On Sun, Jul 15, 2012 at 7:49 PM, Yun Tao <[email protected]> wrote: > Another interesting observation: > > Solution leakage or inflation is most noticeable when the diffusion > coefficient > convection coefficient whereas it's almost negligible when > the inequality is flipped. One way to ameliorate the effect is to use > linearGMRES as solver, however, in contrast to the constant memory use by > the default solver, it accumulates machine memory over time steps at a > radical rate (90MB vs. 2GB over 2000 time step). Pretty intriguing stuff. >
I am not sure what a systematic error means regarding the conservation of the field. If it is influenced by the choice of solver, the iteration count and the tolerance then I think it is just solver round off that you are dealing with. If the tolerance and iterations are increased does that improve things? Is the error magnitude you are seeing consistent with solver round off for the tolerance that is being used in the solver? Rerunning your code from your first email. It looks like there is only 2e-10 after 2000 steps. I think the solver tolerance is set at 1e-10 by default, so this isn't really surprising. As for the memory leak that could well be. For some reason, I rarely use the Pysparse GMRES solver because it has given me problems in the past. Try using Trilinos's GMRES solver if you can and see if it helps. Cheers -- Daniel Wheeler
_______________________________________________ fipy mailing list [email protected] http://www.ctcms.nist.gov/fipy [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ]
