[aspect-devel] free surf in 3D

Wolfgang Bangerth bangerth at tamu.edu
Sun May 31 13:26:30 PDT 2015


> I have been trying to do some simple crustal 3D runs with my visco-plastic
> rheology and I am now using the free surface.
> Since my rheology is nonlinear I use Iterated Stokes, so I had to ‘force’
> Ian’s routines to accept this solver.

Ian's patch to allow Iterated Stokes with free surface was merged yesterday. 
Want to retry?

> When I use no AMR, things go rather well, I get enormous mountains, very large
> (surface) deformation, very happy (see pic).
> When I turn AMR on, the model does some tens of time steps and then stalls:
> the code does not crash and
> my job still ‘runs’ on the cluster and uses 10’s of cores but their activity
> is zero and nothing happens for hours.
> (I have checked the mesh, the fields, etc … everything looks as expected, no
> visually suspicious clue).
> I have been putting flags all over the place to figure out where precisely it
> seems to stop and I have narrowed it down to
> the highlighted line on the attached pic (taken
> from Simulator<dim>::FreeSurfaceHandler::make_constraints()
> in  source/simulator/freesurface.cc <http://freesurface.cc> )
> The code does not seem to pass this line and stalls there.
> Would anybody have an idea of what could go wrong there ? and how to fix it ? :)

It would very useful to know where inside that call things go wrong. How many 
processors do you need to reproduce this?

If it's a manageably small number, the usual approach is to run the program as in
   mpirun -np 4 xterm -e gdb ./aspect
and then run every MPI instance inside their own gdb inside their own xterm 
window. When the program hangs, you can Ctrl-C all of them and get a 
backtrace. Typically in cases such as yours, the 4 programs are stopped in 
different places in a deadlock.

A demonstration of this is here:

Wolfgang Bangerth               email:            bangerth at math.tamu.edu
                                 www: http://www.math.tamu.edu/~bangerth/

More information about the Aspect-devel mailing list