In my own work, I usually perform a final self-consistent run after a relaxation with ISMEAR = -5 to get the total energy for such comparisons due cited reasoning that the tetrahedron method with Blochl corrections (ISMEAR = -5) is generally regarded as most accurate.  I have found that, at least in metals (which are relaxed with ISMEAR = 1), that the difference can be a few meV.  The computational cost of doing so is usually negligible since you can reuse the wavefunction and charge density from the relaxation run.

That being said, as long as your SIGMA is sufficiently small for your system, there isn't anything "wrong" with using ISMEAR = 0 for total energy.  What I mean by this is that the difference dE between F (free energy) and E0 (sigma->0 extrapolation) is smaller than the accuracy you expect from you convergence of k-points, cutoff energy, etc.    This was, in fact, the primary method to use before the papers introducing the tetrahedron method (and tetrahedron method with Blochl corrections) were introduced.  Furthermore, there are cases where you may have no choice but to use ISMEAR in cases where the tetrehedron method can't be used (e.g. if one constituent of your system is a molecule or if  you have such a large cell that you have too few k-points).

If you have a metallic system (or even a small band-gap) the k-point sampling will need to be increased if you reduce the smearing width, in order to get the fully converged result.

Thank you very much for the answers. Unfortunately, I am not sure if my system is metallic; that is why I used Gaussian smearing as a safe option. But I think I will repeat the static calculations and compare the energies.