A case study in top-down performance estimation for a large-scale parallel application