How many iterations are required for the Lanczos algorithm to converge?

I am trying to find the n smallest eigenvalues and eigenvectors of a NxN SPD matrix using Lanczos method. What is the number of iterations usually required? I mean, does it scale as $O(N)$ or $O(\sqrt{N})$ assuming n << N (n is usually <=10 and N ~ $10^6$ or $10^7$)?

-
See my short note: mathoverflow.net/questions/75370/… – Suvrit Mar 8 '12 at 0:52
Thanks Suvrit. From "Estimating the Largest Eigenvalue by the Power and Lanczos Algorithms with a Random Start by J. Kuczyński and H. Woźniakowski", it looks like for a fixed error, the number of iterations for the Lanczos method scales as $O(ln(N))$ for finding the eigenvalues. Is there a difference if I need to calculate the eigenvectors too? I assume the complexity is not different, but would like to know if that is not so. – user21969 Mar 9 '12 at 1:20