[aspect-devel] Compositional fields with different densities and viscosities

[Wolfgang Bangerth]

Cedric & Anne,

> I'll resubmit a job tomorrow morning on 1 core and see how the memory is evolving.
> Right now, I have a 1024x1024 (no grid ref , i.e. the input file I sent you) running, claiming 5Gb more by the hour.

I've played with this a bit but I can't really reproduce it. Here's what 
I tried:

1/ Single processor, 4 and 6 levels of global refinement: I'm running 
this under valgrind's massif tool which gives me a graphical 
representation of memory consumption as a function of time (or, rather, 
CPU instructions). This turns out to be completely flat -- no increase 
in memory over time.

2/ 8-processor MPI job, 6 levels, running under massif: also, no 
increase in memory consumption for any of the 8 MPI jobs. The command I 
used is
   mpirun -np 8 valgrind --tool=massif ./lib/aspect vkk.prm
which yields 8 output files that I can annotate via ms_print (part of 
valgrind's massif).

3/ 32-processors, 8 levels of refinement, no massif: massif is too slow 
(runs ~20 times slower than the original program) to run very large 
jobs, but I'm just running this job just so and monitor memory 
consumption with 'top'. I started the job 3 hours ago, it's at 
approximately time=1500 now (so 3/4 done) and the memory consumption is 
+- the same as right after starting it.

That leaves me with only three hypotheses:
1/ It's something specific to your environment.
2/ It's got something to do with the Trilinos/deal.II/Aspect version you 
are using. Can you send the versions/svn revisions of each?
3/ Or, maybe, there is a problem with refining 10 instead of 8 times. Do 
you see the same effect when you refine only 8 times? 6 times? 4 times?

It would be interesting to see if there is a difference between your 
tools and mine. I don't have ganglia (which requires significant 
infrastructure to install on a cluster), so can't test this. But could 
you possibly run your jobs as I did above with valgrind/massif? It's 
going to be very slow, so it would be useful to know up front whether 
the same problem exists on coarser meshes as well.

Best & sorry for not being more help here...
  W.
-- 
------------------------------------------------------------------------
Wolfgang Bangerth               email:            bangerth at math.tamu.edu
                                 www: http://www.math.tamu.edu/~bangerth/