Parallel multigrid for large-scale simulations presents unique challenges, due to the need to simultaneously address parallel and algorithmic performance. System resources are often constrained because of demands from other parts of the application, coarse-level problems can have large communication to computation ratios, and load-balancing of the linear systems may be optimized for something other than the linear solver. We consider approaches, such as using more expensive smoothing with more aggressive coarsening, and present numerical results on thousands of processors.