[aspect-devel] Problems with recent deal.ii

[Ian Rose]

H
i all,

Here's a bit of a doozy of a problem.  I noticed the other day a sharp
increase in the number of iterations aspect required for the Stokes solver
to converge.  As in, the cheap solver never converges, and the expensive
one takes ~50-200 iterations.  After a bunch of experimentation, I think it
may be due to some recent changes in deal.ii.  If I run
simple_2d_shells.prm on 8 cores with deal.ii 7.3.0, I get the following
representative output:

*** Timestep 148:  t=4.6211e+06 years
   Solving temperature system... 13 iterations.
   Solving Stokes system... 13 iterations.

If, however, I run it using deal.ii r29892 I get the following

*** Timestep 27:  t=2.68009e+06 years
   Solving temperature system... 13 iterations.
   Solving Stokes system... 30+219 iterations.

I have no idea what may be causing this, or how reproducible it is across
different systems (though I have noticed the same increase in iterations on
TACC Stampede).  In both cases, I am using trilinos 11.2.3 and p4est 0.3.4.
 Everything is built with icc-12.1, mkl math libraries, and using myrinet
MPI.  Any insight would be greatly appreciated.
Best,
Ian
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://geodynamics.org/pipermail/aspect-devel/attachments/20130701/0cf8e002/attachment.html>