           a = 3.643       # Lattice constant from the convergence
>>> calculations
>>>
>>>           surface, calc = restart('bulk-8-initial.gpw',
>>>                           txt = name + '.txt',
>>>                           eigensolver = 'cg',
>>>                           mixer=Mixer(beta=0.1, nmaxold=2, weight=100.0),
>>>                           poissonsolver=PoissonSolver(nn=3, relax='J'),
>>>                           basis='szp(dzp)',
>>>                           parallel={'sl_default': (5,1,64),
>>>                                     'domain': (1,1,5)},
>>>                           kpts = (k, k, 1))
>>>
>>>           # Fixing of the two bottom layers
>>>
>>>           fix = 2
>>>
>>>           positions = surface.get_positions()
>>>           number = positions.size / 3
>>>
>>>           array = zeros([number])
>>>           for i in range(0, number):
>>>                   if (surface.positions[i][2]<   fix * a / 2):
>>>                           array[i] = 1
>>>                   surface.set_tags(array)
>>>           c = FixAtoms(mask=[atom.tag == 1 for atom in surface])
>>>           surface.set_constraint(c)
>>>
>>>           surface.pbc = (1, 1, 0)                         # Set PBC
>>> false in z-direction
>>>           surface.center(axis = 2, vacuum = vac)          # Center the
>>> system in z-direction and add vacuum
>>>
>>>           surface.set_calculator(calc)
>>>
>>>           # Relaxation of the surface
>>>
>>>           relax = QuasiNewton(surface, trajectory = name + '.traj')
>>>           relax.run(fmax = 0.05)
>>>
>>> Equivalent code worked with 4 and 6 layers just fine. Then of course
>>> there were no need for different eigensolver or basis etc. For some
>>> reason the only error message I get, is Louhi sending me this to my
>>> email:
>>>
>>> PBS Job Id: 1098436.sdb
>>> Job Name:   8-layer-cg
>>> Post job file processing error; job 1098436.sdb on host nid00143
>>>
>>>
>>> Unable to copy file /var/spool/PBS/spool/1098436.sdb.OU to
>>> nid00139://wrk/arjoran/8-layer-cg.o1098436
>>> error from copy
>>> nid00139: Connection refused
>>> end error output
>>> Output retained on that host in: /var/spool/PBS/undelivered/1098436.sdb.OU
>>>
>>> Any ideas how the get that running?
>>>
>>> Juho Arjoranta
>>>
>>>
>>> _______________________________________________
>>> gpaw-users mailing list
>>> gpaw-users at listserv.fysik.dtu.dk
>>> https://listserv.fysik.dtu.dk/mailman/listinfo/gpaw-users
>>>
>>
>>
>


-- 
***********************************

Marcin Dulak
Technical University of Denmark
Department of Physics
Building 307, Room 229
DK-2800 Kongens Lyngby
Denmark
Tel.: (+45) 4525 3157
Fax.: (+45) 4593 2399
email: Marcin.Dulak at fysik.dtu.dk

***********************************