[CIG-SEISMO] global_s362ani instability with certian configuration
Dimitri Komatitsch
komatitsch at lma.cnrs-mrs.fr
Fri Jun 28 08:18:06 PDT 2013
Hello Ping,
It is probably a configuration that nobody has ever tried before and for
which the time step that the code computes automatically is a bit too
big and thus slightly above the code stability limit (the CFL number).
A quick patch is to edit file src/shared/read_compute_parameters.f90, go
to line 754:
! takes a 5% safety margin on the maximum stable time step
! which was obtained by trial and error
DT = DT * (1.d0 - 0.05d0)
and add
DT = DT * 0.9d0
right after it.
Most users of the code use the same configurations (I'd say 5 to 10 in
total are probably used quite often, the others are sometimes / often
untested; and for 3D models there is no exact (mathematical) way of
determining the CFL limit, thus the code computes an estimate that can
sometimes be slightly too high, and over the years for configurations
that are widely used we have improved that and made them stable by
slightly reducing the time step).
Thanks,
Dimitri.
On 06/28/2013 04:52 PM, Luo, Ping wrote:
> Hello,
>
> I obtained the newest development version of SPECFEM3D_GLOBE (not the tar.gz one) and tried to run global_s362ani in the EXAMPLES directory. My intention is to keep the number of elements per core to a constant (48*48) and vary the number of cores. The parameters I set in the Par_file are listed in the table below.
>
> #elements/core = 48*48
>
> NPROC_XI NPROC_ETA NEX_XI NEX_ETA total cores
> 2 2 96 96 6*2*2=24
> 4 4 192 192 6*4*4=96
> 6 6 288 288 6*6*6=196
> 8 8 384 384 6*8*8=384
> 10 10 480 480 6*10*10=600
>
>
> The tests ran to completion except for NPROC 8x8, in which the solver became unstable and blew up. The norm and potential from the timetamp files is listed below. It seems the computation blow up before step 60.
>
> timestamp000005: Max norm displacement vector U in solid in all slices (m) = 1.2461744E-03
> timestamp000005: Max non-dimensional potential Ufluid in fluid in all slices = 1.0009440E-24
> timestamp000020: Max norm displacement vector U in solid in all slices (m) = 9.6447496E-03
> timestamp000020: Max non-dimensional potential Ufluid in fluid in all slices = 1.0034338E-24
> timestamp000040: Max norm displacement vector U in solid in all slices (m) = 1.8385876E-02
> timestamp000040: Max non-dimensional potential Ufluid in fluid in all slices = 1.0065002E-24
> timestamp000060: Max norm displacement vector U in solid in all slices (m) = 154.3319
> timestamp000060: Max non-dimensional potential Ufluid in fluid in all slices = 1.0093262E-24
> timestamp000080: Max norm displacement vector U in solid in all slices (m) = 1.6437107E+11
> timestamp000080: Max non-dimensional potential Ufluid in fluid in all slices = 1.0125660E-24
> timestamp000100: Max norm displacement vector U in solid in all slices (m) = 1.7506350E+20
> timestamp000100: Max non-dimensional potential Ufluid in fluid in all slices = 1.0156563E-24
> timestamp000120: Max norm displacement vector U in solid in all slices (m) = 1.1403834E+26
> timestamp000120: Max non-dimensional potential Ufluid in fluid in all slices = 1.0184948E-24
>
> My questions are: why the 8x8 case crashed, but all other cases worked fine, and how to fix it.
>
>
> Thanks,
>
> Ping
> _______________________________________________
> CIG-SEISMO mailing list
> CIG-SEISMO at geodynamics.org
> http://geodynamics.org/cgi-bin/mailman/listinfo/cig-seismo
>
--
Dimitri Komatitsch
CNRS Research Director (DR CNRS), Laboratory of Mechanics and Acoustics,
UPR 7051, Marseille, France http://komatitsch.free.fr
More information about the CIG-SEISMO
mailing list