To generate large-scale meshes with highly stretched elements for complex geometries, we propose a distributed parallel Newton-GMRES penalty solver. For each non-linear iteration, we solve a pre-conditioned sparse linear problem. To choose the parallel pre-conditioner, we compare an algebraic multi-grid implementation with a restricted additive Schwarz domain decomposition with one level of overlap and local problems approximated with symmetric successive over-relaxation. We show that domain decomposition is faster and more energy efficient. Furthermore, to accelerate the penalty based solver, we propose a novel p-continuation technique with two unique features. First, it has an early termination criterion to stop the optimization of the initial polynomial degrees. Second, it estimates the initial penalty parameter for each polynomial degree. We conclude that this continuation can reduce four times (eight times) the wall clock time (energy per core) required to curve a whole boundary layer quartic mesh using the chosen domain decomposition.