I am running the FEAST algorithm on Mathematica/10.0. I have sparse matrices of dimension around 500,000, looking for around 50 eigenvalues. I have noticed the runtime depends massively on what the entries are.
For instance, if I have two sparse matrices with precisely the same non-zero entries, I just modify them slightly, the runtime can slow down 5x or more. The specified intervals are the same and the results return very nearly the same number of eigenvalues.
Is this a known issue with FEAST?