On Wed, 24 Feb 2010 10:15:21 -0600, Andrea Hawkins <[email protected]> wrote: > I am using the SNES, I just utilized the SNESLineSearchSet to set my own > update.
Okay, is this better than the standard one for your problem? > So, if I do a diff between the output after the final newton step in > the zero-th time step with the the output after the first newton step > of the 1st time step I get > > broglie $ diff step0 step1 > 19c19 > < total: nonzeros=32153, allocated nonzeros=33489 > --- > > total: nonzeros=32153, allocated nonzeros=40509 > 24c24 > < total: nonzeros=32153, allocated nonzeros=38169 > --- > > total: nonzeros=32153, allocated nonzeros=40509 > 27d26 > < This is odd. Do you intend for the sparsity pattern of the matrix to change? If it really shouldn't change, then try setting MAT_NEW_NONZERO_ALLOCATION_ERR and see what triggers the error. > That wouldn't explain it working after restarts though, would it? > Because with a restart, I'm seeding the initial condition with the > previous solution. So, the linear solver should be using this as an > initial guess, correct? Do you mean that you're calling SNESSolve with the current solution in X? That just specifies the first Newton iterate, the Krylov solve will still start with a zero initial guess. (As a general rule with Newton, if you know a better initial guess than 0 for the linear solve, then you should have evaluated the function there.) PETSc solvers should be completely configured before the first XXXSolve. Modification after this should only be done if your code is taking active control (i.e. checking how convergence is proceeding and changing tolerances/preconditioners/etc in some problem-specific way). Maybe if you describe all the ways you modify the SNES/KSP/PC after the first call to SNESSolve, I can guess what is going on. Jed ------------------------------------------------------------------------------ Download Intel® Parallel Studio Eval Try the new software tools for yourself. Speed compiling, find bugs proactively, and fine-tune applications for parallel performance. See why Intel Parallel Studio got high marks during beta. http://p.sf.net/sfu/intel-sw-dev _______________________________________________ Libmesh-users mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/libmesh-users
