When a large number of basis functions is used for dense bulk systems with fcc, hcp, and bcc like structures, the basis set tends to be overcomplete. In such a case, you may observe erratic eigenvalues. To avoid the overcompleteness, a small number of optimized basis functions should be used.
For a system with a highly symmetric structure, the lapack diagonalization routines may fail to find the correct eigenvectors. while this phenomenon strongly depends on the computational environment. For such a case, try to find a nicely working routine by 'scf.lapack.dste'.
For large-scale systems with a complex (non-collinear) magnetic structure, a metallic electric structure, or the mixture, it is quite difficult to get the SCF convergence. In such a case, one has to mix the charge density very slowly, indicating that the number of SCF steps to get the convergence becomes large unfortunately.
For weak interacting systems such as molecular systems, it is not easy to obtain a completely optimized structure, leading that the large number of iteration steps is required. Although the default value of criterion for geometrical optimization is Hartree/bohr for the largest force, it would be a compromise to increase the criterion from to in such a case.