> I suspect your custom preconditioner is computing NaN for the call to the > inner solver. Run with -fp_trap -start_in_debugger to see where NaN was > first generated.
Very useful hint that with -fp_trap. I get trapped at: Program received signal SIGFPE, Arithmetic exception.0x00000000010fad4e in hypre_BoomerAMGRelax (A=0x1b56e9a0, f=0x195487b0,?? ? cf_marker=0x692b900, relax_type=6, relax_points=1, relax_weight=1,?? ? omega=1, l1_norms=0x0, u=0xb0f8aa0, Vtemp=0x18fd1900, Ztemp=0x0)? ? at par_relax.c:19251925 ? ? ?u_data[i] = res / A_diag_data[A_diag_i[i]]; with res = -2.4683099805590231e+303 and A_diag_data[A_diag_i[i]] = 9.287739686222747e-06. The full backtrace goes through many boomeramg lines to point back at the end to my code where (in a custom preconditioner) I solve a linear system with KSPGMRES and PCHYPRE/boomeramg. I guess I will proceed now with inspecting the values by hand, but if there still are some useful Petsc tricks I am eager to hear.
